]> Pileus Git - ~andy/linux/blob - drivers/dma/ioat/dma.c
dmaengine: consolidate tx_status functions
[~andy/linux] / 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/slab.h>
31 #include <linux/pci.h>
32 #include <linux/interrupt.h>
33 #include <linux/dmaengine.h>
34 #include <linux/delay.h>
35 #include <linux/dma-mapping.h>
36 #include <linux/workqueue.h>
37 #include <linux/prefetch.h>
38 #include <linux/i7300_idle.h>
39 #include "dma.h"
40 #include "registers.h"
41 #include "hw.h"
42
43 #include "../dmaengine.h"
44
45 int ioat_pending_level = 4;
46 module_param(ioat_pending_level, int, 0644);
47 MODULE_PARM_DESC(ioat_pending_level,
48                  "high-water mark for pushing ioat descriptors (default: 4)");
49
50 /* internal functions */
51 static void ioat1_cleanup(struct ioat_dma_chan *ioat);
52 static void ioat1_dma_start_null_desc(struct ioat_dma_chan *ioat);
53
54 /**
55  * ioat_dma_do_interrupt - handler used for single vector interrupt mode
56  * @irq: interrupt id
57  * @data: interrupt data
58  */
59 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data)
60 {
61         struct ioatdma_device *instance = data;
62         struct ioat_chan_common *chan;
63         unsigned long attnstatus;
64         int bit;
65         u8 intrctrl;
66
67         intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET);
68
69         if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN))
70                 return IRQ_NONE;
71
72         if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) {
73                 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
74                 return IRQ_NONE;
75         }
76
77         attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET);
78         for_each_set_bit(bit, &attnstatus, BITS_PER_LONG) {
79                 chan = ioat_chan_by_index(instance, bit);
80                 tasklet_schedule(&chan->cleanup_task);
81         }
82
83         writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
84         return IRQ_HANDLED;
85 }
86
87 /**
88  * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
89  * @irq: interrupt id
90  * @data: interrupt data
91  */
92 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data)
93 {
94         struct ioat_chan_common *chan = data;
95
96         tasklet_schedule(&chan->cleanup_task);
97
98         return IRQ_HANDLED;
99 }
100
101 /* common channel initialization */
102 void ioat_init_channel(struct ioatdma_device *device, struct ioat_chan_common *chan, int idx)
103 {
104         struct dma_device *dma = &device->common;
105         struct dma_chan *c = &chan->common;
106         unsigned long data = (unsigned long) c;
107
108         chan->device = device;
109         chan->reg_base = device->reg_base + (0x80 * (idx + 1));
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         init_timer(&chan->timer);
115         chan->timer.function = device->timer_fn;
116         chan->timer.data = data;
117         tasklet_init(&chan->cleanup_task, device->cleanup_fn, data);
118         tasklet_disable(&chan->cleanup_task);
119 }
120
121 /**
122  * ioat1_dma_enumerate_channels - find and initialize the device's channels
123  * @device: the device to be enumerated
124  */
125 static int ioat1_enumerate_channels(struct ioatdma_device *device)
126 {
127         u8 xfercap_scale;
128         u32 xfercap;
129         int i;
130         struct ioat_dma_chan *ioat;
131         struct device *dev = &device->pdev->dev;
132         struct dma_device *dma = &device->common;
133
134         INIT_LIST_HEAD(&dma->channels);
135         dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
136         dma->chancnt &= 0x1f; /* bits [4:0] valid */
137         if (dma->chancnt > ARRAY_SIZE(device->idx)) {
138                 dev_warn(dev, "(%d) exceeds max supported channels (%zu)\n",
139                          dma->chancnt, ARRAY_SIZE(device->idx));
140                 dma->chancnt = ARRAY_SIZE(device->idx);
141         }
142         xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
143         xfercap_scale &= 0x1f; /* bits [4:0] valid */
144         xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale));
145         dev_dbg(dev, "%s: xfercap = %d\n", __func__, xfercap);
146
147 #ifdef  CONFIG_I7300_IDLE_IOAT_CHANNEL
148         if (i7300_idle_platform_probe(NULL, NULL, 1) == 0)
149                 dma->chancnt--;
150 #endif
151         for (i = 0; i < dma->chancnt; i++) {
152                 ioat = devm_kzalloc(dev, sizeof(*ioat), GFP_KERNEL);
153                 if (!ioat)
154                         break;
155
156                 ioat_init_channel(device, &ioat->base, i);
157                 ioat->xfercap = xfercap;
158                 spin_lock_init(&ioat->desc_lock);
159                 INIT_LIST_HEAD(&ioat->free_desc);
160                 INIT_LIST_HEAD(&ioat->used_desc);
161         }
162         dma->chancnt = i;
163         return i;
164 }
165
166 /**
167  * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
168  *                                 descriptors to hw
169  * @chan: DMA channel handle
170  */
171 static inline void
172 __ioat1_dma_memcpy_issue_pending(struct ioat_dma_chan *ioat)
173 {
174         void __iomem *reg_base = ioat->base.reg_base;
175
176         dev_dbg(to_dev(&ioat->base), "%s: pending: %d\n",
177                 __func__, ioat->pending);
178         ioat->pending = 0;
179         writeb(IOAT_CHANCMD_APPEND, reg_base + IOAT1_CHANCMD_OFFSET);
180 }
181
182 static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan)
183 {
184         struct ioat_dma_chan *ioat = to_ioat_chan(chan);
185
186         if (ioat->pending > 0) {
187                 spin_lock_bh(&ioat->desc_lock);
188                 __ioat1_dma_memcpy_issue_pending(ioat);
189                 spin_unlock_bh(&ioat->desc_lock);
190         }
191 }
192
193 /**
194  * ioat1_reset_channel - restart a channel
195  * @ioat: IOAT DMA channel handle
196  */
197 static void ioat1_reset_channel(struct ioat_dma_chan *ioat)
198 {
199         struct ioat_chan_common *chan = &ioat->base;
200         void __iomem *reg_base = chan->reg_base;
201         u32 chansts, chanerr;
202
203         dev_warn(to_dev(chan), "reset\n");
204         chanerr = readl(reg_base + IOAT_CHANERR_OFFSET);
205         chansts = *chan->completion & IOAT_CHANSTS_STATUS;
206         if (chanerr) {
207                 dev_err(to_dev(chan),
208                         "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
209                         chan_num(chan), chansts, chanerr);
210                 writel(chanerr, reg_base + IOAT_CHANERR_OFFSET);
211         }
212
213         /*
214          * whack it upside the head with a reset
215          * and wait for things to settle out.
