Merge git://git.kernel.org/pub/scm/linux/kernel/git/bart/ide-2.6
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_sdma.c
1 /*
2  * Copyright (c) 2007, 2008 QLogic Corporation. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32
33 #include <linux/spinlock.h>
34
35 #include "ipath_kernel.h"
36 #include "ipath_verbs.h"
37 #include "ipath_common.h"
38
39 #define SDMA_DESCQ_SZ PAGE_SIZE /* 256 entries per 4KB page */
40
41 static void vl15_watchdog_enq(struct ipath_devdata *dd)
42 {
43         /* ipath_sdma_lock must already be held */
44         if (atomic_inc_return(&dd->ipath_sdma_vl15_count) == 1) {
45                 unsigned long interval = (HZ + 19) / 20;
46                 dd->ipath_sdma_vl15_timer.expires = jiffies + interval;
47                 add_timer(&dd->ipath_sdma_vl15_timer);
48         }
49 }
50
51 static void vl15_watchdog_deq(struct ipath_devdata *dd)
52 {
53         /* ipath_sdma_lock must already be held */
54         if (atomic_dec_return(&dd->ipath_sdma_vl15_count) != 0) {
55                 unsigned long interval = (HZ + 19) / 20;
56                 mod_timer(&dd->ipath_sdma_vl15_timer, jiffies + interval);
57         } else {
58                 del_timer(&dd->ipath_sdma_vl15_timer);
59         }
60 }
61
62 static void vl15_watchdog_timeout(unsigned long opaque)
63 {
64         struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
65
66         if (atomic_read(&dd->ipath_sdma_vl15_count) != 0) {
67                 ipath_dbg("vl15 watchdog timeout - clearing\n");
68                 ipath_cancel_sends(dd, 1);
69                 ipath_hol_down(dd);
70         } else {
71                 ipath_dbg("vl15 watchdog timeout - "
72                           "condition already cleared\n");
73         }
74 }
75
76 static void unmap_desc(struct ipath_devdata *dd, unsigned head)
77 {
78         __le64 *descqp = &dd->ipath_sdma_descq[head].qw[0];
79         u64 desc[2];
80         dma_addr_t addr;
81         size_t len;
82
83         desc[0] = le64_to_cpu(descqp[0]);
84         desc[1] = le64_to_cpu(descqp[1]);
85
86         addr = (desc[1] << 32) | (desc[0] >> 32);
87         len = (desc[0] >> 14) & (0x7ffULL << 2);
88         dma_unmap_single(&dd->pcidev->dev, addr, len, DMA_TO_DEVICE);
89 }
90
91 /*
92  * ipath_sdma_lock should be locked before calling this.
93  */
94 int ipath_sdma_make_progress(struct ipath_devdata *dd)
95 {
96         struct list_head *lp = NULL;
97         struct ipath_sdma_txreq *txp = NULL;
98         u16 dmahead;
99         u16 start_idx = 0;
100         int progress = 0;
101
102         if (!list_empty(&dd->ipath_sdma_activelist)) {
103                 lp = dd->ipath_sdma_activelist.next;
104                 txp = list_entry(lp, struct ipath_sdma_txreq, list);
105                 start_idx = txp->start_idx;
106         }
107
108         /*
109          * Read the SDMA head register in order to know that the
110          * interrupt clear has been written to the chip.
111          * Otherwise, we may not get an interrupt for the last
112          * descriptor in the queue.
