Merge branch 'x86/pebs' into x86/unify-cpu-detect
[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 ((hwstatus & (IPATH_SDMA_STATUS_SCORE_BOARD_DRAIN_IN_PROG |
267                                  IPATH_SDMA_STATUS_ABORT_IN_PROG             |
268                                  IPATH_SDMA_STATUS_INTERNAL_SDMA_ENABLE)) ||
269                     !(hwstatus & IPATH_SDMA_STATUS_SCB_EMPTY)) {
270                         if (dd->ipath_sdma_reset_wait > 0) {
271                                 /* not done shutting down sdma */
272                                 --dd->ipath_sdma_reset_wait;
273                                 goto resched;
274                         }
275                         ipath_cdbg(VERBOSE, "gave up waiting for quiescent "
276                                 "status after SDMA reset, continuing\n");
277                         dump_sdma_state(dd);
278                 }
279
280                 /* dequeue all "sent" requests */
281                 list_for_each_entry_safe(txp, txpnext,
282                                          &dd->ipath_sdma_activelist, list) {
283                         txp->callback_status = IPATH_SDMA_TXREQ_S_ABORTED;
284                         if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
285                                 vl15_watchdog_deq(dd);
286                         list_move_tail(&txp->list, &dd->ipath_sdma_notifylist);
287                         notify = 1;
288                 }
289                 if (notify)
290                         tasklet_hi_schedule(&dd->ipath_sdma_notify_task);
291
292                 /* reset our notion of head and tail */
293                 dd->ipath_sdma_descq_tail = 0;
294                 dd->ipath_sdma_descq_head = 0;
295                 dd->ipath_sdma_head_dma[0] = 0;
296                 dd->ipath_sdma_generation = 0;
297                 dd->ipath_sdma_descq_removed = dd->ipath_sdma_descq_added;
298
299                 /* Reset SendDmaLenGen */
300                 ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen,
301                         (u64) dd->ipath_sdma_descq_cnt | (1ULL << 18));
302
303                 /* done with sdma state for a bit */
304                 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
305
306                 /*
307                  * Don't restart sdma here (with the exception
308                  * below). Wait until link is up to ACTIVE.  VL15 MADs
309                  * used to bring the link up use PIO, and multiple link
310                  * transitions otherwise cause the sdma engine to be
311                  * stopped and started multiple times.
312                  * The disable is done here, including the shadow,
313                  * so the state is kept consistent.
314                  * See ipath_restart_sdma() for the actual starting
315                  * of sdma.
316                  */
317                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
318                 dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
319                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
320                                  dd->ipath_sendctrl);
321                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
322                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
323
324                 /* make sure I see next message */
325                 dd->ipath_sdma_abort_jiffies = 0;
326
327                 /*
328                  * Not everything that takes SDMA offline is a link
329                  * status change.  If the link was up, restart SDMA.
330                  */
331                 if (dd->ipath_flags & IPATH_LINKACTIVE)
332                         ipath_restart_sdma(dd);
333
334                 goto done;
335         }
336
337 resched:
338         /*
339          * for now, keep spinning
340          * JAG - this is bad to just have default be a loop without
341          * state change
342          */
343         if (jiffies > dd->ipath_sdma_abort_jiffies) {
344                 ipath_dbg("looping with status 0x%08lx\n",
345                           dd->ipath_sdma_status);
346                 dd->ipath_sdma_abort_jiffies = jiffies + 5 * HZ;
347         }
348 resched_noprint:
349         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
350         if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
351                 tasklet_hi_schedule(&dd->ipath_sdma_abort_task);
352         return;
353
354 unlock:
355         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
356 done:
357         return;
358 }
359
360 /*
361  * This is called from interrupt context.
