Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
dma_v2.c
Go to the documentation of this file.
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/prefetch.h>
38 #include <linux/i7300_idle.h>
39 #include "dma.h"
40 #include "dma_v2.h"
41 #include "registers.h"
42 #include "hw.h"
43 
44 #include "../dmaengine.h"
45 
49  "ioat2+: allocate 2^n descriptors per channel"
50  " (default: 8 max: 16)");
51 static int ioat_ring_max_alloc_order = IOAT_MAX_ORDER;
52 module_param(ioat_ring_max_alloc_order, int, 0644);
53 MODULE_PARM_DESC(ioat_ring_max_alloc_order,
54  "ioat2+: upper limit for ring size (default: 16)");
55 
57 {
58  struct ioat_chan_common *chan = &ioat->base;
59 
60  ioat->dmacount += ioat2_ring_pending(ioat);
61  ioat->issued = ioat->head;
63  dev_dbg(to_dev(chan),
64  "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
65  __func__, ioat->head, ioat->tail, ioat->issued, ioat->dmacount);
66 }
67 
69 {
70  struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
71 
72  if (ioat2_ring_pending(ioat)) {
73  spin_lock_bh(&ioat->prep_lock);
75  spin_unlock_bh(&ioat->prep_lock);
76  }
77 }
78 
86 static void ioat2_update_pending(struct ioat2_dma_chan *ioat)
87 {
88  if (ioat2_ring_pending(ioat) > ioat_pending_level)
90 }
91 
92 static void __ioat2_start_null_desc(struct ioat2_dma_chan *ioat)
93 {
94  struct ioat_ring_ent *desc;
95  struct ioat_dma_descriptor *hw;
96 
97  if (ioat2_ring_space(ioat) < 1) {
98  dev_err(to_dev(&ioat->base),
99  "Unable to start null desc - ring full\n");
100  return;
101  }
102 
103  dev_dbg(to_dev(&ioat->base), "%s: head: %#x tail: %#x issued: %#x\n",
104  __func__, ioat->head, ioat->tail, ioat->issued);
105  desc = ioat2_get_ring_ent(ioat, ioat->head);
106 
107  hw = desc->hw;
108  hw->ctl = 0;
109  hw->ctl_f.null = 1;
110  hw->ctl_f.int_en = 1;
111  hw->ctl_f.compl_write = 1;
112  /* set size to non-zero value (channel returns error when size is 0) */
114  hw->src_addr = 0;
115  hw->dst_addr = 0;
116  async_tx_ack(&desc->txd);
117  ioat2_set_chainaddr(ioat, desc->txd.phys);
118  dump_desc_dbg(ioat, desc);
119  wmb();
120  ioat->head += 1;
121  __ioat2_issue_pending(ioat);
122 }
123 
124 static void ioat2_start_null_desc(struct ioat2_dma_chan *ioat)
125 {
126  spin_lock_bh(&ioat->prep_lock);
127  __ioat2_start_null_desc(ioat);
128  spin_unlock_bh(&ioat->prep_lock);
129 }
130 
131 static void __cleanup(struct ioat2_dma_chan *ioat, dma_addr_t phys_complete)
132 {
133  struct ioat_chan_common *chan = &ioat->base;
134  struct dma_async_tx_descriptor *tx;
135  struct ioat_ring_ent *desc;
136  bool seen_current = false;
137  u16 active;
138  int idx = ioat->tail, i;
139 
140  dev_dbg(to_dev(chan), "%s: head: %#x tail: %#x issued: %#x\n",
141  __func__, ioat->head, ioat->tail, ioat->issued);
142 
143  active = ioat2_ring_active(ioat);
144  for (i = 0; i < active && !seen_current; i++) {
146  prefetch(ioat2_get_ring_ent(ioat, idx + i + 1));
147  desc = ioat2_get_ring_ent(ioat, idx + i);
148  tx = &desc->txd;
149  dump_desc_dbg(ioat, desc);
150  if (tx->cookie) {
151  ioat_dma_unmap(chan, tx->flags, desc->len, desc->hw);
152  dma_cookie_complete(tx);
153  if (tx->callback) {
154  tx->callback(tx->callback_param);
155  tx->callback = NULL;
156  }
157  }
158 
159  if (tx->phys == phys_complete)
160  seen_current = true;
161  }
