Merge git://git.kernel.org/pub/scm/linux/kernel/git/hirofumi/fatfs-2.6
[pandora-kernel.git] / drivers / dma / ioat / dma_v2.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 (versions >= 2), which
25  * does asynchronous data movement and checksumming 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/i7300_idle.h>
38 #include "dma.h"
39 #include "dma_v2.h"
40 #include "registers.h"
41 #include "hw.h"
42
43 int ioat_ring_alloc_order = 8;
44 module_param(ioat_ring_alloc_order, int, 0644);
45 MODULE_PARM_DESC(ioat_ring_alloc_order,
46                  "ioat2+: allocate 2^n descriptors per channel"
47                  " (default: 8 max: 16)");
48 static int ioat_ring_max_alloc_order = IOAT_MAX_ORDER;
49 module_param(ioat_ring_max_alloc_order, int, 0644);
50 MODULE_PARM_DESC(ioat_ring_max_alloc_order,
51                  "ioat2+: upper limit for ring size (default: 16)");
52
53 void __ioat2_issue_pending(struct ioat2_dma_chan *ioat)
54 {
55         struct ioat_chan_common *chan = &ioat->base;
56
57         ioat->dmacount += ioat2_ring_pending(ioat);
58         ioat->issued = ioat->head;
59         writew(ioat->dmacount, chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
60         dev_dbg(to_dev(chan),
61                 "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
62                 __func__, ioat->head, ioat->tail, ioat->issued, ioat->dmacount);
63 }
64
65 void ioat2_issue_pending(struct dma_chan *c)
66 {
67         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
68
69         if (ioat2_ring_pending(ioat)) {
70                 spin_lock_bh(&ioat->prep_lock);
71                 __ioat2_issue_pending(ioat);
72                 spin_unlock_bh(&ioat->prep_lock);
73         }
74 }
75
76 /**
77  * ioat2_update_pending - log pending descriptors
78  * @ioat: ioat2+ channel
79  *
80  * Check if the number of unsubmitted descriptors has exceeded the
81  * watermark.  Called with prep_lock held
82  */
83 static void ioat2_update_pending(struct ioat2_dma_chan *ioat)
84 {
85         if (ioat2_ring_pending(ioat) > ioat_pending_level)
86                 __ioat2_issue_pending(ioat);
87 }
88
89 static void __ioat2_start_null_desc(struct ioat2_dma_chan *ioat)
90 {
91         struct ioat_ring_ent *desc;
92         struct ioat_dma_descriptor *hw;
93
94         if (ioat2_ring_space(ioat) < 1) {
95                 dev_err(to_dev(&ioat->base),
96                         "Unable to start null desc - ring full\n");
97                 return;
98         }
99
100         dev_dbg(to_dev(&ioat->base), "%s: head: %#x tail: %#x issued: %#x\n",
101                 __func__, ioat->head, ioat->tail, ioat->issued);
102         desc = ioat2_get_ring_ent(ioat, ioat->head);
103
104         hw = desc->hw;
105         hw->ctl = 0;
106         hw->ctl_f.null = 1;
107         hw->ctl_f.int_en = 1;
108         hw->ctl_f.compl_write = 1;
109         /* set size to non-zero value (channel returns error when size is 0) */
110         hw->size = NULL_DESC_BUFFER_SIZE;
111         hw->src_addr = 0;
112         hw->dst_addr = 0;
113         async_tx_ack(&desc->txd);
114         ioat2_set_chainaddr(ioat, desc->txd.phys);
115         dump_desc_dbg(ioat, desc);
116         wmb();
117         ioat->head += 1;
118         __ioat2_issue_pending(ioat);
119 }
120
121 static void ioat2_start_null_desc(struct ioat2_dma_chan *ioat)
122 {
123         spin_lock_bh(&ioat->prep_lock);
124         __ioat2_start_null_desc(ioat);
125         spin_unlock_bh(&ioat->prep_lock);
126 }
127
128 static void __cleanup(struct ioat2_dma_chan *ioat, unsigned long phys_complete)
129 {
130         struct ioat_chan_common *chan = &ioat->base;
131         struct dma_async_tx_descriptor *tx;
132         struct ioat_ring_ent *desc;
133         bool seen_current = false;
134         u16 active;
135         int idx = ioat->tail, i;
136
137         dev_dbg(to_dev(chan), "%s: head: %#x tail: %#x issued: %#x\n",
138                 __func__, ioat->head, ioat->tail, ioat->issued);
139
140         active = ioat2_ring_active(ioat);
141         for (i = 0; i < active && !seen_current; i++) {