216          * force the pending count to a really big negative
217          * to make sure no one forces an issue_pending
218          * while we're waiting.
219          */
220
221         ioat->pending = INT_MIN;
222         writeb(IOAT_CHANCMD_RESET,
223                reg_base + IOAT_CHANCMD_OFFSET(chan->device->version));
224         set_bit(IOAT_RESET_PENDING, &chan->state);
225         mod_timer(&chan->timer, jiffies + RESET_DELAY);
226 }
227
228 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
229 {
230         struct dma_chan *c = tx->chan;
231         struct ioat_dma_chan *ioat = to_ioat_chan(c);
232         struct ioat_desc_sw *desc = tx_to_ioat_desc(tx);
233         struct ioat_chan_common *chan = &ioat->base;
234         struct ioat_desc_sw *first;
235         struct ioat_desc_sw *chain_tail;
236         dma_cookie_t cookie;
237
238         spin_lock_bh(&ioat->desc_lock);
239         /* cookie incr and addition to used_list must be atomic */
240         cookie = dma_cookie_assign(tx);
241         dev_dbg(to_dev(&ioat->base), "%s: cookie: %d\n", __func__, cookie);
242
243         /* write address into NextDescriptor field of last desc in chain */
244         first = to_ioat_desc(desc->tx_list.next);
245         chain_tail = to_ioat_desc(ioat->used_desc.prev);
246         /* make descriptor updates globally visible before chaining */
247         wmb();
248         chain_tail->hw->next = first->txd.phys;
249         list_splice_tail_init(&desc->tx_list, &ioat->used_desc);
250         dump_desc_dbg(ioat, chain_tail);
251         dump_desc_dbg(ioat, first);
252
253         if (!test_and_set_bit(IOAT_COMPLETION_PENDING, &chan->state))
254                 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
255
256         ioat->active += desc->hw->tx_cnt;
257         ioat->pending += desc->hw->tx_cnt;
258         if (ioat->pending >= ioat_pending_level)
259                 __ioat1_dma_memcpy_issue_pending(ioat);
260         spin_unlock_bh(&ioat->desc_lock);
261
262         return cookie;
263 }
264
265 /**
266  * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair
267  * @ioat: the channel supplying the memory pool for the descriptors
268  * @flags: allocation flags
269  */
270 static struct ioat_desc_sw *
271 ioat_dma_alloc_descriptor(struct ioat_dma_chan *ioat, gfp_t flags)
272 {
273         struct ioat_dma_descriptor *desc;
274         struct ioat_desc_sw *desc_sw;
275         struct ioatdma_device *ioatdma_device;
276         dma_addr_t phys;
277
278         ioatdma_device = ioat->base.device;
279         desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys);
280         if (unlikely(!desc))
281                 return NULL;
282
283         desc_sw = kzalloc(sizeof(*desc_sw), flags);
284         if (unlikely(!desc_sw)) {
285                 pci_pool_free(ioatdma_device->dma_pool, desc, phys);
286                 return NULL;
287         }
288
289         memset(desc, 0, sizeof(*desc));
290
291         INIT_LIST_HEAD(&desc_sw->tx_list);
292         dma_async_tx_descriptor_init(&desc_sw->txd, &ioat->base.common);
293         desc_sw->txd.tx_submit = ioat1_tx_submit;
294         desc_sw->hw = desc;
295         desc_sw->txd.phys = phys;
296         set_desc_id(desc_sw, -1);
297
298         return desc_sw;
299 }
300
301 static int ioat_initial_desc_count = 256;
302 module_param(ioat_initial_desc_count, int, 0644);
303 MODULE_PARM_DESC(ioat_initial_desc_count,
304                  "ioat1: initial descriptors per channel (default: 256)");
305 /**
306  * ioat1_dma_alloc_chan_resources - returns the number of allocated descriptors
307  * @chan: the channel to be filled out
308  */
309 static int ioat1_dma_alloc_chan_resources(struct dma_chan *c)
310 {
311         struct ioat_dma_chan *ioat = to_ioat_chan(c);
312         struct ioat_chan_common *chan = &ioat->base;
313         struct ioat_desc_sw *desc;
314         u32 chanerr;
315         int i;
316         LIST_HEAD(tmp_list);
317
318         /* have we already been set up? */
319         if (!list_empty(&ioat->free_desc))
320                 return ioat->desccount;
321
322         /* Setup register to interrupt and write completion status on error */
323         writew(IOAT_CHANCTRL_RUN, chan->reg_base + IOAT_CHANCTRL_OFFSET);
324
325         chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
326         if (chanerr) {
327                 dev_err(to_dev(chan), "CHANERR = %x, clearing\n", chanerr);