162  smp_mb(); /* finish all descriptor reads before incrementing tail */
163  ioat->tail = idx + i;
164  BUG_ON(active && !seen_current); /* no active descs have written a completion? */
165 
166  chan->last_completion = phys_complete;
167  if (active - i == 0) {
168  dev_dbg(to_dev(chan), "%s: cancel completion timeout\n",
169  __func__);
171  mod_timer(&chan->timer, jiffies + IDLE_TIMEOUT);
172  }
173 }
174 
179 static void ioat2_cleanup(struct ioat2_dma_chan *ioat)
180 {
181  struct ioat_chan_common *chan = &ioat->base;
182  dma_addr_t phys_complete;
183 
184  spin_lock_bh(&chan->cleanup_lock);
185  if (ioat_cleanup_preamble(chan, &phys_complete))
186  __cleanup(ioat, phys_complete);
187  spin_unlock_bh(&chan->cleanup_lock);
188 }
189 
190 void ioat2_cleanup_event(unsigned long data)
191 {
192  struct ioat2_dma_chan *ioat = to_ioat2_chan((void *) data);
193 
194  ioat2_cleanup(ioat);
196 }
197 
199 {
200  struct ioat_chan_common *chan = &ioat->base;
201 
202  /* set the tail to be re-issued */
203  ioat->issued = ioat->tail;
204  ioat->dmacount = 0;
207 
208  dev_dbg(to_dev(chan),
209  "%s: head: %#x tail: %#x issued: %#x count: %#x\n",
210  __func__, ioat->head, ioat->tail, ioat->issued, ioat->dmacount);
211 
212  if (ioat2_ring_pending(ioat)) {
213  struct ioat_ring_ent *desc;
214 
215  desc = ioat2_get_ring_ent(ioat, ioat->tail);
216  ioat2_set_chainaddr(ioat, desc->txd.phys);
217  __ioat2_issue_pending(ioat);
218  } else
219  __ioat2_start_null_desc(ioat);
220 }
221 
222 int ioat2_quiesce(struct ioat_chan_common *chan, unsigned long tmo)
223 {
224  unsigned long end = jiffies + tmo;
225  int err = 0;
226  u32 status;
227 
228  status = ioat_chansts(chan);
229  if (is_ioat_active(status) || is_ioat_idle(status))
230  ioat_suspend(chan);
231  while (is_ioat_active(status) || is_ioat_idle(status)) {
232  if (tmo && time_after(jiffies, end)) {
233  err = -ETIMEDOUT;
234  break;
235  }
236  status = ioat_chansts(chan);
237  cpu_relax();
238  }
239 
240  return err;
241 }
242 
243 int ioat2_reset_sync(struct ioat_chan_common *chan, unsigned long tmo)
244 {
245  unsigned long end = jiffies + tmo;
246  int err = 0;
247 
248  ioat_reset(chan);
249  while (ioat_reset_pending(chan)) {
250  if (end && time_after(jiffies, end)) {
251  err = -ETIMEDOUT;
252  break;
253  }
254  cpu_relax();
255  }
256 
257  return err;
258 }
259 
260 static void ioat2_restart_channel(struct ioat2_dma_chan *ioat)
261 {
262  struct ioat_chan_common *chan = &ioat->base;
263  dma_addr_t phys_complete;
264 
265  ioat2_quiesce(chan, 0);
266  if (ioat_cleanup_preamble(chan, &phys_complete))
267  __cleanup(ioat, phys_complete);
268 
269  __ioat2_restart_chan(ioat);
270 }
271 
272 void ioat2_timer_event(unsigned long data)
273 {
274  struct ioat2_dma_chan *ioat = to_ioat2_chan((void *) data);
275  struct ioat_chan_common *chan = &ioat->base;
276 
277  if (test_bit(IOAT_COMPLETION_PENDING, &chan->state)) {
278  dma_addr_t phys_complete;
279  u64 status;
280 
281  status = ioat_chansts(chan);
282 
283  /* when halted due to errors check for channel
284  * programming errors before advancing the completion state
285  */
286  if (is_ioat_halted(status)) {
287  u32 chanerr;
288 
289  chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
290  dev_err(to_dev(chan), "%s: Channel halted (%x)\n",
291  __func__, chanerr);
292  if (test_bit(IOAT_RUN, &chan->state))