142                 smp_read_barrier_depends();
143                 prefetch(ioat2_get_ring_ent(ioat, idx + i + 1));
144                 desc = ioat2_get_ring_ent(ioat, idx + i);
145                 tx = &desc->txd;
146                 dump_desc_dbg(ioat, desc);
147                 if (tx->cookie) {
148                         ioat_dma_unmap(chan, tx->flags, desc->len, desc->hw);
149                         chan->completed_cookie = tx->cookie;
150                         tx->cookie = 0;
151                         if (tx->callback) {
152                                 tx->callback(tx->callback_param);
153                                 tx->callback = NULL;
154                         }
155                 }
156
157                 if (tx->phys == phys_complete)
158                         seen_current = true;
159         }
160         smp_mb(); /* finish all descriptor reads before incrementing tail */
161         ioat->tail = idx + i;
162         BUG_ON(active && !seen_current); /* no active descs have written a completion? */
163
164         chan->last_completion = phys_complete;
165         if (active - i == 0) {
166                 dev_dbg(to_dev(chan), "%s: cancel completion timeout\n",
167                         __func__);
168                 clear_bit(IOAT_COMPLETION_PENDING, &chan->state);
169                 mod_timer(&chan->timer, jiffies + IDLE_TIMEOUT);
170         }
171 }
172
173 /**
174  * ioat2_cleanup - clean finished descriptors (advance tail pointer)
175  * @chan: ioat channel to be cleaned up
176  */
177 static void ioat2_cleanup(struct ioat2_dma_chan *ioat)
178 {
179         struct ioat_chan_common *chan = &ioat->base;
180         unsigned long phys_complete;
181
182         spin_lock_bh(&chan->cleanup_lock);
183         if (ioat_cleanup_preamble(chan, &phys_complete))
184                 __cleanup(ioat, phys_complete);
185         spin_unlock_bh(&chan->cleanup_lock);
186 }
187
188 void ioat2_cleanup_event(unsigned long data)
189 {
190         struct ioat2_dma_chan *ioat = to_ioat2_chan((void *) data);
191
192         ioat2_cleanup(ioat);
193         writew(IOAT_CHANCTRL_RUN, ioat->base.reg_base + IOAT_CHANCTRL_OFFSET);
194 }
195
196 void __ioat2_restart_chan(struct ioat2_dma_chan *ioat)
197 {
198         struct ioat_chan_common *chan = &ioat->base;
199
200         /* set the tail to be re-issued */
201         ioat->issued = ioat->tail;
202         ioat->dmacount = 0;
203         set_bit(IOAT_COMPLETION_PENDING, &chan->state);
204         mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
205
206         dev_dbg(to_dev(chan),
207                 "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
208                 __func__, ioat->head, ioat->tail, ioat->issued, ioat->dmacount);
209
210         if (ioat2_ring_pending(ioat)) {
211                 struct ioat_ring_ent *desc;
212
213                 desc = ioat2_get_ring_ent(ioat, ioat->tail);
214                 ioat2_set_chainaddr(ioat, desc->txd.phys);
215                 __ioat2_issue_pending(ioat);
216         } else
217                 __ioat2_start_null_desc(ioat);
218 }
219
220 int ioat2_quiesce(struct ioat_chan_common *chan, unsigned long tmo)
221 {
222         unsigned long end = jiffies + tmo;
223         int err = 0;
224         u32 status;
225
226         status = ioat_chansts(chan);
227         if (is_ioat_active(status) || is_ioat_idle(status))
228                 ioat_suspend(chan);
229         while (is_ioat_active(status) || is_ioat_idle(status)) {
230                 if (tmo && time_after(jiffies, end)) {
231                         err = -ETIMEDOUT;
232                         break;
233                 }
234                 status = ioat_chansts(chan);
235                 cpu_relax();
236         }
237
238         return err;
239 }
240
241 int ioat2_reset_sync(struct ioat_chan_common *chan, unsigned long tmo)
242 {
243         unsigned long end = jiffies + tmo;
244         int err = 0;
245
246         ioat_reset(chan);
247         while (ioat_reset_pending(chan)) {
248                 if (end && time_after(jiffies, end)) {