328                 writel(chanerr, chan->reg_base + IOAT_CHANERR_OFFSET);
329         }
330
331         /* Allocate descriptors */
332         for (i = 0; i < ioat_initial_desc_count; i++) {
333                 desc = ioat_dma_alloc_descriptor(ioat, GFP_KERNEL);
334                 if (!desc) {
335                         dev_err(to_dev(chan), "Only %d initial descriptors\n", i);
336                         break;
337                 }
338                 set_desc_id(desc, i);
339                 list_add_tail(&desc->node, &tmp_list);
340         }
341         spin_lock_bh(&ioat->desc_lock);
342         ioat->desccount = i;
343         list_splice(&tmp_list, &ioat->free_desc);
344         spin_unlock_bh(&ioat->desc_lock);
345
346         /* allocate a completion writeback area */
347         /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
348         chan->completion = pci_pool_alloc(chan->device->completion_pool,
349                                           GFP_KERNEL, &chan->completion_dma);
350         memset(chan->completion, 0, sizeof(*chan->completion));
351         writel(((u64) chan->completion_dma) & 0x00000000FFFFFFFF,
352                chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
353         writel(((u64) chan->completion_dma) >> 32,
354                chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
355
356         tasklet_enable(&chan->cleanup_task);
357         ioat1_dma_start_null_desc(ioat);  /* give chain to dma device */
358         dev_dbg(to_dev(chan), "%s: allocated %d descriptors\n",
359                 __func__, ioat->desccount);
360         return ioat->desccount;
361 }
362
363 /**
364  * ioat1_dma_free_chan_resources - release all the descriptors
365  * @chan: the channel to be cleaned
366  */
367 static void ioat1_dma_free_chan_resources(struct dma_chan *c)
368 {
369         struct ioat_dma_chan *ioat = to_ioat_chan(c);
370         struct ioat_chan_common *chan = &ioat->base;
371         struct ioatdma_device *ioatdma_device = chan->device;
372         struct ioat_desc_sw *desc, *_desc;
373         int in_use_descs = 0;
374
375         /* Before freeing channel resources first check
376          * if they have been previously allocated for this channel.
377          */
378         if (ioat->desccount == 0)
379                 return;
380
381         tasklet_disable(&chan->cleanup_task);
382         del_timer_sync(&chan->timer);
383         ioat1_cleanup(ioat);
384
385         /* Delay 100ms after reset to allow internal DMA logic to quiesce
386          * before removing DMA descriptor resources.
387          */
388         writeb(IOAT_CHANCMD_RESET,
389                chan->reg_base + IOAT_CHANCMD_OFFSET(chan->device->version));
390         mdelay(100);
391
392         spin_lock_bh(&ioat->desc_lock);
393         list_for_each_entry_safe(desc, _desc, &ioat->used_desc, node) {
394                 dev_dbg(to_dev(chan), "%s: freeing %d from used list\n",
395                         __func__, desc_id(desc));
396                 dump_desc_dbg(ioat, desc);
397                 in_use_descs++;
398                 list_del(&desc->node);
399                 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
400                               desc->txd.phys);
401                 kfree(desc);
402         }
403         list_for_each_entry_safe(desc, _desc,
404                                  &ioat->free_desc, node) {
405                 list_del(&desc->node);
406                 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
407                               desc->txd.phys);
408                 kfree(desc);
409         }
410         spin_unlock_bh(&ioat->desc_lock);
411
412         pci_pool_free(ioatdma_device->completion_pool,
413                       chan->completion,
414                       chan->completion_dma);
415
416         /* one is ok since we left it on there on purpose */
417         if (in_use_descs > 1)
418                 dev_err(to_dev(chan), "Freeing %d in use descriptors!\n",
419                         in_use_descs - 1);
420
421         chan->last_completion = 0;
422         chan->completion_dma = 0;
423         ioat->pending = 0;
424         ioat->desccount = 0;
425 }
426
427 /**
428  * ioat1_dma_get_next_descriptor - return the next available descriptor
429  * @ioat: IOAT DMA channel handle
430  *
431  * Gets the next descriptor from the chain, and must be called with the
432  * channel's desc_lock held.  Allocates more descriptors if the channel
433  * has run out.