113          */
114         dmahead = (u16)ipath_read_kreg32(dd, dd->ipath_kregs->kr_senddmahead);
115         /* sanity check return value for error handling (chip reset, etc.) */
116         if (dmahead >= dd->ipath_sdma_descq_cnt)
117                 goto done;
118
119         while (dd->ipath_sdma_descq_head != dmahead) {
120                 if (txp && txp->flags & IPATH_SDMA_TXREQ_F_FREEDESC &&
121                     dd->ipath_sdma_descq_head == start_idx) {
122                         unmap_desc(dd, dd->ipath_sdma_descq_head);
123                         start_idx++;
124                         if (start_idx == dd->ipath_sdma_descq_cnt)
125                                 start_idx = 0;
126                 }
127
128                 /* increment free count and head */
129                 dd->ipath_sdma_descq_removed++;
130                 if (++dd->ipath_sdma_descq_head == dd->ipath_sdma_descq_cnt)
131                         dd->ipath_sdma_descq_head = 0;
132
133                 if (txp && txp->next_descq_idx == dd->ipath_sdma_descq_head) {
134                         /* move to notify list */
135                         if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
136                                 vl15_watchdog_deq(dd);
137                         list_move_tail(lp, &dd->ipath_sdma_notifylist);
138                         if (!list_empty(&dd->ipath_sdma_activelist)) {
139                                 lp = dd->ipath_sdma_activelist.next;
140                                 txp = list_entry(lp, struct ipath_sdma_txreq,
141                                                  list);
142                                 start_idx = txp->start_idx;
143                         } else {
144                                 lp = NULL;
145                                 txp = NULL;
146                         }
147                 }
148                 progress = 1;
149         }
150
151         if (progress)
152                 tasklet_hi_schedule(&dd->ipath_sdma_notify_task);
153
154 done:
155         return progress;
156 }
157
158 static void ipath_sdma_notify(struct ipath_devdata *dd, struct list_head *list)
159 {
160         struct ipath_sdma_txreq *txp, *txp_next;
161
162         list_for_each_entry_safe(txp, txp_next, list, list) {
163                 list_del_init(&txp->list);
164
165                 if (txp->callback)
166                         (*txp->callback)(txp->callback_cookie,
167                                          txp->callback_status);
168         }
169 }
170
171 static void sdma_notify_taskbody(struct ipath_devdata *dd)
172 {
173         unsigned long flags;
174         struct list_head list;
175
176         INIT_LIST_HEAD(&list);
177
178         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
179
180         list_splice_init(&dd->ipath_sdma_notifylist, &list);
181
182         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
183
184         ipath_sdma_notify(dd, &list);
185
186         /*
187          * The IB verbs layer needs to see the callback before getting
188          * the call to ipath_ib_piobufavail() because the callback
189          * handles releasing resources the next send will need.
190          * Otherwise, we could do these calls in
191          * ipath_sdma_make_progress().
192          */
193         ipath_ib_piobufavail(dd->verbs_dev);
194 }
195
196 static void sdma_notify_task(unsigned long opaque)
197 {
198         struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
199
200         if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
201                 sdma_notify_taskbody(dd);
202 }
203
204 static void dump_sdma_state(struct ipath_devdata *dd)
205 {
206         unsigned long reg;
207
208         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmastatus);
209         ipath_cdbg(VERBOSE, "kr_senddmastatus: 0x%016lx\n", reg);
210
211         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendctrl);
212         ipath_cdbg(VERBOSE, "kr_sendctrl: 0x%016lx\n", reg);
213
214         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask0);
215         ipath_cdbg(VERBOSE, "kr_senddmabufmask0: 0x%016lx\n", reg);
216
217         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask1);
218         ipath_cdbg(VERBOSE, "kr_senddmabufmask1: 0x%016lx\n", reg);
219
220         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask2);
221         ipath_cdbg(VERBOSE, "kr_senddmabufmask2: 0x%016lx\n", reg);
222
223         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmatail);
224         ipath_cdbg(VERBOSE, "kr_senddmatail: 0x%016lx\n", reg);
225
226         reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmahead);
227         ipath_cdbg(VERBOSE, "kr_senddmahead: 0x%016lx\n", reg);
228 }
229
230 static void sdma_abort_task(unsigned long opaque)
231 {
232         struct ipath_devdata *dd = (struct ipath_devdata *) opaque;
233         u64 status;
234         unsigned long flags;
235
236         if (test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
237                 return;
238
239         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
240
241         status = dd->ipath_sdma_status & IPATH_SDMA_ABORT_MASK;
242
243         /* nothing to do */