249                         err = -ETIMEDOUT;
250                         break;
251                 }
252                 cpu_relax();
253         }
254
255         return err;
256 }
257
258 static void ioat2_restart_channel(struct ioat2_dma_chan *ioat)
259 {
260         struct ioat_chan_common *chan = &ioat->base;
261         unsigned long phys_complete;
262
263         ioat2_quiesce(chan, 0);
264         if (ioat_cleanup_preamble(chan, &phys_complete))
265                 __cleanup(ioat, phys_complete);
266
267         __ioat2_restart_chan(ioat);
268 }
269
270 void ioat2_timer_event(unsigned long data)
271 {
272         struct ioat2_dma_chan *ioat = to_ioat2_chan((void *) data);
273         struct ioat_chan_common *chan = &ioat->base;
274
275         if (test_bit(IOAT_COMPLETION_PENDING, &chan->state)) {
276                 unsigned long phys_complete;
277                 u64 status;
278
279                 status = ioat_chansts(chan);
280
281                 /* when halted due to errors check for channel
282                  * programming errors before advancing the completion state
283                  */
284                 if (is_ioat_halted(status)) {
285                         u32 chanerr;
286
287                         chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
288                         dev_err(to_dev(chan), "%s: Channel halted (%x)\n",
289                                 __func__, chanerr);
290                         BUG_ON(is_ioat_bug(chanerr));
291                 }
292
293                 /* if we haven't made progress and we have already
294                  * acknowledged a pending completion once, then be more
295                  * forceful with a restart
296                  */
297                 spin_lock_bh(&chan->cleanup_lock);
298                 if (ioat_cleanup_preamble(chan, &phys_complete)) {
299                         __cleanup(ioat, phys_complete);
300                 } else if (test_bit(IOAT_COMPLETION_ACK, &chan->state)) {
301                         spin_lock_bh(&ioat->prep_lock);
302                         ioat2_restart_channel(ioat);
303                         spin_unlock_bh(&ioat->prep_lock);
304                 } else {
305                         set_bit(IOAT_COMPLETION_ACK, &chan->state);
306                         mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
307                 }
308                 spin_unlock_bh(&chan->cleanup_lock);
309         } else {
310                 u16 active;
311
312                 /* if the ring is idle, empty, and oversized try to step
313                  * down the size
314                  */
315                 spin_lock_bh(&chan->cleanup_lock);
316                 spin_lock_bh(&ioat->prep_lock);
317                 active = ioat2_ring_active(ioat);
318                 if (active == 0 && ioat->alloc_order > ioat_get_alloc_order())
319                         reshape_ring(ioat, ioat->alloc_order-1);
320                 spin_unlock_bh(&ioat->prep_lock);
321                 spin_unlock_bh(&chan->cleanup_lock);
322
323                 /* keep shrinking until we get back to our minimum
324                  * default size
325                  */
326                 if (ioat->alloc_order > ioat_get_alloc_order())
327                         mod_timer(&chan->timer, jiffies + IDLE_TIMEOUT);
328         }
329 }
330
331 static int ioat2_reset_hw(struct ioat_chan_common *chan)
332 {
333         /* throw away whatever the channel was doing and get it initialized */
334         u32 chanerr;
335
336         ioat2_quiesce(chan, msecs_to_jiffies(100));
337
338         chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
339         writel(chanerr, chan->reg_base + IOAT_CHANERR_OFFSET);
340
341         return ioat2_reset_sync(chan, msecs_to_jiffies(200));
342 }
343
344 /**
345  * ioat2_enumerate_channels - find and initialize the device's channels
346  * @device: the device to be enumerated
347  */
348 int ioat2_enumerate_channels(struct ioatdma_device *device)
349 {
350         struct ioat2_dma_chan *ioat;