434  */
435 static struct ioat_desc_sw *
436 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat)
437 {
438         struct ioat_desc_sw *new;
439
440         if (!list_empty(&ioat->free_desc)) {
441                 new = to_ioat_desc(ioat->free_desc.next);
442                 list_del(&new->node);
443         } else {
444                 /* try to get another desc */
445                 new = ioat_dma_alloc_descriptor(ioat, GFP_ATOMIC);
446                 if (!new) {
447                         dev_err(to_dev(&ioat->base), "alloc failed\n");
448                         return NULL;
449                 }
450         }
451         dev_dbg(to_dev(&ioat->base), "%s: allocated: %d\n",
452                 __func__, desc_id(new));
453         prefetch(new->hw);
454         return new;
455 }
456
457 static struct dma_async_tx_descriptor *
458 ioat1_dma_prep_memcpy(struct dma_chan *c, dma_addr_t dma_dest,
459                       dma_addr_t dma_src, size_t len, unsigned long flags)
460 {
461         struct ioat_dma_chan *ioat = to_ioat_chan(c);
462         struct ioat_desc_sw *desc;
463         size_t copy;
464         LIST_HEAD(chain);
465         dma_addr_t src = dma_src;
466         dma_addr_t dest = dma_dest;
467         size_t total_len = len;
468         struct ioat_dma_descriptor *hw = NULL;
469         int tx_cnt = 0;
470
471         spin_lock_bh(&ioat->desc_lock);
472         desc = ioat1_dma_get_next_descriptor(ioat);
473         do {
474                 if (!desc)
475                         break;
476
477                 tx_cnt++;
478                 copy = min_t(size_t, len, ioat->xfercap);
479
480                 hw = desc->hw;
481                 hw->size = copy;
482                 hw->ctl = 0;
483                 hw->src_addr = src;
484                 hw->dst_addr = dest;
485
486                 list_add_tail(&desc->node, &chain);
487
488                 len -= copy;
489                 dest += copy;
490                 src += copy;
491                 if (len) {
492                         struct ioat_desc_sw *next;
493
494                         async_tx_ack(&desc->txd);
495                         next = ioat1_dma_get_next_descriptor(ioat);
496                         hw->next = next ? next->txd.phys : 0;
497                         dump_desc_dbg(ioat, desc);
498                         desc = next;
499                 } else
500                         hw->next = 0;
501         } while (len);
502
503         if (!desc) {
504                 struct ioat_chan_common *chan = &ioat->base;
505
506                 dev_err(to_dev(chan),
507                         "chan%d - get_next_desc failed\n", chan_num(chan));
508                 list_splice(&chain, &ioat->free_desc);
509                 spin_unlock_bh(&ioat->desc_lock);
510                 return NULL;
511         }
512         spin_unlock_bh(&ioat->desc_lock);
513
514         desc->txd.flags = flags;
515         desc->len = total_len;
516         list_splice(&chain, &desc->tx_list);
517         hw->ctl_f.int_en = !!(flags & DMA_PREP_INTERRUPT);
518         hw->ctl_f.compl_write = 1;
519         hw->tx_cnt = tx_cnt;
520         dump_desc_dbg(ioat, desc);
521
522         return &desc->txd;
523 }
524
525 static void ioat1_cleanup_event(unsigned long data)
526 {
527         struct ioat_dma_chan *ioat = to_ioat_chan((void *) data);
528
529         ioat1_cleanup(ioat);
530         writew(IOAT_CHANCTRL_RUN, ioat->base.reg_base + IOAT_CHANCTRL_OFFSET);
531 }
532
533 void ioat_dma_unmap(struct ioat_chan_common *chan, enum dma_ctrl_flags flags,
534                     size_t len, struct ioat_dma_descriptor *hw)
535 {
536         struct pci_dev *pdev = chan->device->pdev;
537         size_t offset = len - hw->size;
538
539         if (!(flags & DMA_COMPL_SKIP_DEST_UNMAP))
540                 ioat_unmap(pdev, hw->dst_addr - offset, len,
541                            PCI_DMA_FROMDEVICE, flags, 1);
542
543         if (!(flags & DMA_COMPL_SKIP_SRC_UNMAP))
544                 ioat_unmap(pdev, hw->src_addr - offset, len,
545                            PCI_DMA_TODEVICE, flags, 0);
546 }
547
548 unsigned long ioat_get_current_completion(struct ioat_chan_common *chan)
549 {
550         unsigned long phys_complete;
551         u64 completion;
552
553         completion = *chan->completion;
554         phys_complete = ioat_chansts_to_addr(completion);
555
556         dev_dbg(to_dev(chan), "%s: phys_complete: %#llx\n", __func__,
557                 (unsigned long long) phys_complete);
558
559         if (is_ioat_halted(completion)) {
560                 u32 chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
561                 dev_err(to_dev(chan), "Channel halted, chanerr = %x\n",
562                         chanerr);
563
564                 /* TODO do something to salvage the situation */
565         }
566
567         return phys_complete;
568 }
569
570 bool ioat_cleanup_preamble(struct ioat_chan_common *chan,
571                            unsigned long *phys_complete)
572 {
573         *phys_complete = ioat_get_current_completion(chan);
574         if (*phys_complete == chan->last_completion)
575                 return false;
576         clear_bit(IOAT_COMPLETION_ACK, &chan->state);
577         mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
578
579         return true;
580 }
581
582 static void __cleanup(struct ioat_dma_chan *ioat, unsigned long phys_complete)
583 {
584         struct ioat_chan_common *chan = &ioat->base;
585         struct list_head *_desc, *n;
586         struct dma_async_tx_descriptor *tx;
587
588         dev_dbg(to_dev(chan), "%s: phys_complete: %lx\n",
589                  __func__, phys_complete);
590         list_for_each_safe(_desc, n, &ioat->used_desc) {
591                 struct ioat_desc_sw *desc;
592
593                 prefetch(n);
594                 desc = list_entry(_desc, typeof(*desc), node);
595                 tx = &desc->txd;
596                 /*
597                  * Incoming DMA requests may use multiple descriptors,
598                  * due to exceeding xfercap, perhaps. If so, only the
599                  * last one will have a cookie, and require unmapping.
600                  */
601                 dump_desc_dbg(ioat, desc);
602                 if (tx->cookie) {
603                         dma_cookie_complete(tx);
604                         ioat_dma_unmap(chan, tx->flags, desc->len, desc->hw);
605                         ioat->active -= desc->hw->tx_cnt;
606                         if (tx->callback) {
607                                 tx->callback(tx->callback_param);
608                                 tx->callback = NULL;
609                         }
610                 }
611
612                 if (tx->phys != phys_complete) {
613                         /*
614                          * a completed entry, but not the last, so clean
615                          * up if the client is done with the descriptor
616                          */
617                         if (async_tx_test_ack(tx))
618                                 list_move_tail(&desc->node, &ioat->free_desc);
619                 } else {
620                         /*
621                          * last used desc. Do not remove, so we can
622                          * append from it.