244         if (status == IPATH_SDMA_ABORT_NONE)
245                 goto unlock;
246
247         /* ipath_sdma_abort() is done, waiting for interrupt */
248         if (status == IPATH_SDMA_ABORT_DISARMED) {
249                 if (jiffies < dd->ipath_sdma_abort_intr_timeout)
250                         goto resched_noprint;
251                 /* give up, intr got lost somewhere */
252                 ipath_dbg("give up waiting for SDMADISABLED intr\n");
253                 __set_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
254                 status = IPATH_SDMA_ABORT_ABORTED;
255         }
256
257         /* everything is stopped, time to clean up and restart */
258         if (status == IPATH_SDMA_ABORT_ABORTED) {
259                 struct ipath_sdma_txreq *txp, *txpnext;
260                 u64 hwstatus;
261                 int notify = 0;
262
263                 hwstatus = ipath_read_kreg64(dd,
264                                 dd->ipath_kregs->kr_senddmastatus);
265
266                 if (/* ScoreBoardDrainInProg */
267                     test_bit(63, &hwstatus) ||
268                     /* AbortInProg */
269                     test_bit(62, &hwstatus) ||
270                     /* InternalSDmaEnable */
271                     test_bit(61, &hwstatus) ||
272                     /* ScbEmpty */
273                     !test_bit(30, &hwstatus)) {
274                         if (dd->ipath_sdma_reset_wait > 0) {
275                                 /* not done shutting down sdma */
276                                 --dd->ipath_sdma_reset_wait;
277                                 goto resched;
278                         }
279                         ipath_cdbg(VERBOSE, "gave up waiting for quiescent "
280                                 "status after SDMA reset, continuing\n");
281                         dump_sdma_state(dd);
282                 }
283
284                 /* dequeue all "sent" requests */
285                 list_for_each_entry_safe(txp, txpnext,
286                                          &dd->ipath_sdma_activelist, list) {
287                         txp->callback_status = IPATH_SDMA_TXREQ_S_ABORTED;
288                         if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
289                                 vl15_watchdog_deq(dd);
290                         list_move_tail(&txp->list, &dd->ipath_sdma_notifylist);
291                         notify = 1;
292                 }
293                 if (notify)
294                         tasklet_hi_schedule(&dd->ipath_sdma_notify_task);
295
296                 /* reset our notion of head and tail */
297                 dd->ipath_sdma_descq_tail = 0;
298                 dd->ipath_sdma_descq_head = 0;
299                 dd->ipath_sdma_head_dma[0] = 0;
300                 dd->ipath_sdma_generation = 0;
301                 dd->ipath_sdma_descq_removed = dd->ipath_sdma_descq_added;
302
303                 /* Reset SendDmaLenGen */
304                 ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen,
305                         (u64) dd->ipath_sdma_descq_cnt | (1ULL << 18));
306
307                 /* done with sdma state for a bit */
308                 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
309
310                 /*
311                  * Don't restart sdma here. Wait until link is up to ACTIVE.
312                  * VL15 MADs used to bring the link up use PIO, and multiple
313                  * link transitions otherwise cause the sdma engine to be
314                  * stopped and started multiple times.
315                  * The disable is done here, including the shadow, so the
316                  * state is kept consistent.
317                  * See ipath_restart_sdma() for the actual starting of sdma.
318                  */
319                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
320                 dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
321                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
322                                  dd->ipath_sendctrl);
323                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
324                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
325
326                 /* make sure I see next message */
327                 dd->ipath_sdma_abort_jiffies = 0;
328
329                 goto done;
330         }
331
332 resched:
333         /*
334          * for now, keep spinning
335          * JAG - this is bad to just have default be a loop without
336          * state change
337          */
338         if (jiffies > dd->ipath_sdma_abort_jiffies) {
339                 ipath_dbg("looping with status 0x%016llx\n",
340                           dd->ipath_sdma_status);
341                 dd->ipath_sdma_abort_jiffies = jiffies + 5 * HZ;
342         }
343 resched_noprint:
344         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
345         if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
346                 tasklet_hi_schedule(&dd->ipath_sdma_abort_task);
347         return;
348
349 unlock:
350         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
351 done:
352         return;
353 }
354
355 /*
356  * This is called from interrupt context.