293  BUG_ON(is_ioat_bug(chanerr));
294  else /* we never got off the ground */
295  return;
296  }
297 
298  /* if we haven't made progress and we have already
299  * acknowledged a pending completion once, then be more
300  * forceful with a restart
301  */
302  spin_lock_bh(&chan->cleanup_lock);
303  if (ioat_cleanup_preamble(chan, &phys_complete)) {
304  __cleanup(ioat, phys_complete);
305  } else if (test_bit(IOAT_COMPLETION_ACK, &chan->state)) {
306  spin_lock_bh(&ioat->prep_lock);
307  ioat2_restart_channel(ioat);
308  spin_unlock_bh(&ioat->prep_lock);
309  } else {
312  }
313  spin_unlock_bh(&chan->cleanup_lock);
314  } else {
315  u16 active;
316 
317  /* if the ring is idle, empty, and oversized try to step
318  * down the size
319  */
320  spin_lock_bh(&chan->cleanup_lock);
321  spin_lock_bh(&ioat->prep_lock);
322  active = ioat2_ring_active(ioat);
323  if (active == 0 && ioat->alloc_order > ioat_get_alloc_order())
324  reshape_ring(ioat, ioat->alloc_order-1);
325  spin_unlock_bh(&ioat->prep_lock);
326  spin_unlock_bh(&chan->cleanup_lock);
327 
328  /* keep shrinking until we get back to our minimum
329  * default size
330  */
331  if (ioat->alloc_order > ioat_get_alloc_order())
332  mod_timer(&chan->timer, jiffies + IDLE_TIMEOUT);
333  }
334 }
335 
336 static int ioat2_reset_hw(struct ioat_chan_common *chan)
337 {
338  /* throw away whatever the channel was doing and get it initialized */
339  u32 chanerr;
340 
341  ioat2_quiesce(chan, msecs_to_jiffies(100));
342 
343  chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
344  writel(chanerr, chan->reg_base + IOAT_CHANERR_OFFSET);
345 
346  return ioat2_reset_sync(chan, msecs_to_jiffies(200));
347 }
348 
354 {
355  struct ioat2_dma_chan *ioat;
356  struct device *dev = &device->pdev->dev;
357  struct dma_device *dma = &device->common;
358  u8 xfercap_log;
359  int i;
360 
361  INIT_LIST_HEAD(&dma->channels);
362  dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
363  dma->chancnt &= 0x1f; /* bits [4:0] valid */
364  if (dma->chancnt > ARRAY_SIZE(device->idx)) {
365  dev_warn(dev, "(%d) exceeds max supported channels (%zu)\n",
366  dma->chancnt, ARRAY_SIZE(device->idx));
367  dma->chancnt = ARRAY_SIZE(device->idx);
368  }
369  xfercap_log = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
370  xfercap_log &= 0x1f; /* bits [4:0] valid */
371  if (xfercap_log == 0)
372  return 0;
373  dev_dbg(dev, "%s: xfercap = %d\n", __func__, 1 << xfercap_log);
374 
375  /* FIXME which i/oat version is i7300? */
376 #ifdef CONFIG_I7300_IDLE_IOAT_CHANNEL
377  if (i7300_idle_platform_probe(NULL, NULL, 1) == 0)
378  dma->chancnt--;
379 #endif
380  for (i = 0; i < dma->chancnt; i++) {
381  ioat = devm_kzalloc(dev, sizeof(*ioat), GFP_KERNEL);
382  if (!ioat)
383  break;
384 
385  ioat_init_channel(device, &ioat->base, i);
386  ioat->xfercap_log = xfercap_log;
387  spin_lock_init(&ioat->prep_lock);
388  if (device->reset_hw(&ioat->base)) {
389  i = 0;
390  break;
391  }
392  }
393  dma->chancnt = i;
394  return i;
395 }
396 
397 static dma_cookie_t ioat2_tx_submit_unlock(struct dma_async_tx_descriptor *tx)
398 {
399  struct dma_chan *c = tx->chan;
400  struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
401  struct ioat_chan_common *chan = &ioat->base;
403 
404  cookie = dma_cookie_assign(tx);