351         struct device *dev = &device->pdev->dev;
352         struct dma_device *dma = &device->common;
353         u8 xfercap_log;
354         int i;
355
356         INIT_LIST_HEAD(&dma->channels);
357         dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
358         dma->chancnt &= 0x1f; /* bits [4:0] valid */
359         if (dma->chancnt > ARRAY_SIZE(device->idx)) {
360                 dev_warn(dev, "(%d) exceeds max supported channels (%zu)\n",
361                          dma->chancnt, ARRAY_SIZE(device->idx));
362                 dma->chancnt = ARRAY_SIZE(device->idx);
363         }
364         xfercap_log = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
365         xfercap_log &= 0x1f; /* bits [4:0] valid */
366         if (xfercap_log == 0)
367                 return 0;
368         dev_dbg(dev, "%s: xfercap = %d\n", __func__, 1 << xfercap_log);
369
370         /* FIXME which i/oat version is i7300? */
371 #ifdef CONFIG_I7300_IDLE_IOAT_CHANNEL
372         if (i7300_idle_platform_probe(NULL, NULL, 1) == 0)
373                 dma->chancnt--;
374 #endif
375         for (i = 0; i < dma->chancnt; i++) {
376                 ioat = devm_kzalloc(dev, sizeof(*ioat), GFP_KERNEL);
377                 if (!ioat)
378                         break;
379
380                 ioat_init_channel(device, &ioat->base, i);
381                 ioat->xfercap_log = xfercap_log;
382                 spin_lock_init(&ioat->prep_lock);
383                 if (device->reset_hw(&ioat->base)) {
384                         i = 0;
385                         break;
386                 }
387         }
388         dma->chancnt = i;
389         return i;
390 }
391
392 static dma_cookie_t ioat2_tx_submit_unlock(struct dma_async_tx_descriptor *tx)
393 {
394         struct dma_chan *c = tx->chan;
395         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
396         struct ioat_chan_common *chan = &ioat->base;
397         dma_cookie_t cookie = c->cookie;
398
399         cookie++;
400         if (cookie < 0)
401                 cookie = 1;
402         tx->cookie = cookie;
403         c->cookie = cookie;
404         dev_dbg(to_dev(&ioat->base), "%s: cookie: %d\n", __func__, cookie);
405
406         if (!test_and_set_bit(IOAT_COMPLETION_PENDING, &chan->state))
407                 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
408
409         /* make descriptor updates visible before advancing ioat->head,
410          * this is purposefully not smp_wmb() since we are also
411          * publishing the descriptor updates to a dma device
412          */
413         wmb();
414
415         ioat->head += ioat->produce;
416
417         ioat2_update_pending(ioat);
418         spin_unlock_bh(&ioat->prep_lock);
419
420         return cookie;
421 }
422
423 static struct ioat_ring_ent *ioat2_alloc_ring_ent(struct dma_chan *chan, gfp_t flags)
424 {
425         struct ioat_dma_descriptor *hw;
426         struct ioat_ring_ent *desc;
427         struct ioatdma_device *dma;
428         dma_addr_t phys;
429
430         dma = to_ioatdma_device(chan->device);
431         hw = pci_pool_alloc(dma->dma_pool, flags, &phys);
432         if (!hw)
433                 return NULL;
434         memset(hw, 0, sizeof(*hw));
435
436         desc = kmem_cache_alloc(ioat2_cache, flags);
437         if (!desc) {
438                 pci_pool_free(dma->dma_pool, hw, phys);
439                 return NULL;
440         }
441         memset(desc, 0, sizeof(*desc));
442
443         dma_async_tx_descriptor_init(&desc->txd, chan);
444         desc->txd.tx_submit = ioat2_tx_submit_unlock;
445         desc->hw = hw;
446         desc->txd.phys = phys;
447         return desc;
448 }
449
450 static void ioat2_free_ring_ent(struct ioat_ring_ent *desc, struct dma_chan *chan)
451 {
452         struct ioatdma_device *dma;
453
454         dma = to_ioatdma_device(chan->device);
455         pci_pool_free(dma->dma_pool, desc->hw, desc->txd.phys);
456         kmem_cache_free(ioat2_cache, desc);
457 }
458
459 static struct ioat_ring_ent **ioat2_alloc_ring(struct dma_chan *c, int order, gfp_t flags)
460 {