362  */
363 void ipath_sdma_intr(struct ipath_devdata *dd)
364 {
365         unsigned long flags;
366
367         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
368
369         (void) ipath_sdma_make_progress(dd);
370
371         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
372 }
373
374 static int alloc_sdma(struct ipath_devdata *dd)
375 {
376         int ret = 0;
377
378         /* Allocate memory for SendDMA descriptor FIFO */
379         dd->ipath_sdma_descq = dma_alloc_coherent(&dd->pcidev->dev,
380                 SDMA_DESCQ_SZ, &dd->ipath_sdma_descq_phys, GFP_KERNEL);
381
382         if (!dd->ipath_sdma_descq) {
383                 ipath_dev_err(dd, "failed to allocate SendDMA descriptor "
384                         "FIFO memory\n");
385                 ret = -ENOMEM;
386                 goto done;
387         }
388
389         dd->ipath_sdma_descq_cnt =
390                 SDMA_DESCQ_SZ / sizeof(struct ipath_sdma_desc);
391
392         /* Allocate memory for DMA of head register to memory */
393         dd->ipath_sdma_head_dma = dma_alloc_coherent(&dd->pcidev->dev,
394                 PAGE_SIZE, &dd->ipath_sdma_head_phys, GFP_KERNEL);
395         if (!dd->ipath_sdma_head_dma) {
396                 ipath_dev_err(dd, "failed to allocate SendDMA head memory\n");
397                 ret = -ENOMEM;
398                 goto cleanup_descq;
399         }
400         dd->ipath_sdma_head_dma[0] = 0;
401
402         init_timer(&dd->ipath_sdma_vl15_timer);
403         dd->ipath_sdma_vl15_timer.function = vl15_watchdog_timeout;
404         dd->ipath_sdma_vl15_timer.data = (unsigned long)dd;
405         atomic_set(&dd->ipath_sdma_vl15_count, 0);
406
407         goto done;
408
409 cleanup_descq:
410         dma_free_coherent(&dd->pcidev->dev, SDMA_DESCQ_SZ,
411                 (void *)dd->ipath_sdma_descq, dd->ipath_sdma_descq_phys);
412         dd->ipath_sdma_descq = NULL;
413         dd->ipath_sdma_descq_phys = 0;
414 done:
415         return ret;
416 }
417
418 int setup_sdma(struct ipath_devdata *dd)
419 {
420         int ret = 0;
421         unsigned i, n;
422         u64 tmp64;
423         u64 senddmabufmask[3] = { 0 };
424         unsigned long flags;
425
426         ret = alloc_sdma(dd);
427         if (ret)
428                 goto done;
429
430         if (!dd->ipath_sdma_descq) {
431                 ipath_dev_err(dd, "SendDMA memory not allocated\n");
432                 goto done;
433         }
434
435         /*
436          * Set initial status as if we had been up, then gone down.
437          * This lets initial start on transition to ACTIVE be the
438          * same as restart after link flap.
439          */
440         dd->ipath_sdma_status = IPATH_SDMA_ABORT_ABORTED;
441         dd->ipath_sdma_abort_jiffies = 0;
442         dd->ipath_sdma_generation = 0;
443         dd->ipath_sdma_descq_tail = 0;
444         dd->ipath_sdma_descq_head = 0;
445         dd->ipath_sdma_descq_removed = 0;
446         dd->ipath_sdma_descq_added = 0;
447
448         /* Set SendDmaBase */
449         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabase,
450                          dd->ipath_sdma_descq_phys);
451         /* Set SendDmaLenGen */
452         tmp64 = dd->ipath_sdma_descq_cnt;
453         tmp64 |= 1<<18; /* enable generation checking */
454         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen, tmp64);
455         /* Set SendDmaTail */
456         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail,
457                          dd->ipath_sdma_descq_tail);
458         /* Set SendDmaHeadAddr */
459         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmaheadaddr,
460                          dd->ipath_sdma_head_phys);
461
462         /*
463          * Reserve all the former "kernel" piobufs, using high number range
464          * so we get as many 4K buffers as possible
465          */
466         n = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
467         i = dd->ipath_lastport_piobuf + dd->ipath_pioreserved;
468         ipath_chg_pioavailkernel(dd, i, n - i , 0);
469         for (; i < n; ++i) {
470                 unsigned word = i / 64;
471                 unsigned bit = i & 63;
472                 BUG_ON(word >= 3);
473                 senddmabufmask[word] |= 1ULL << bit;
474         }
475         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask0,
476                          senddmabufmask[0]);
477         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask1,
478                          senddmabufmask[1]);
479         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask2,
480                          senddmabufmask[2]);
481
482         INIT_LIST_HEAD(&dd->ipath_sdma_activelist);
483         INIT_LIST_HEAD(&dd->ipath_sdma_notifylist);
484
485         tasklet_init(&dd->ipath_sdma_notify_task, sdma_notify_task,
486                      (unsigned long) dd);
487         tasklet_init(&dd->ipath_sdma_abort_task, sdma_abort_task,
488                      (unsigned long) dd);
489
490         /*
491          * No use to turn on SDMA here, as link is probably not ACTIVE
492          * Just mark it RUNNING and enable the interrupt, and let the
493          * ipath_restart_sdma() on link transition to ACTIVE actually
494          * enable it.