405  dev_dbg(to_dev(&ioat->base), "%s: cookie: %d\n", __func__, cookie);
406 
408  mod_timer(&chan->timer, jiffies + COMPLETION_TIMEOUT);
409 
410  /* make descriptor updates visible before advancing ioat->head,
411  * this is purposefully not smp_wmb() since we are also
412  * publishing the descriptor updates to a dma device
413  */
414  wmb();
415 
416  ioat->head += ioat->produce;
417 
418  ioat2_update_pending(ioat);
419  spin_unlock_bh(&ioat->prep_lock);
420 
421  return cookie;
422 }
423 
424 static struct ioat_ring_ent *ioat2_alloc_ring_ent(struct dma_chan *chan, gfp_t flags)
425 {
426  struct ioat_dma_descriptor *hw;
427  struct ioat_ring_ent *desc;
428  struct ioatdma_device *dma;
430 
431  dma = to_ioatdma_device(chan->device);
432  hw = pci_pool_alloc(dma->dma_pool, flags, &phys);
433  if (!hw)
434  return NULL;
435  memset(hw, 0, sizeof(*hw));
436 
437  desc = kmem_cache_zalloc(ioat2_cache, flags);
438  if (!desc) {
439  pci_pool_free(dma->dma_pool, hw, phys);
440  return NULL;
441  }
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);
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 void ioat2_free_chan_resources(struct dma_chan *c);
496 
497 /* ioat2_alloc_chan_resources - allocate/initialize ioat2 descriptor ring
498  * @chan: channel to be initialized
499  */
501 {
502  struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
503  struct ioat_chan_common *chan = &ioat->base;
504  struct ioat_ring_ent **ring;
505  u64 status;
506  int order;
507  int i = 0;
508 
509  /* have we already been set up? */
510  if (ioat->ring)
511  return 1 << ioat->alloc_order;
512 
513  /* Setup register to interrupt and write completion status on error */
515 
516  /* allocate a completion writeback area */
517  /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
518  chan->completion = pci_pool_alloc(chan->device->completion_pool,
519  GFP_KERNEL, &chan->completion_dma);
520  if (!chan->completion)
521  return -ENOMEM;
522 
523  memset(chan->completion, 0, sizeof(*chan->completion));
524  writel(((u64) chan->completion_dma) & 0x00000000FFFFFFFF,
526  writel(((u64) chan->completion_dma) >> 32,
528 
529  order = ioat_get_alloc_order();
530  ring = ioat2_alloc_ring(c, order, GFP_KERNEL);
531  if (!ring)
532  return -ENOMEM;
533 
534  spin_lock_bh(&chan->cleanup_lock);
535  spin_lock_bh(&ioat->prep_lock);
536  ioat->ring = ring;
537  ioat->head = 0;
538  ioat->issued = 0;
539  ioat->tail = 0;
540  ioat->alloc_order = order;
541  spin_unlock_bh(&ioat->prep_lock);
542  spin_unlock_bh(&chan->cleanup_lock);
543 
544  tasklet_enable(&chan->cleanup_task);
545  ioat2_start_null_desc(ioat);
546 
547  /* check that we got off the ground */
548  do {
549  udelay(1);
550  status = ioat_chansts(chan);
551  } while (i++ < 20 && !is_ioat_active(status) && !is_ioat_idle(status));
552 
553  if (is_ioat_active(status) || is_ioat_idle(status)) {
554  set_bit(IOAT_RUN, &chan->state);
555  return 1 << ioat->alloc_order;
556  } else {
557  u32 chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
558 
559  dev_WARN(to_dev(chan),
560  "failed to start channel chanerr: %#x\n", chanerr);
562  return -EFAULT;
563  }
564 }
565 
566 bool reshape_ring(struct ioat2_dma_chan *ioat, int order)
567 {
568  /* reshape differs from normal ring allocation in that we want
569  * to allocate a new software ring while only