461         struct ioat_ring_ent **ring;
462         int descs = 1 << order;
463         int i;
464
465         if (order > ioat_get_max_alloc_order())
466                 return NULL;
467
468         /* allocate the array to hold the software ring */
469         ring = kcalloc(descs, sizeof(*ring), flags);
470         if (!ring)
471                 return NULL;
472         for (i = 0; i < descs; i++) {
473                 ring[i] = ioat2_alloc_ring_ent(c, flags);
474                 if (!ring[i]) {
475                         while (i--)
476                                 ioat2_free_ring_ent(ring[i], c);
477                         kfree(ring);
478                         return NULL;
479                 }
480                 set_desc_id(ring[i], i);
481         }
482
483         /* link descs */
484         for (i = 0; i < descs-1; i++) {
485                 struct ioat_ring_ent *next = ring[i+1];
486                 struct ioat_dma_descriptor *hw = ring[i]->hw;
487
488                 hw->next = next->txd.phys;
489         }
490         ring[i]->hw->next = ring[0]->txd.phys;
491
492         return ring;
493 }
494
495 /* ioat2_alloc_chan_resources - allocate/initialize ioat2 descriptor ring
496  * @chan: channel to be initialized
497  */
498 int ioat2_alloc_chan_resources(struct dma_chan *c)
499 {
500         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
501         struct ioat_chan_common *chan = &ioat->base;
502         struct ioat_ring_ent **ring;
503         int order;
504
505         /* have we already been set up? */
506         if (ioat->ring)
507                 return 1 << ioat->alloc_order;
508
509         /* Setup register to interrupt and write completion status on error */
510         writew(IOAT_CHANCTRL_RUN, chan->reg_base + IOAT_CHANCTRL_OFFSET);
511
512         /* allocate a completion writeback area */
513         /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
514         chan->completion = pci_pool_alloc(chan->device->completion_pool,
515                                           GFP_KERNEL, &chan->completion_dma);
516         if (!chan->completion)
517                 return -ENOMEM;
518
519         memset(chan->completion, 0, sizeof(*chan->completion));
520         writel(((u64) chan->completion_dma) & 0x00000000FFFFFFFF,
521                chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
522         writel(((u64) chan->completion_dma) >> 32,
523                chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
524
525         order = ioat_get_alloc_order();
526         ring = ioat2_alloc_ring(c, order, GFP_KERNEL);
527         if (!ring)
528                 return -ENOMEM;
529
530         spin_lock_bh(&chan->cleanup_lock);
531         spin_lock_bh(&ioat->prep_lock);
532         ioat->ring = ring;
533         ioat->head = 0;
534         ioat->issued = 0;
535         ioat->tail = 0;
536         ioat->alloc_order = order;
537         spin_unlock_bh(&ioat->prep_lock);
538         spin_unlock_bh(&chan->cleanup_lock);
539
540         tasklet_enable(&chan->cleanup_task);
541         ioat2_start_null_desc(ioat);
542
543         return 1 << ioat->alloc_order;
544 }
545
546 bool reshape_ring(struct ioat2_dma_chan *ioat, int order)
547 {
548         /* reshape differs from normal ring allocation in that we want
549          * to allocate a new software ring while only
550          * extending/truncating the hardware ring
551          */
552         struct ioat_chan_common *chan = &ioat->base;
553         struct dma_chan *c = &chan->common;
554         const u16 curr_size = ioat2_ring_size(ioat);
555         const u16 active = ioat2_ring_active(ioat);
556         const u16 new_size = 1 << order;
557         struct ioat_ring_ent **ring;
558         u16 i;
559
560         if (order > ioat_get_max_alloc_order())
561                 return false;
562
563         /* double check that we have at least 1 free descriptor */
564         if (active == curr_size)
565                 return false;
566
567         /* when shrinking, verify that we can hold the current active
568          * set in the new ring