623                          */
624
625                         /* if nothing else is pending, cancel the
626                          * completion timeout
627                          */
628                         if (n == &ioat->used_desc) {
629                                 dev_dbg(to_dev(chan),
630                                         "%s cancel completion timeout\n",
631                                         __func__);
632                                 clear_bit(IOAT_COMPLETION_PENDING, &chan->state);
633                         }
634
635                         /* TODO check status bits? */
636                         break;
637                 }
638         }
639
640         chan->last_completion = phys_complete;
641 }
642
643 /**
644  * ioat1_cleanup - cleanup up finished descriptors
645  * @chan: ioat channel to be cleaned up
646  *
647  * To prevent lock contention we defer cleanup when the locks are
648  * contended with a terminal timeout that forces cleanup and catches
649  * completion notification errors.
650  */
651 static void ioat1_cleanup(struct ioat_dma_chan *ioat)
652 {
653         struct ioat_chan_common *chan = &ioat->base;
654         unsigned long phys_complete;
655
656         prefetch(chan->completion);
657
658         if (!spin_trylock_bh(&chan->cleanup_lock))
659                 return;
660
661         if (!ioat_cleanup_preamble(chan, &phys_complete)) {
662                 spin_unlock_bh(&chan->cleanup_lock);
663                 return;
664         }
665
666         if (!spin_trylock_bh(&ioat->desc_lock)) {
667                 spin_unlock_bh(&chan->cleanup_lock);
668                 return;
669         }
670
671         __cleanup(ioat, phys_complete);
672
673         spin_unlock_bh(&ioat->desc_lock);
674         spin_unlock_bh(&chan->cleanup_lock);
675 }
676
677 static void ioat1_timer_event(unsigned long data)
678 {
679         struct ioat_dma_chan *ioat = to_ioat_chan((void *) data);
680         struct ioat_chan_common *chan = &ioat->base;
681
682         dev_dbg(to_dev(chan), "%s: state: %lx\n", __func__, chan->state);
683
684         spin_lock_bh(&chan->cleanup_lock);
685         if (test_and_clear_bit(IOAT_RESET_PENDING, &chan->state)) {
686                 struct ioat_desc_sw *desc;
687
688                 spin_lock_bh(&ioat->desc_lock);
689
690                 /* restart active descriptors */
691                 desc = to_ioat_desc(ioat->used_desc.prev);
692                 ioat_set_chainaddr(ioat, desc->txd.phys);
693                 ioat_start(chan);
694
695                 ioat->pending = 0;
696                 set_bit(IOAT_COMPLETION_PENDING, &chan->state);
697                 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
698                 spin_unlock_bh(&ioat->desc_lock);
699         } else if (test_bit(IOAT_COMPLETION_PENDING, &chan->state)) {
700                 unsigned long phys_complete;
701
702                 spin_lock_bh(&ioat->desc_lock);
703                 /* if we haven't made progress and we have already
704                  * acknowledged a pending completion once, then be more
705                  * forceful with a restart
706                  */
707                 if (ioat_cleanup_preamble(chan, &phys_complete))
708                         __cleanup(ioat, phys_complete);
709                 else if (test_bit(IOAT_COMPLETION_ACK, &chan->state))
710                         ioat1_reset_channel(ioat);
711                 else {
712                         u64 status = ioat_chansts(chan);
713
714                         /* manually update the last completion address */
715                         if (ioat_chansts_to_addr(status) != 0)
716                                 *chan->completion = status;
717
718                         set_bit(IOAT_COMPLETION_ACK, &chan->state);
719                         mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
720                 }
721                 spin_unlock_bh(&ioat->desc_lock);
722         }
723         spin_unlock_bh(&chan->cleanup_lock);
724 }
725
726 enum dma_status
727 ioat_dma_tx_status(struct dma_chan *c, dma_cookie_t cookie,
728                    struct dma_tx_state *txstate)
729 {
730         struct ioat_chan_common *chan = to_chan_common(c);
731         struct ioatdma_device *device = chan->device;
732         enum dma_status ret;
733
734         ret = dma_cookie_status(c, cookie, txstate);
735         if (ret == DMA_SUCCESS)
736                 return ret;
737
738         device->cleanup_fn((unsigned long) c);
739
740         return dma_cookie_status(c, cookie, txstate);
741 }
742
743 static void ioat1_dma_start_null_desc(struct ioat_dma_chan *ioat)
744 {
745         struct ioat_chan_common *chan = &ioat->base;
746         struct ioat_desc_sw *desc;
747         struct ioat_dma_descriptor *hw;
748
749         spin_lock_bh(&ioat->desc_lock);
750
751         desc = ioat1_dma_get_next_descriptor(ioat);
752
753         if (!desc) {
754                 dev_err(to_dev(chan),
755                         "Unable to start null desc - get next desc failed\n");
756                 spin_unlock_bh(&ioat->desc_lock);
757                 return;
758         }
759
760         hw = desc->hw;
761         hw->ctl = 0;
762         hw->ctl_f.null = 1;
763         hw->ctl_f.int_en = 1;
764         hw->ctl_f.compl_write = 1;
765         /* set size to non-zero value (channel returns error when size is 0) */
766         hw->size = NULL_DESC_BUFFER_SIZE;
767         hw->src_addr = 0;
768         hw->dst_addr = 0;
769         async_tx_ack(&desc->txd);
770         hw->next = 0;
771         list_add_tail(&desc->node, &ioat->used_desc);
772         dump_desc_dbg(ioat, desc);
773
774         ioat_set_chainaddr(ioat, desc->txd.phys);
775         ioat_start(chan);
776         spin_unlock_bh(&ioat->desc_lock);
777 }
778
779 /*
780  * Perform a IOAT transaction to verify the HW works.