570  * extending/truncating the hardware ring
571  */
572  struct ioat_chan_common *chan = &ioat->base;
573  struct dma_chan *c = &chan->common;
574  const u32 curr_size = ioat2_ring_size(ioat);
575  const u16 active = ioat2_ring_active(ioat);
576  const u32 new_size = 1 << order;
577  struct ioat_ring_ent **ring;
578  u16 i;
579 
580  if (order > ioat_get_max_alloc_order())
581  return false;
582 
583  /* double check that we have at least 1 free descriptor */
584  if (active == curr_size)
585  return false;
586 
587  /* when shrinking, verify that we can hold the current active
588  * set in the new ring
589  */
590  if (active >= new_size)
591  return false;
592 
593  /* allocate the array to hold the software ring */
594  ring = kcalloc(new_size, sizeof(*ring), GFP_NOWAIT);
595  if (!ring)
596  return false;
597 
598  /* allocate/trim descriptors as needed */
599  if (new_size > curr_size) {
600  /* copy current descriptors to the new ring */
601  for (i = 0; i < curr_size; i++) {
602  u16 curr_idx = (ioat->tail+i) & (curr_size-1);
603  u16 new_idx = (ioat->tail+i) & (new_size-1);
604 
605  ring[new_idx] = ioat->ring[curr_idx];
606  set_desc_id(ring[new_idx], new_idx);
607  }
608 
609  /* add new descriptors to the ring */
610  for (i = curr_size; i < new_size; i++) {
611  u16 new_idx = (ioat->tail+i) & (new_size-1);
612 
613  ring[new_idx] = ioat2_alloc_ring_ent(c, GFP_NOWAIT);
614  if (!ring[new_idx]) {
615  while (i--) {
616  u16 new_idx = (ioat->tail+i) & (new_size-1);
617 
618  ioat2_free_ring_ent(ring[new_idx], c);
619  }
620  kfree(ring);
621  return false;
622  }
623  set_desc_id(ring[new_idx], new_idx);
624  }
625 
626  /* hw link new descriptors */
627  for (i = curr_size-1; i < new_size; i++) {
628  u16 new_idx = (ioat->tail+i) & (new_size-1);
629  struct ioat_ring_ent *next = ring[(new_idx+1) & (new_size-1)];
630  struct ioat_dma_descriptor *hw = ring[new_idx]->hw;
631 
632  hw->next = next->txd.phys;
633  }
634  } else {
635  struct ioat_dma_descriptor *hw;
636  struct ioat_ring_ent *next;
637 
638  /* copy current descriptors to the new ring, dropping the
639  * removed descriptors
640  */
641  for (i = 0; i < new_size; i++) {
642  u16 curr_idx = (ioat->tail+i) & (curr_size-1);
643  u16 new_idx = (ioat->tail+i) & (new_size-1);
644 
645  ring[new_idx] = ioat->ring[curr_idx];
646  set_desc_id(ring[new_idx], new_idx);
647  }
648 
649  /* free deleted descriptors */
650  for (i = new_size; i < curr_size; i++) {
651  struct ioat_ring_ent *ent;
652 
653  ent = ioat2_get_ring_ent(ioat, ioat->tail+i);
654  ioat2_free_ring_ent(ent, c);
655  }
656 
657  /* fix up hardware ring */
658  hw = ring[(ioat->tail+new_size-1) & (new_size-1)]->hw;
659  next = ring[(ioat->tail+new_size) & (new_size-1)];
660  hw->next = next->txd.phys;
661  }
662 
663  dev_dbg(to_dev(chan), "%s: allocated %d descriptors\n",
664  __func__, new_size);
665 
666  kfree(ioat->ring);
667  ioat->ring = ring;
668  ioat->alloc_order = order;
669 
670  return true;
671 }
672 
678 int ioat2_check_space_lock(struct ioat2_dma_chan *ioat, int num_descs)
679 {
680  struct ioat_chan_common *chan = &ioat->base;
681  bool retry;
682 
683  retry:
684  spin_lock_bh(&ioat->prep_lock);
685  /* never allow the last descriptor to be consumed, we need at
686  * least one free at all times to allow for on-the-fly ring
687  * resizing.