569          */
570         if (active >= new_size)
571                 return false;
572
573         /* allocate the array to hold the software ring */
574         ring = kcalloc(new_size, sizeof(*ring), GFP_NOWAIT);
575         if (!ring)
576                 return false;
577
578         /* allocate/trim descriptors as needed */
579         if (new_size > curr_size) {
580                 /* copy current descriptors to the new ring */
581                 for (i = 0; i < curr_size; i++) {
582                         u16 curr_idx = (ioat->tail+i) & (curr_size-1);
583                         u16 new_idx = (ioat->tail+i) & (new_size-1);
584
585                         ring[new_idx] = ioat->ring[curr_idx];
586                         set_desc_id(ring[new_idx], new_idx);
587                 }
588
589                 /* add new descriptors to the ring */
590                 for (i = curr_size; i < new_size; i++) {
591                         u16 new_idx = (ioat->tail+i) & (new_size-1);
592
593                         ring[new_idx] = ioat2_alloc_ring_ent(c, GFP_NOWAIT);
594                         if (!ring[new_idx]) {
595                                 while (i--) {
596                                         u16 new_idx = (ioat->tail+i) & (new_size-1);
597
598                                         ioat2_free_ring_ent(ring[new_idx], c);
599                                 }
600                                 kfree(ring);
601                                 return false;
602                         }
603                         set_desc_id(ring[new_idx], new_idx);
604                 }
605
606                 /* hw link new descriptors */
607                 for (i = curr_size-1; i < new_size; i++) {
608                         u16 new_idx = (ioat->tail+i) & (new_size-1);
609                         struct ioat_ring_ent *next = ring[(new_idx+1) & (new_size-1)];
610                         struct ioat_dma_descriptor *hw = ring[new_idx]->hw;
611
612                         hw->next = next->txd.phys;
613                 }
614         } else {
615                 struct ioat_dma_descriptor *hw;
616                 struct ioat_ring_ent *next;
617
618                 /* copy current descriptors to the new ring, dropping the
619                  * removed descriptors
620                  */
621                 for (i = 0; i < new_size; i++) {
622                         u16 curr_idx = (ioat->tail+i) & (curr_size-1);
623                         u16 new_idx = (ioat->tail+i) & (new_size-1);
624
625                         ring[new_idx] = ioat->ring[curr_idx];
626                         set_desc_id(ring[new_idx], new_idx);
627                 }
628
629                 /* free deleted descriptors */
630                 for (i = new_size; i < curr_size; i++) {
631                         struct ioat_ring_ent *ent;
632
633                         ent = ioat2_get_ring_ent(ioat, ioat->tail+i);
634                         ioat2_free_ring_ent(ent, c);
635                 }
636
637                 /* fix up hardware ring */
638                 hw = ring[(ioat->tail+new_size-1) & (new_size-1)]->hw;
639                 next = ring[(ioat->tail+new_size) & (new_size-1)];
640                 hw->next = next->txd.phys;
641         }
642
643         dev_dbg(to_dev(chan), "%s: allocated %d descriptors\n",
644                 __func__, new_size);
645
646         kfree(ioat->ring);
647         ioat->ring = ring;
648         ioat->alloc_order = order;
649
650         return true;
651 }
652
653 /**
654  * ioat2_check_space_lock - verify space and grab ring producer lock
655  * @ioat: ioat2,3 channel (ring) to operate on
656  * @num_descs: allocation length
657  */
658 int ioat2_check_space_lock(struct ioat2_dma_chan *ioat, int num_descs)
659 {
660         struct ioat_chan_common *chan = &ioat->base;
661         bool retry;
662
663  retry:
664         spin_lock_bh(&ioat->prep_lock);
665         /* never allow the last descriptor to be consumed, we need at
666          * least one free at all times to allow for on-the-fly ring
667          * resizing.