357  */
358 void ipath_sdma_intr(struct ipath_devdata *dd)
359 {
360         unsigned long flags;
361
362         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
363
364         (void) ipath_sdma_make_progress(dd);
365
366         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
367 }
368
369 static int alloc_sdma(struct ipath_devdata *dd)
370 {
371         int ret = 0;
372
373         /* Allocate memory for SendDMA descriptor FIFO */
374         dd->ipath_sdma_descq = dma_alloc_coherent(&dd->pcidev->dev,
375                 SDMA_DESCQ_SZ, &dd->ipath_sdma_descq_phys, GFP_KERNEL);
376
377         if (!dd->ipath_sdma_descq) {
378                 ipath_dev_err(dd, "failed to allocate SendDMA descriptor "
379                         "FIFO memory\n");
380                 ret = -ENOMEM;
381                 goto done;
382         }
383
384         dd->ipath_sdma_descq_cnt =
385                 SDMA_DESCQ_SZ / sizeof(struct ipath_sdma_desc);
386
387         /* Allocate memory for DMA of head register to memory */
388         dd->ipath_sdma_head_dma = dma_alloc_coherent(&dd->pcidev->dev,
389                 PAGE_SIZE, &dd->ipath_sdma_head_phys, GFP_KERNEL);
390         if (!dd->ipath_sdma_head_dma) {
391                 ipath_dev_err(dd, "failed to allocate SendDMA head memory\n");
392                 ret = -ENOMEM;
393                 goto cleanup_descq;
394         }
395         dd->ipath_sdma_head_dma[0] = 0;
396
397         init_timer(&dd->ipath_sdma_vl15_timer);
398         dd->ipath_sdma_vl15_timer.function = vl15_watchdog_timeout;
399         dd->ipath_sdma_vl15_timer.data = (unsigned long)dd;
400         atomic_set(&dd->ipath_sdma_vl15_count, 0);
401
402         goto done;
403
404 cleanup_descq:
405         dma_free_coherent(&dd->pcidev->dev, SDMA_DESCQ_SZ,
406                 (void *)dd->ipath_sdma_descq, dd->ipath_sdma_descq_phys);
407         dd->ipath_sdma_descq = NULL;
408         dd->ipath_sdma_descq_phys = 0;
409 done:
410         return ret;
411 }
412
413 int setup_sdma(struct ipath_devdata *dd)
414 {
415         int ret = 0;
416         unsigned i, n;
417         u64 tmp64;
418         u64 senddmabufmask[3] = { 0 };
419         unsigned long flags;
420
421         ret = alloc_sdma(dd);
422         if (ret)
423                 goto done;
424
425         if (!dd->ipath_sdma_descq) {
426                 ipath_dev_err(dd, "SendDMA memory not allocated\n");
427                 goto done;
428         }
429
430         dd->ipath_sdma_status = 0;
431         dd->ipath_sdma_abort_jiffies = 0;
432         dd->ipath_sdma_generation = 0;
433         dd->ipath_sdma_descq_tail = 0;
434         dd->ipath_sdma_descq_head = 0;
435         dd->ipath_sdma_descq_removed = 0;
436         dd->ipath_sdma_descq_added = 0;
437
438         /* Set SendDmaBase */
439         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabase,
440                          dd->ipath_sdma_descq_phys);
441         /* Set SendDmaLenGen */
442         tmp64 = dd->ipath_sdma_descq_cnt;
443         tmp64 |= 1<<18; /* enable generation checking */
444         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen, tmp64);
445         /* Set SendDmaTail */
446         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail,
447                          dd->ipath_sdma_descq_tail);
448         /* Set SendDmaHeadAddr */
449         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmaheadaddr,
450                          dd->ipath_sdma_head_phys);
451
452         /* Reserve all the former "kernel" piobufs */
453         n = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - dd->ipath_pioreserved;
454         for (i = dd->ipath_lastport_piobuf; i < n; ++i) {
455                 unsigned word = i / 64;
456                 unsigned bit = i & 63;
457                 BUG_ON(word >= 3);
458                 senddmabufmask[word] |= 1ULL << bit;
459         }
460         ipath_chg_pioavailkernel(dd, dd->ipath_lastport_piobuf,
461                 n - dd->ipath_lastport_piobuf, 0);
462         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask0,
463                          senddmabufmask[0]);
464         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask1,
465                          senddmabufmask[1]);
466         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask2,
467                          senddmabufmask[2]);
468
469         INIT_LIST_HEAD(&dd->ipath_sdma_activelist);
470         INIT_LIST_HEAD(&dd->ipath_sdma_notifylist);
471
472         tasklet_init(&dd->ipath_sdma_notify_task, sdma_notify_task,
473                      (unsigned long) dd);
474         tasklet_init(&dd->ipath_sdma_abort_task, sdma_abort_task,
475                      (unsigned long) dd);
476
477         /*
478          * No use to turn on SDMA here, as link is probably not ACTIVE
479          * Just mark it RUNNING and enable the interrupt, and let the
480          * ipath_restart_sdma() on link transition to ACTIVE actually
481          * enable it.