688  */
689  if (likely(ioat2_ring_space(ioat) > num_descs)) {
690  dev_dbg(to_dev(chan), "%s: num_descs: %d (%x:%x:%x)\n",
691  __func__, num_descs, ioat->head, ioat->tail, ioat->issued);
692  ioat->produce = num_descs;
693  return 0; /* with ioat->prep_lock held */
694  }
696  spin_unlock_bh(&ioat->prep_lock);
697 
698  /* is another cpu already trying to expand the ring? */
699  if (retry)
700  goto retry;
701 
702  spin_lock_bh(&chan->cleanup_lock);
703  spin_lock_bh(&ioat->prep_lock);
704  retry = reshape_ring(ioat, ioat->alloc_order + 1);
706  spin_unlock_bh(&ioat->prep_lock);
707  spin_unlock_bh(&chan->cleanup_lock);
708 
709  /* if we were able to expand the ring retry the allocation */
710  if (retry)
711  goto retry;
712 
713  if (printk_ratelimit())
714  dev_dbg(to_dev(chan), "%s: ring full! num_descs: %d (%x:%x:%x)\n",
715  __func__, num_descs, ioat->head, ioat->tail, ioat->issued);
716 
717  /* progress reclaim in the allocation failure case we may be
718  * called under bh_disabled so we need to trigger the timer
719  * event directly
720  */
721  if (jiffies > chan->timer.expires && timer_pending(&chan->timer)) {
722  struct ioatdma_device *device = chan->device;
723 
725  device->timer_fn((unsigned long) &chan->common);
726  }
727 
728  return -ENOMEM;
729 }
730 
733  dma_addr_t dma_src, size_t len, unsigned long flags)
734 {
735  struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
736  struct ioat_dma_descriptor *hw;
737  struct ioat_ring_ent *desc;
738  dma_addr_t dst = dma_dest;
739  dma_addr_t src = dma_src;
740  size_t total_len = len;
741  int num_descs, idx, i;
742 
743  num_descs = ioat2_xferlen_to_descs(ioat, len);
744  if (likely(num_descs) && ioat2_check_space_lock(ioat, num_descs) == 0)
745  idx = ioat->head;
746  else
747  return NULL;
748  i = 0;
749  do {
750  size_t copy = min_t(size_t, len, 1 << ioat->xfercap_log);
751 
752  desc = ioat2_get_ring_ent(ioat, idx + i);
753  hw = desc->hw;
754 
755  hw->size = copy;
756  hw->ctl = 0;
757  hw->src_addr = src;
758  hw->dst_addr = dst;
759 
760  len -= copy;
761  dst += copy;
762  src += copy;
763  dump_desc_dbg(ioat, desc);
764  } while (++i < num_descs);
765 
766  desc->txd.flags = flags;
767  desc->len = total_len;
768  hw->ctl_f.int_en = !!(flags & DMA_PREP_INTERRUPT);
769  hw->ctl_f.fence = !!(flags & DMA_PREP_FENCE);
770  hw->ctl_f.compl_write = 1;
771  dump_desc_dbg(ioat, desc);
772  /* we leave the channel locked to ensure in order submission */
773 
774  return &desc->txd;
775 }
776 
782 {
783  struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
784  struct ioat_chan_common *chan = &ioat->base;
785  struct ioatdma_device *device = chan->device;
786  struct ioat_ring_ent *desc;
787  const u16 total_descs = 1 << ioat->alloc_order;
788  int descs;
789  int i;
790 
791  /* Before freeing channel resources first check
792  * if they have been previously allocated for this channel.