495          */
496         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
497         dd->ipath_sendctrl |= INFINIPATH_S_SDMAINTENABLE;
498         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
499         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
500         __set_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status);
501         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
502
503 done:
504         return ret;
505 }
506
507 void teardown_sdma(struct ipath_devdata *dd)
508 {
509         struct ipath_sdma_txreq *txp, *txpnext;
510         unsigned long flags;
511         dma_addr_t sdma_head_phys = 0;
512         dma_addr_t sdma_descq_phys = 0;
513         void *sdma_descq = NULL;
514         void *sdma_head_dma = NULL;
515
516         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
517         __clear_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status);
518         __set_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
519         __set_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status);
520         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
521
522         tasklet_kill(&dd->ipath_sdma_abort_task);
523         tasklet_kill(&dd->ipath_sdma_notify_task);
524
525         /* turn off sdma */
526         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
527         dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
528         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
529                 dd->ipath_sendctrl);
530         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
531         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
532
533         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
534         /* dequeue all "sent" requests */
535         list_for_each_entry_safe(txp, txpnext, &dd->ipath_sdma_activelist,
536                                  list) {
537                 txp->callback_status = IPATH_SDMA_TXREQ_S_SHUTDOWN;
538                 if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
539                         vl15_watchdog_deq(dd);
540                 list_move_tail(&txp->list, &dd->ipath_sdma_notifylist);
541         }
542         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
543
544         sdma_notify_taskbody(dd);
545
546         del_timer_sync(&dd->ipath_sdma_vl15_timer);
547
548         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
549
550         dd->ipath_sdma_abort_jiffies = 0;
551
552         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabase, 0);
553         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen, 0);
554         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, 0);
555         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmaheadaddr, 0);
556         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask0, 0);
557         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask1, 0);
558         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask2, 0);
559
560         if (dd->ipath_sdma_head_dma) {
561                 sdma_head_dma = (void *) dd->ipath_sdma_head_dma;
562                 sdma_head_phys = dd->ipath_sdma_head_phys;
563                 dd->ipath_sdma_head_dma = NULL;
564                 dd->ipath_sdma_head_phys = 0;
565         }
566
567         if (dd->ipath_sdma_descq) {
568                 sdma_descq = dd->ipath_sdma_descq;
569                 sdma_descq_phys = dd->ipath_sdma_descq_phys;
570                 dd->ipath_sdma_descq = NULL;
571                 dd->ipath_sdma_descq_phys = 0;
572         }
573
574         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
575
576         if (sdma_head_dma)
577                 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
578                                   sdma_head_dma, sdma_head_phys);
579
580         if (sdma_descq)
581                 dma_free_coherent(&dd->pcidev->dev, SDMA_DESCQ_SZ,
582                                   sdma_descq, sdma_descq_phys);
583 }
584
585 /*
586  * [Re]start SDMA, if we use it, and it's not already OK.
587  * This is called on transition to link ACTIVE, either the first or
588  * subsequent times.