482          */
483         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
484         dd->ipath_sendctrl |= INFINIPATH_S_SDMAINTENABLE;
485         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
486         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
487         __set_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status);
488         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
489
490 done:
491         return ret;
492 }
493
494 void teardown_sdma(struct ipath_devdata *dd)
495 {
496         struct ipath_sdma_txreq *txp, *txpnext;
497         unsigned long flags;
498         dma_addr_t sdma_head_phys = 0;
499         dma_addr_t sdma_descq_phys = 0;
500         void *sdma_descq = NULL;
501         void *sdma_head_dma = NULL;
502
503         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
504         __clear_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status);
505         __set_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
506         __set_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status);
507         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
508
509         tasklet_kill(&dd->ipath_sdma_abort_task);
510         tasklet_kill(&dd->ipath_sdma_notify_task);
511
512         /* turn off sdma */
513         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
514         dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
515         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
516                 dd->ipath_sendctrl);
517         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
518         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
519
520         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
521         /* dequeue all "sent" requests */
522         list_for_each_entry_safe(txp, txpnext, &dd->ipath_sdma_activelist,
523                                  list) {
524                 txp->callback_status = IPATH_SDMA_TXREQ_S_SHUTDOWN;
525                 if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
526                         vl15_watchdog_deq(dd);
527                 list_move_tail(&txp->list, &dd->ipath_sdma_notifylist);
528         }
529         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
530
531         sdma_notify_taskbody(dd);
532
533         del_timer_sync(&dd->ipath_sdma_vl15_timer);
534
535         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
536
537         dd->ipath_sdma_abort_jiffies = 0;
538
539         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabase, 0);
540         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen, 0);
541         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, 0);
542         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmaheadaddr, 0);
543         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask0, 0);
544         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask1, 0);
545         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask2, 0);
546
547         if (dd->ipath_sdma_head_dma) {
548                 sdma_head_dma = (void *) dd->ipath_sdma_head_dma;
549                 sdma_head_phys = dd->ipath_sdma_head_phys;
550                 dd->ipath_sdma_head_dma = NULL;
551                 dd->ipath_sdma_head_phys = 0;
552         }
553
554         if (dd->ipath_sdma_descq) {
555                 sdma_descq = dd->ipath_sdma_descq;
556                 sdma_descq_phys = dd->ipath_sdma_descq_phys;
557                 dd->ipath_sdma_descq = NULL;
558                 dd->ipath_sdma_descq_phys = 0;
559         }
560
561         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
562
563         if (sdma_head_dma)
564                 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
565                                   sdma_head_dma, sdma_head_phys);
566
567         if (sdma_descq)
568                 dma_free_coherent(&dd->pcidev->dev, SDMA_DESCQ_SZ,
569                                   sdma_descq, sdma_descq_phys);
570 }
571
572 /*
573  * [Re]start SDMA, if we use it, and it's not already OK.