793  */
794  if (!ioat->ring)
795  return;
796 
797  tasklet_disable(&chan->cleanup_task);
798  del_timer_sync(&chan->timer);
799  device->cleanup_fn((unsigned long) c);
800  device->reset_hw(chan);
801  clear_bit(IOAT_RUN, &chan->state);
802 
803  spin_lock_bh(&chan->cleanup_lock);
804  spin_lock_bh(&ioat->prep_lock);
805  descs = ioat2_ring_space(ioat);
806  dev_dbg(to_dev(chan), "freeing %d idle descriptors\n", descs);
807  for (i = 0; i < descs; i++) {
808  desc = ioat2_get_ring_ent(ioat, ioat->head + i);
809  ioat2_free_ring_ent(desc, c);
810  }
811 
812  if (descs < total_descs)
813  dev_err(to_dev(chan), "Freeing %d in use descriptors!\n",
814  total_descs - descs);
815 
816  for (i = 0; i < total_descs - descs; i++) {
817  desc = ioat2_get_ring_ent(ioat, ioat->tail + i);
818  dump_desc_dbg(ioat, desc);
819  ioat2_free_ring_ent(desc, c);
820  }
821 
822  kfree(ioat->ring);
823  ioat->ring = NULL;
824  ioat->alloc_order = 0;
825  pci_pool_free(device->completion_pool, chan->completion,
826  chan->completion_dma);
827  spin_unlock_bh(&ioat->prep_lock);
828  spin_unlock_bh(&chan->cleanup_lock);
829 
830  chan->last_completion = 0;
831  chan->completion_dma = 0;
832  ioat->dmacount = 0;
833 }
834 
835 static ssize_t ring_size_show(struct dma_chan *c, char *page)
836 {
837  struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
838 
839  return sprintf(page, "%d\n", (1 << ioat->alloc_order) & ~1);
840 }
841 static struct ioat_sysfs_entry ring_size_attr = __ATTR_RO(ring_size);
842 
843 static ssize_t ring_active_show(struct dma_chan *c, char *page)
844 {
845  struct ioat2_dma_chan *ioat = to_ioat2_chan(c);
846 
847  /* ...taken outside the lock, no need to be precise */
848  return sprintf(page, "%d\n", ioat2_ring_active(ioat));
849 }
850 static struct ioat_sysfs_entry ring_active_attr = __ATTR_RO(ring_active);
851 
852 static struct attribute *ioat2_attrs[] = {
853  &ring_size_attr.attr,
854  &ring_active_attr.attr,
855  &ioat_cap_attr.attr,
856  &ioat_version_attr.attr,
857  NULL,
858 };
859 
861  .sysfs_ops = &ioat_sysfs_ops,
862  .default_attrs = ioat2_attrs,
863 };
864 
866 {
867  struct pci_dev *pdev = device->pdev;
868  struct dma_device *dma;
869  struct dma_chan *c;
870  struct ioat_chan_common *chan;
871  int err;
872 
874  device->reset_hw = ioat2_reset_hw;
876  device->timer_fn = ioat2_timer_event;
877  device->self_test = ioat_dma_self_test;
878  dma = &device->common;
884 
885  err = ioat_probe(device);
886  if (err)
887  return err;
888  ioat_set_tcp_copy_break(2048);
889 
891  chan = to_chan_common(c);
893  chan->reg_base + IOAT_DCACTRL_OFFSET);
894  }
895 
896  err = ioat_register(device);
897  if (err)
898  return err;
899 
900  ioat_kobject_add(device, &ioat2_ktype);
901 
902  if (dca)
903  device->dca = ioat2_dca_init(pdev, device->reg_base);
904 
905  return err;
906 }