589  */
590 void ipath_restart_sdma(struct ipath_devdata *dd)
591 {
592         unsigned long flags;
593         int needed = 1;
594
595         if (!(dd->ipath_flags & IPATH_HAS_SEND_DMA))
596                 goto bail;
597
598         /*
599          * First, make sure we should, which is to say,
600          * check that we are "RUNNING" (not in teardown)
601          * and not "SHUTDOWN"
602          */
603         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
604         if (!test_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status)
605                 || test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
606                         needed = 0;
607         else {
608                 __clear_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
609                 __clear_bit(IPATH_SDMA_DISARMED, &dd->ipath_sdma_status);
610                 __clear_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
611         }
612         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
613         if (!needed) {
614                 ipath_dbg("invalid attempt to restart SDMA, status 0x%08lx\n",
615                         dd->ipath_sdma_status);
616                 goto bail;
617         }
618         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
619         /*
620          * First clear, just to be safe. Enable is only done
621          * in chip on 0->1 transition
622          */
623         dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
624         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
625         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
626         dd->ipath_sendctrl |= INFINIPATH_S_SDMAENABLE;
627         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
628         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
629         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
630
631         /* notify upper layers */
632         ipath_ib_piobufavail(dd->verbs_dev);
633
634 bail:
635         return;
636 }
637
638 static inline void make_sdma_desc(struct ipath_devdata *dd,
639         u64 *sdmadesc, u64 addr, u64 dwlen, u64 dwoffset)
640 {
641         WARN_ON(addr & 3);
642         /* SDmaPhyAddr[47:32] */
643         sdmadesc[1] = addr >> 32;
644         /* SDmaPhyAddr[31:0] */
645         sdmadesc[0] = (addr & 0xfffffffcULL) << 32;
646         /* SDmaGeneration[1:0] */
647         sdmadesc[0] |= (dd->ipath_sdma_generation & 3ULL) << 30;
648         /* SDmaDwordCount[10:0] */
649         sdmadesc[0] |= (dwlen & 0x7ffULL) << 16;
650         /* SDmaBufOffset[12:2] */
651         sdmadesc[0] |= dwoffset & 0x7ffULL;
652 }
653
654 /*
655  * This function queues one IB packet onto the send DMA queue per call.
656  * The caller is responsible for checking:
657  * 1) The number of send DMA descriptor entries is less than the size of
658  *    the descriptor queue.
659  * 2) The IB SGE addresses and lengths are 32-bit aligned
660  *    (except possibly the last SGE's length)
661  * 3) The SGE addresses are suitable for passing to dma_map_single().
662  */
663 int ipath_sdma_verbs_send(struct ipath_devdata *dd,
664         struct ipath_sge_state *ss, u32 dwords,
665         struct ipath_verbs_txreq *tx)
666 {
667
668         unsigned long flags;
669         struct ipath_sge *sge;
670         int ret = 0;
671         u16 tail;
672         __le64 *descqp;
673         u64 sdmadesc[2];
674         u32 dwoffset;
675         dma_addr_t addr;
676
677         if ((tx->map_len + (dwords<<2)) > dd->ipath_ibmaxlen) {
678                 ipath_dbg("packet size %X > ibmax %X, fail\n",
679                         tx->map_len + (dwords<<2), dd->ipath_ibmaxlen);
680                 ret = -EMSGSIZE;
681                 goto fail;
682         }
683
684         spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
685
686 retry:
687         if (unlikely(test_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status))) {
688                 ret = -EBUSY;
689                 goto unlock;
690         }
691
692         if (tx->txreq.sg_count > ipath_sdma_descq_freecnt(dd)) {
693                 if (ipath_sdma_make_progress(dd))
694                         goto retry;
695                 ret = -ENOBUFS;
696                 goto unlock;
697         }
698
699         addr = dma_map_single(&dd->pcidev->dev, tx->txreq.map_addr,
700                               tx->map_len, DMA_TO_DEVICE);
701         if (dma_mapping_error(&dd->pcidev->dev, addr)) {