574  * This is called on transition to link ACTIVE, either the first or
575  * subsequent times.
576  */
577 void ipath_restart_sdma(struct ipath_devdata *dd)
578 {
579         unsigned long flags;
580         int needed = 1;
581
582         if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA))
583                 goto bail;
584
585         /*
586          * First, make sure we should, which is to say,
587          * check that we are "RUNNING" (not in teardown)
588          * and not "SHUTDOWN"
589          */
590         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
591         if (!test_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status)
592                 || test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
593                         needed = 0;
594         else {
595                 __clear_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
596                 __clear_bit(IPATH_SDMA_DISARMED, &dd->ipath_sdma_status);
597                 __clear_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
598         }
599         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
600         if (!needed) {
601                 ipath_dbg("invalid attempt to restart SDMA, status 0x%016llx\n",
602                         dd->ipath_sdma_status);
603                 goto bail;
604         }
605         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
606         /*
607          * First clear, just to be safe. Enable is only done
608          * in chip on 0->1 transition
609          */
610         dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
611         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
612         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
613         dd->ipath_sendctrl |= INFINIPATH_S_SDMAENABLE;
614         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
615         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
616         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
617
618 bail:
619         return;
620 }
621
622 static inline void make_sdma_desc(struct ipath_devdata *dd,
623         u64 *sdmadesc, u64 addr, u64 dwlen, u64 dwoffset)
624 {
625         WARN_ON(addr & 3);
626         /* SDmaPhyAddr[47:32] */
627         sdmadesc[1] = addr >> 32;
628         /* SDmaPhyAddr[31:0] */
629         sdmadesc[0] = (addr & 0xfffffffcULL) << 32;
630         /* SDmaGeneration[1:0] */
631         sdmadesc[0] |= (dd->ipath_sdma_generation & 3ULL) << 30;
632         /* SDmaDwordCount[10:0] */
633         sdmadesc[0] |= (dwlen & 0x7ffULL) << 16;
634         /* SDmaBufOffset[12:2] */
635         sdmadesc[0] |= dwoffset & 0x7ffULL;
636 }
637
638 /*
639  * This function queues one IB packet onto the send DMA queue per call.
640  * The caller is responsible for checking:
641  * 1) The number of send DMA descriptor entries is less than the size of
642  *    the descriptor queue.
643  * 2) The IB SGE addresses and lengths are 32-bit aligned
644  *    (except possibly the last SGE's length)
645  * 3) The SGE addresses are suitable for passing to dma_map_single().
646  */
647 int ipath_sdma_verbs_send(struct ipath_devdata *dd,
648         struct ipath_sge_state *ss, u32 dwords,
649         struct ipath_verbs_txreq *tx)
650 {
651
652         unsigned long flags;
653         struct ipath_sge *sge;
654         int ret = 0;
655         u16 tail;
656         __le64 *descqp;
657         u64 sdmadesc[2];
658         u32 dwoffset;
659         dma_addr_t addr;
660
661         if ((tx->map_len + (dwords<<2)) > dd->ipath_ibmaxlen) {
662                 ipath_dbg("packet size %X > ibmax %X, fail\n",
663                         tx->map_len + (dwords<<2), dd->ipath_ibmaxlen);
664                 ret = -EMSGSIZE;
665                 goto fail;
666         }
667
668         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
669
670 retry:
671         if (unlikely(test_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status))) {
672                 ret = -EBUSY;
673                 goto unlock;
674         }
675
676         if (tx->txreq.sg_count > ipath_sdma_descq_freecnt(dd)) {
677                 if (ipath_sdma_make_progress(dd))
678                         goto retry;
679                 ret = -ENOBUFS;
680                 goto unlock;
681         }
682
683         addr = dma_map_single(&dd->pcidev->dev, tx->txreq.map_addr,
684                               tx->map_len, DMA_TO_DEVICE);