781  */
782 #define IOAT_TEST_SIZE 2000
783
784 static void __devinit ioat_dma_test_callback(void *dma_async_param)
785 {
786         struct completion *cmp = dma_async_param;
787
788         complete(cmp);
789 }
790
791 /**
792  * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
793  * @device: device to be tested
794  */
795 int __devinit ioat_dma_self_test(struct ioatdma_device *device)
796 {
797         int i;
798         u8 *src;
799         u8 *dest;
800         struct dma_device *dma = &device->common;
801         struct device *dev = &device->pdev->dev;
802         struct dma_chan *dma_chan;
803         struct dma_async_tx_descriptor *tx;
804         dma_addr_t dma_dest, dma_src;
805         dma_cookie_t cookie;
806         int err = 0;
807         struct completion cmp;
808         unsigned long tmo;
809         unsigned long flags;
810
811         src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
812         if (!src)
813                 return -ENOMEM;
814         dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
815         if (!dest) {
816                 kfree(src);
817                 return -ENOMEM;
818         }
819
820         /* Fill in src buffer */
821         for (i = 0; i < IOAT_TEST_SIZE; i++)
822                 src[i] = (u8)i;
823
824         /* Start copy, using first DMA channel */
825         dma_chan = container_of(dma->channels.next, struct dma_chan,
826                                 device_node);
827         if (dma->device_alloc_chan_resources(dma_chan) < 1) {
828                 dev_err(dev, "selftest cannot allocate chan resource\n");
829                 err = -ENODEV;
830                 goto out;
831         }
832
833         dma_src = dma_map_single(dev, src, IOAT_TEST_SIZE, DMA_TO_DEVICE);
834         dma_dest = dma_map_single(dev, dest, IOAT_TEST_SIZE, DMA_FROM_DEVICE);
835         flags = DMA_COMPL_SRC_UNMAP_SINGLE | DMA_COMPL_DEST_UNMAP_SINGLE |
836                 DMA_PREP_INTERRUPT;
837         tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src,
838                                                    IOAT_TEST_SIZE, flags);
839         if (!tx) {
840                 dev_err(dev, "Self-test prep failed, disabling\n");
841                 err = -ENODEV;
842                 goto free_resources;
843         }
844
845         async_tx_ack(tx);
846         init_completion(&cmp);
847         tx->callback = ioat_dma_test_callback;
848         tx->callback_param = &cmp;
849         cookie = tx->tx_submit(tx);
850         if (cookie < 0) {
851                 dev_err(dev, "Self-test setup failed, disabling\n");
852                 err = -ENODEV;
853                 goto free_resources;
854         }
855         dma->device_issue_pending(dma_chan);
856
857         tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000));
858
859         if (tmo == 0 ||
860             dma->device_tx_status(dma_chan, cookie, NULL)
861                                         != DMA_SUCCESS) {
862                 dev_err(dev, "Self-test copy timed out, disabling\n");
863                 err = -ENODEV;
864                 goto free_resources;
865         }
866         if (memcmp(src, dest, IOAT_TEST_SIZE)) {
867                 dev_err(dev, "Self-test copy failed compare, disabling\n");
868                 err = -ENODEV;
869                 goto free_resources;
870         }
871
872 free_resources:
873         dma->device_free_chan_resources(dma_chan);
874 out:
875         kfree(src);
876         kfree(dest);
877         return err;
878 }
879
880 static char ioat_interrupt_style[32] = "msix";
881 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
882                     sizeof(ioat_interrupt_style), 0644);
883 MODULE_PARM_DESC(ioat_interrupt_style,
884                  "set ioat interrupt style: msix (default), "
885                  "msix-single-vector, msi, intx)");
886
887 /**
888  * ioat_dma_setup_interrupts - setup interrupt handler
889  * @device: ioat device
890  */
891 static int ioat_dma_setup_interrupts(struct ioatdma_device *device)
892 {
893         struct ioat_chan_common *chan;
894         struct pci_dev *pdev = device->pdev;
895         struct device *dev = &pdev->dev;
896         struct msix_entry *msix;
897         int i, j, msixcnt;
898         int err = -EINVAL;
899         u8 intrctrl = 0;
900
901         if (!strcmp(ioat_interrupt_style, "msix"))
902                 goto msix;
903         if (!strcmp(ioat_interrupt_style, "msix-single-vector"))
904                 goto msix_single_vector;
905         if (!strcmp(ioat_interrupt_style, "msi"))
906                 goto msi;
907         if (!strcmp(ioat_interrupt_style, "intx"))
908                 goto intx;
909         dev_err(dev, "invalid ioat_interrupt_style %s\n", ioat_interrupt_style);
910         goto err_no_irq;
911
912 msix:
913         /* The number of MSI-X vectors should equal the number of channels */
914         msixcnt = device->common.chancnt;
915         for (i = 0; i < msixcnt; i++)
916                 device->msix_entries[i].entry = i;
917
918         err = pci_enable_msix(pdev, device->msix_entries, msixcnt);
919         if (err < 0)
920                 goto msi;
921         if (err > 0)
922                 goto msix_single_vector;
923
924         for (i = 0; i < msixcnt; i++) {
925                 msix = &device->msix_entries[i];
926                 chan = ioat_chan_by_index(device, i);
927                 err = devm_request_irq(dev, msix->vector,
928                                        ioat_dma_do_interrupt_msix, 0,