702                 ret = -EIO;
703                 goto unlock;
704         }
705
706         dwoffset = tx->map_len >> 2;
707         make_sdma_desc(dd, sdmadesc, (u64) addr, dwoffset, 0);
708
709         /* SDmaFirstDesc */
710         sdmadesc[0] |= 1ULL << 12;
711         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_USELARGEBUF)
712                 sdmadesc[0] |= 1ULL << 14;      /* SDmaUseLargeBuf */
713
714         /* write to the descq */
715         tail = dd->ipath_sdma_descq_tail;
716         descqp = &dd->ipath_sdma_descq[tail].qw[0];
717         *descqp++ = cpu_to_le64(sdmadesc[0]);
718         *descqp++ = cpu_to_le64(sdmadesc[1]);
719
720         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEDESC)
721                 tx->txreq.start_idx = tail;
722
723         /* increment the tail */
724         if (++tail == dd->ipath_sdma_descq_cnt) {
725                 tail = 0;
726                 descqp = &dd->ipath_sdma_descq[0].qw[0];
727                 ++dd->ipath_sdma_generation;
728         }
729
730         sge = &ss->sge;
731         while (dwords) {
732                 u32 dw;
733                 u32 len;
734
735                 len = dwords << 2;
736                 if (len > sge->length)
737                         len = sge->length;
738                 if (len > sge->sge_length)
739                         len = sge->sge_length;
740                 BUG_ON(len == 0);
741                 dw = (len + 3) >> 2;
742                 addr = dma_map_single(&dd->pcidev->dev, sge->vaddr, dw << 2,
743                                       DMA_TO_DEVICE);
744                 make_sdma_desc(dd, sdmadesc, (u64) addr, dw, dwoffset);
745                 /* SDmaUseLargeBuf has to be set in every descriptor */
746                 if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_USELARGEBUF)
747                         sdmadesc[0] |= 1ULL << 14;
748                 /* write to the descq */
749                 *descqp++ = cpu_to_le64(sdmadesc[0]);
750                 *descqp++ = cpu_to_le64(sdmadesc[1]);
751
752                 /* increment the tail */
753                 if (++tail == dd->ipath_sdma_descq_cnt) {
754                         tail = 0;
755                         descqp = &dd->ipath_sdma_descq[0].qw[0];
756                         ++dd->ipath_sdma_generation;
757                 }
758                 sge->vaddr += len;
759                 sge->length -= len;
760                 sge->sge_length -= len;
761                 if (sge->sge_length == 0) {
762                         if (--ss->num_sge)
763                                 *sge = *ss->sg_list++;
764                 } else if (sge->length == 0 && sge->mr != NULL) {
765                         if (++sge->n >= IPATH_SEGSZ) {
766                                 if (++sge->m >= sge->mr->mapsz)
767                                         break;
768                                 sge->n = 0;
769                         }
770                         sge->vaddr =
771                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
772                         sge->length =
773                                 sge->mr->map[sge->m]->segs[sge->n].length;
774                 }
775
776                 dwoffset += dw;
777                 dwords -= dw;
778         }
779
780         if (!tail)
781                 descqp = &dd->ipath_sdma_descq[dd->ipath_sdma_descq_cnt].qw[0];
782         descqp -= 2;
783         /* SDmaLastDesc */
784         descqp[0] |= __constant_cpu_to_le64(1ULL << 11);
785         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_INTREQ) {
786                 /* SDmaIntReq */
787                 descqp[0] |= __constant_cpu_to_le64(1ULL << 15);
788         }
789
790         /* Commit writes to memory and advance the tail on the chip */
791         wmb();
792         ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, tail);
793
794         tx->txreq.next_descq_idx = tail;
795         tx->txreq.callback_status = IPATH_SDMA_TXREQ_S_OK;
796         dd->ipath_sdma_descq_tail = tail;
797         dd->ipath_sdma_descq_added += tx->txreq.sg_count;
798         list_add_tail(&tx->txreq.list, &dd->ipath_sdma_activelist);
799         if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_VL15)
800                 vl15_watchdog_enq(dd);
801
802 unlock:
803         spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
804 fail:
805         return ret;
806 }