668          */
669         if (likely(ioat2_ring_space(ioat) > num_descs)) {
670                 dev_dbg(to_dev(chan), "%s: num_descs: %d (%x:%x:%x)\n",
671                         __func__, num_descs, ioat->head, ioat->tail, ioat->issued);
672                 ioat->produce = num_descs;
673                 return 0;  /* with ioat->prep_lock held */
674         }
675         retry = test_and_set_bit(IOAT_RESHAPE_PENDING, &chan->state);
676         spin_unlock_bh(&ioat->prep_lock);
677
678         /* is another cpu already trying to expand the ring? */
679         if (retry)
680                 goto retry;
681
682         spin_lock_bh(&chan->cleanup_lock);
683         spin_lock_bh(&ioat->prep_lock);
684         retry = reshape_ring(ioat, ioat->alloc_order + 1);
685         clear_bit(IOAT_RESHAPE_PENDING, &chan->state);
686         spin_unlock_bh(&ioat->prep_lock);
687         spin_unlock_bh(&chan->cleanup_lock);
688
689         /* if we were able to expand the ring retry the allocation */
690         if (retry)
691                 goto retry;
692
693         if (printk_ratelimit())
694                 dev_dbg(to_dev(chan), "%s: ring full! num_descs: %d (%x:%x:%x)\n",
695                         __func__, num_descs, ioat->head, ioat->tail, ioat->issued);
696
697         /* progress reclaim in the allocation failure case we may be
698          * called under bh_disabled so we need to trigger the timer
699          * event directly
700          */
701         if (jiffies > chan->timer.expires && timer_pending(&chan->timer)) {
702                 struct ioatdma_device *device = chan->device;
703
704                 mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
705                 device->timer_fn((unsigned long) &chan->common);
706         }
707
708         return -ENOMEM;
709 }
710
711 struct dma_async_tx_descriptor *
712 ioat2_dma_prep_memcpy_lock(struct dma_chan *c, dma_addr_t dma_dest,
713                            dma_addr_t dma_src, size_t len, unsigned long flags)
714 {
715         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
716         struct ioat_dma_descriptor *hw;
717         struct ioat_ring_ent *desc;
718         dma_addr_t dst = dma_dest;
719         dma_addr_t src = dma_src;
720         size_t total_len = len;
721         int num_descs, idx, i;
722
723         num_descs = ioat2_xferlen_to_descs(ioat, len);
724         if (likely(num_descs) && ioat2_check_space_lock(ioat, num_descs) == 0)
725                 idx = ioat->head;
726         else
727                 return NULL;
728         i = 0;
729         do {
730                 size_t copy = min_t(size_t, len, 1 << ioat->xfercap_log);
731
732                 desc = ioat2_get_ring_ent(ioat, idx + i);
733                 hw = desc->hw;
734
735                 hw->size = copy;
736                 hw->ctl = 0;
737                 hw->src_addr = src;
738                 hw->dst_addr = dst;
739
740                 len -= copy;
741                 dst += copy;
742                 src += copy;
743                 dump_desc_dbg(ioat, desc);
744         } while (++i < num_descs);
745
746         desc->txd.flags = flags;
747         desc->len = total_len;
748         hw->ctl_f.int_en = !!(flags & DMA_PREP_INTERRUPT);
749         hw->ctl_f.fence = !!(flags & DMA_PREP_FENCE);
750         hw->ctl_f.compl_write = 1;
751         dump_desc_dbg(ioat, desc);
752         /* we leave the channel locked to ensure in order submission */
753
754         return &desc->txd;
755 }
756
757 /**
758  * ioat2_free_chan_resources - release all the descriptors
759  * @chan: the channel to be cleaned
760  */
761 void ioat2_free_chan_resources(struct dma_chan *c)
762 {
763         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
764         struct ioat_chan_common *chan = &ioat->base;
765         struct ioatdma_device *device = chan->device;
766         struct ioat_ring_ent *desc;
767         const u16 total_descs = 1 << ioat->alloc_order;
768         int descs;
769         int i;
770
771         /* Before freeing channel resources first check
772          * if they have been previously allocated for this channel.