929                                        "ioat-msix", chan);
930                 if (err) {
931                         for (j = 0; j < i; j++) {
932                                 msix = &device->msix_entries[j];
933                                 chan = ioat_chan_by_index(device, j);
934                                 devm_free_irq(dev, msix->vector, chan);
935                         }
936                         goto msix_single_vector;
937                 }
938         }
939         intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
940         goto done;
941
942 msix_single_vector:
943         msix = &device->msix_entries[0];
944         msix->entry = 0;
945         err = pci_enable_msix(pdev, device->msix_entries, 1);
946         if (err)
947                 goto msi;
948
949         err = devm_request_irq(dev, msix->vector, ioat_dma_do_interrupt, 0,
950                                "ioat-msix", device);
951         if (err) {
952                 pci_disable_msix(pdev);
953                 goto msi;
954         }
955         goto done;
956
957 msi:
958         err = pci_enable_msi(pdev);
959         if (err)
960                 goto intx;
961
962         err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt, 0,
963                                "ioat-msi", device);
964         if (err) {
965                 pci_disable_msi(pdev);
966                 goto intx;
967         }
968         goto done;
969
970 intx:
971         err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt,
972                                IRQF_SHARED, "ioat-intx", device);
973         if (err)
974                 goto err_no_irq;
975
976 done:
977         if (device->intr_quirk)
978                 device->intr_quirk(device);
979         intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
980         writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
981         return 0;
982
983 err_no_irq:
984         /* Disable all interrupt generation */
985         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
986         dev_err(dev, "no usable interrupts\n");
987         return err;
988 }
989
990 static void ioat_disable_interrupts(struct ioatdma_device *device)
991 {
992         /* Disable all interrupt generation */
993         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
994 }
995
996 int __devinit ioat_probe(struct ioatdma_device *device)
997 {
998         int err = -ENODEV;
999         struct dma_device *dma = &device->common;
1000         struct pci_dev *pdev = device->pdev;
1001         struct device *dev = &pdev->dev;
1002
1003         /* DMA coherent memory pool for DMA descriptor allocations */
1004         device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
1005                                            sizeof(struct ioat_dma_descriptor),
1006                                            64, 0);
1007         if (!device->dma_pool) {
1008                 err = -ENOMEM;
1009                 goto err_dma_pool;
1010         }
1011
1012         device->completion_pool = pci_pool_create("completion_pool", pdev,
1013                                                   sizeof(u64), SMP_CACHE_BYTES,
1014                                                   SMP_CACHE_BYTES);
1015
1016         if (!device->completion_pool) {
1017                 err = -ENOMEM;
1018                 goto err_completion_pool;
1019         }
1020
1021         device->enumerate_channels(device);
1022
1023         dma_cap_set(DMA_MEMCPY, dma->cap_mask);
1024         dma->dev = &pdev->dev;
1025
1026         if (!dma->chancnt) {
1027                 dev_err(dev, "channel enumeration error\n");
1028                 goto err_setup_interrupts;
1029         }
1030
1031         err = ioat_dma_setup_interrupts(device);
1032         if (err)
1033                 goto err_setup_interrupts;
1034
1035         err = device->self_test(device);
1036         if (err)
1037                 goto err_self_test;
1038
1039         return 0;
1040
1041 err_self_test:
1042         ioat_disable_interrupts(device);
1043 err_setup_interrupts:
1044         pci_pool_destroy(device->completion_pool);
1045 err_completion_pool:
1046         pci_pool_destroy(device->dma_pool);
1047 err_dma_pool:
1048         return err;
1049 }
1050
1051 int __devinit ioat_register(struct ioatdma_device *device)
1052 {
1053         int err = dma_async_device_register(&device->common);
1054
1055         if (err) {
1056                 ioat_disable_interrupts(device);
1057                 pci_pool_destroy(device->completion_pool);
1058                 pci_pool_destroy(device->dma_pool);
1059         }
1060
1061         return err;
1062 }
1063
1064 /* ioat1_intr_quirk - fix up dma ctrl register to enable / disable msi */
1065 static void ioat1_intr_quirk(struct ioatdma_device *device)
1066 {
1067         struct pci_dev *pdev = device->pdev;
1068         u32 dmactrl;
1069
1070         pci_read_config_dword(pdev, IOAT_PCI_DMACTRL_OFFSET, &dmactrl);
1071         if (pdev->msi_enabled)
1072                 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN;
1073         else
1074                 dmactrl &= ~IOAT_PCI_DMACTRL_MSI_EN;
1075         pci_write_config_dword(pdev, IOAT_PCI_DMACTRL_OFFSET, dmactrl);
1076 }
1077
1078 static ssize_t ring_size_show(struct dma_chan *c, char *page)
1079 {
1080         struct ioat_dma_chan *ioat = to_ioat_chan(c);
1081
1082         return sprintf(page, "%d\n", ioat->desccount);
1083 }
1084 static struct ioat_sysfs_entry ring_size_attr = __ATTR_RO(ring_size);