685         if (dma_mapping_error(addr)) {
686                 ret = -EIO;
687                 goto unlock;
688         }
689
690         dwoffset = tx->map_len >> 2;
691         make_sdma_desc(dd, sdmadesc, (u64) addr, dwoffset, 0);
692
693         /* SDmaFirstDesc */
694         sdmadesc[0] |= 1ULL << 12;
695         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_USELARGEBUF)
696                 sdmadesc[0] |= 1ULL << 14;      /* SDmaUseLargeBuf */
697
698         /* write to the descq */
699         tail = dd->ipath_sdma_descq_tail;
700         descqp = &dd->ipath_sdma_descq[tail].qw[0];
701         *descqp++ = cpu_to_le64(sdmadesc[0]);
702         *descqp++ = cpu_to_le64(sdmadesc[1]);
703
704         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEDESC)
705                 tx->txreq.start_idx = tail;
706
707         /* increment the tail */
708         if (++tail == dd->ipath_sdma_descq_cnt) {
709                 tail = 0;
710                 descqp = &dd->ipath_sdma_descq[0].qw[0];
711                 ++dd->ipath_sdma_generation;
712         }
713
714         sge = &ss->sge;
715         while (dwords) {
716                 u32 dw;
717                 u32 len;
718
719                 len = dwords << 2;
720                 if (len > sge->length)
721                         len = sge->length;
722                 if (len > sge->sge_length)
723                         len = sge->sge_length;
724                 BUG_ON(len == 0);
725                 dw = (len + 3) >> 2;
726                 addr = dma_map_single(&dd->pcidev->dev, sge->vaddr, dw << 2,
727                                       DMA_TO_DEVICE);
728                 make_sdma_desc(dd, sdmadesc, (u64) addr, dw, dwoffset);
729                 /* SDmaUseLargeBuf has to be set in every descriptor */
730                 if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_USELARGEBUF)
731                         sdmadesc[0] |= 1ULL << 14;
732                 /* write to the descq */
733                 *descqp++ = cpu_to_le64(sdmadesc[0]);
734                 *descqp++ = cpu_to_le64(sdmadesc[1]);
735
736                 /* increment the tail */
737                 if (++tail == dd->ipath_sdma_descq_cnt) {
738                         tail = 0;
739                         descqp = &dd->ipath_sdma_descq[0].qw[0];
740                         ++dd->ipath_sdma_generation;
741                 }
742                 sge->vaddr += len;
743                 sge->length -= len;
744                 sge->sge_length -= len;
745                 if (sge->sge_length == 0) {
746                         if (--ss->num_sge)
747                                 *sge = *ss->sg_list++;
748                 } else if (sge->length == 0 && sge->mr != NULL) {
749                         if (++sge->n >= IPATH_SEGSZ) {
750                                 if (++sge->m >= sge->mr->mapsz)
751                                         break;
752                                 sge->n = 0;
753                         }
754                         sge->vaddr =
755                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
756                         sge->length =
757                                 sge->mr->map[sge->m]->segs[sge->n].length;
758                 }
759
760                 dwoffset += dw;
761                 dwords -= dw;
762         }
763
764         if (!tail)
765                 descqp = &dd->ipath_sdma_descq[dd->ipath_sdma_descq_cnt].qw[0];
766         descqp -= 2;
767         /* SDmaLastDesc */
768         descqp[0] |= __constant_cpu_to_le64(1ULL << 11);
769         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_INTREQ) {
770                 /* SDmaIntReq */
771                 descqp[0] |= __constant_cpu_to_le64(1ULL << 15);
772         }
773
774         /* Commit writes to memory and advance the tail on the chip */
775         wmb();
776         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, tail);
777
778         tx->txreq.next_descq_idx = tail;
779         tx->txreq.callback_status = IPATH_SDMA_TXREQ_S_OK;
780         dd->ipath_sdma_descq_tail = tail;
781         dd->ipath_sdma_descq_added += tx->txreq.sg_count;
782         list_add_tail(&tx->txreq.list, &dd->ipath_sdma_activelist);
783         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_VL15)
784                 vl15_watchdog_enq(dd);
785
786 unlock:
787         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
788 fail:
789         return ret;
790 }