773          */
774         if (!ioat->ring)
775                 return;
776
777         tasklet_disable(&chan->cleanup_task);
778         del_timer_sync(&chan->timer);
779         device->cleanup_fn((unsigned long) c);
780         device->reset_hw(chan);
781
782         spin_lock_bh(&chan->cleanup_lock);
783         spin_lock_bh(&ioat->prep_lock);
784         descs = ioat2_ring_space(ioat);
785         dev_dbg(to_dev(chan), "freeing %d idle descriptors\n", descs);
786         for (i = 0; i < descs; i++) {
787                 desc = ioat2_get_ring_ent(ioat, ioat->head + i);
788                 ioat2_free_ring_ent(desc, c);
789         }
790
791         if (descs < total_descs)
792                 dev_err(to_dev(chan), "Freeing %d in use descriptors!\n",
793                         total_descs - descs);
794
795         for (i = 0; i < total_descs - descs; i++) {
796                 desc = ioat2_get_ring_ent(ioat, ioat->tail + i);
797                 dump_desc_dbg(ioat, desc);
798                 ioat2_free_ring_ent(desc, c);
799         }
800
801         kfree(ioat->ring);
802         ioat->ring = NULL;
803         ioat->alloc_order = 0;
804         pci_pool_free(device->completion_pool, chan->completion,
805                       chan->completion_dma);
806         spin_unlock_bh(&ioat->prep_lock);
807         spin_unlock_bh(&chan->cleanup_lock);
808
809         chan->last_completion = 0;
810         chan->completion_dma = 0;
811         ioat->dmacount = 0;
812 }
813
814 static ssize_t ring_size_show(struct dma_chan *c, char *page)
815 {
816         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
817
818         return sprintf(page, "%d\n", (1 << ioat->alloc_order) & ~1);
819 }
820 static struct ioat_sysfs_entry ring_size_attr = __ATTR_RO(ring_size);
821
822 static ssize_t ring_active_show(struct dma_chan *c, char *page)
823 {
824         struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
825
826         /* ...taken outside the lock, no need to be precise */
827         return sprintf(page, "%d\n", ioat2_ring_active(ioat));
828 }
829 static struct ioat_sysfs_entry ring_active_attr = __ATTR_RO(ring_active);
830
831 static struct attribute *ioat2_attrs[] = {
832         &ring_size_attr.attr,
833         &ring_active_attr.attr,
834         &ioat_cap_attr.attr,
835         &ioat_version_attr.attr,
836         NULL,
837 };
838
839 struct kobj_type ioat2_ktype = {
840         .sysfs_ops = &ioat_sysfs_ops,
841         .default_attrs = ioat2_attrs,
842 };
843
844 int __devinit ioat2_dma_probe(struct ioatdma_device *device, int dca)
845 {
846         struct pci_dev *pdev = device->pdev;
847         struct dma_device *dma;
848         struct dma_chan *c;
849         struct ioat_chan_common *chan;
850         int err;
851
852         device->enumerate_channels = ioat2_enumerate_channels;
853         device->reset_hw = ioat2_reset_hw;
854         device->cleanup_fn = ioat2_cleanup_event;
855         device->timer_fn = ioat2_timer_event;
856         device->self_test = ioat_dma_self_test;
857         dma = &device->common;
858         dma->device_prep_dma_memcpy = ioat2_dma_prep_memcpy_lock;
859         dma->device_issue_pending = ioat2_issue_pending;
860         dma->device_alloc_chan_resources = ioat2_alloc_chan_resources;
861         dma->device_free_chan_resources = ioat2_free_chan_resources;
862         dma->device_tx_status = ioat_tx_status;
863
864         err = ioat_probe(device);
865         if (err)
866                 return err;
867         ioat_set_tcp_copy_break(2048);
868
869         list_for_each_entry(c, &dma->channels, device_node) {
870                 chan = to_chan_common(c);
871                 writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE | IOAT_DMA_DCA_ANY_CPU,
872                        chan->reg_base + IOAT_DCACTRL_OFFSET);
873         }
874
875         err = ioat_register(device);
876         if (err)
877                 return err;
878
879         ioat_kobject_add(device, &ioat2_ktype);
880
881         if (dca)
882                 device->dca = ioat2_dca_init(pdev, device->reg_base);
883
884         return err;
885 }