1085
1086 static ssize_t ring_active_show(struct dma_chan *c, char *page)
1087 {
1088         struct ioat_dma_chan *ioat = to_ioat_chan(c);
1089
1090         return sprintf(page, "%d\n", ioat->active);
1091 }
1092 static struct ioat_sysfs_entry ring_active_attr = __ATTR_RO(ring_active);
1093
1094 static ssize_t cap_show(struct dma_chan *c, char *page)
1095 {
1096         struct dma_device *dma = c->device;
1097
1098         return sprintf(page, "copy%s%s%s%s%s%s\n",
1099                        dma_has_cap(DMA_PQ, dma->cap_mask) ? " pq" : "",
1100                        dma_has_cap(DMA_PQ_VAL, dma->cap_mask) ? " pq_val" : "",
1101                        dma_has_cap(DMA_XOR, dma->cap_mask) ? " xor" : "",
1102                        dma_has_cap(DMA_XOR_VAL, dma->cap_mask) ? " xor_val" : "",
1103                        dma_has_cap(DMA_MEMSET, dma->cap_mask)  ? " fill" : "",
1104                        dma_has_cap(DMA_INTERRUPT, dma->cap_mask) ? " intr" : "");
1105
1106 }
1107 struct ioat_sysfs_entry ioat_cap_attr = __ATTR_RO(cap);
1108
1109 static ssize_t version_show(struct dma_chan *c, char *page)
1110 {
1111         struct dma_device *dma = c->device;
1112         struct ioatdma_device *device = to_ioatdma_device(dma);
1113
1114         return sprintf(page, "%d.%d\n",
1115                        device->version >> 4, device->version & 0xf);
1116 }
1117 struct ioat_sysfs_entry ioat_version_attr = __ATTR_RO(version);
1118
1119 static struct attribute *ioat1_attrs[] = {
1120         &ring_size_attr.attr,
1121         &ring_active_attr.attr,
1122         &ioat_cap_attr.attr,
1123         &ioat_version_attr.attr,
1124         NULL,
1125 };
1126
1127 static ssize_t
1128 ioat_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
1129 {
1130         struct ioat_sysfs_entry *entry;
1131         struct ioat_chan_common *chan;
1132
1133         entry = container_of(attr, struct ioat_sysfs_entry, attr);
1134         chan = container_of(kobj, struct ioat_chan_common, kobj);
1135
1136         if (!entry->show)
1137                 return -EIO;
1138         return entry->show(&chan->common, page);
1139 }
1140
1141 const struct sysfs_ops ioat_sysfs_ops = {
1142         .show   = ioat_attr_show,
1143 };
1144
1145 static struct kobj_type ioat1_ktype = {
1146         .sysfs_ops = &ioat_sysfs_ops,
1147         .default_attrs = ioat1_attrs,
1148 };
1149
1150 void ioat_kobject_add(struct ioatdma_device *device, struct kobj_type *type)
1151 {
1152         struct dma_device *dma = &device->common;
1153         struct dma_chan *c;
1154
1155         list_for_each_entry(c, &dma->channels, device_node) {
1156                 struct ioat_chan_common *chan = to_chan_common(c);
1157                 struct kobject *parent = &c->dev->device.kobj;
1158                 int err;
1159
1160                 err = kobject_init_and_add(&chan->kobj, type, parent, "quickdata");
1161                 if (err) {
1162                         dev_warn(to_dev(chan),
1163                                  "sysfs init error (%d), continuing...\n", err);
1164                         kobject_put(&chan->kobj);
1165                         set_bit(IOAT_KOBJ_INIT_FAIL, &chan->state);
1166                 }
1167         }
1168 }
1169
1170 void ioat_kobject_del(struct ioatdma_device *device)
1171 {
1172         struct dma_device *dma = &device->common;
1173         struct dma_chan *c;
1174
1175         list_for_each_entry(c, &dma->channels, device_node) {
1176                 struct ioat_chan_common *chan = to_chan_common(c);
1177
1178                 if (!test_bit(IOAT_KOBJ_INIT_FAIL, &chan->state)) {
1179                         kobject_del(&chan->kobj);
1180                         kobject_put(&chan->kobj);
1181                 }
1182         }
1183 }
1184
1185 int __devinit ioat1_dma_probe(struct ioatdma_device *device, int dca)
1186 {
1187         struct pci_dev *pdev = device->pdev;
1188         struct dma_device *dma;
1189         int err;
1190
1191         device->intr_quirk = ioat1_intr_quirk;
1192         device->enumerate_channels = ioat1_enumerate_channels;
1193         device->self_test = ioat_dma_self_test;
1194         device->timer_fn = ioat1_timer_event;
1195         device->cleanup_fn = ioat1_cleanup_event;
1196         dma = &device->common;
1197         dma->device_prep_dma_memcpy = ioat1_dma_prep_memcpy;
1198         dma->device_issue_pending = ioat1_dma_memcpy_issue_pending;
1199         dma->device_alloc_chan_resources = ioat1_dma_alloc_chan_resources;
1200         dma->device_free_chan_resources = ioat1_dma_free_chan_resources;
1201         dma->device_tx_status = ioat_dma_tx_status;
1202
1203         err = ioat_probe(device);
1204         if (err)
1205                 return err;
1206         ioat_set_tcp_copy_break(4096);
1207         err = ioat_register(device);
1208         if (err)
1209                 return err;
1210         ioat_kobject_add(device, &ioat1_ktype);
1211
1212         if (dca)
1213                 device->dca = ioat_dca_init(pdev, device->reg_base);
1214
1215         return err;
1216 }
1217
1218 void __devexit ioat_dma_remove(struct ioatdma_device *device)
1219 {
1220         struct dma_device *dma = &device->common;
1221
1222         ioat_disable_interrupts(device);
1223
1224         ioat_kobject_del(device);
1225
1226         dma_async_device_unregister(dma);
1227
1228         pci_pool_destroy(device->dma_pool);
1229         pci_pool_destroy(device->completion_pool);
1230
1231         INIT_LIST_HEAD(&dma->channels);
1232 }