diff options
Diffstat (limited to 'drivers/infiniband/hw/qib/qib_sdma.c')
-rw-r--r-- | drivers/infiniband/hw/qib/qib_sdma.c | 973 |
1 files changed, 973 insertions, 0 deletions
diff --git a/drivers/infiniband/hw/qib/qib_sdma.c b/drivers/infiniband/hw/qib/qib_sdma.c new file mode 100644 index 000000000000..b8456881f7f6 --- /dev/null +++ b/drivers/infiniband/hw/qib/qib_sdma.c | |||
@@ -0,0 +1,973 @@ | |||
1 | /* | ||
2 | * Copyright (c) 2007, 2008, 2009, 2010 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 | #include <linux/netdevice.h> | ||
35 | |||
36 | #include "qib.h" | ||
37 | #include "qib_common.h" | ||
38 | |||
39 | /* default pio off, sdma on */ | ||
40 | static ushort sdma_descq_cnt = 256; | ||
41 | module_param_named(sdma_descq_cnt, sdma_descq_cnt, ushort, S_IRUGO); | ||
42 | MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries"); | ||
43 | |||
44 | /* | ||
45 | * Bits defined in the send DMA descriptor. | ||
46 | */ | ||
47 | #define SDMA_DESC_LAST (1ULL << 11) | ||
48 | #define SDMA_DESC_FIRST (1ULL << 12) | ||
49 | #define SDMA_DESC_DMA_HEAD (1ULL << 13) | ||
50 | #define SDMA_DESC_USE_LARGE_BUF (1ULL << 14) | ||
51 | #define SDMA_DESC_INTR (1ULL << 15) | ||
52 | #define SDMA_DESC_COUNT_LSB 16 | ||
53 | #define SDMA_DESC_GEN_LSB 30 | ||
54 | |||
55 | char *qib_sdma_state_names[] = { | ||
56 | [qib_sdma_state_s00_hw_down] = "s00_HwDown", | ||
57 | [qib_sdma_state_s10_hw_start_up_wait] = "s10_HwStartUpWait", | ||
58 | [qib_sdma_state_s20_idle] = "s20_Idle", | ||
59 | [qib_sdma_state_s30_sw_clean_up_wait] = "s30_SwCleanUpWait", | ||
60 | [qib_sdma_state_s40_hw_clean_up_wait] = "s40_HwCleanUpWait", | ||
61 | [qib_sdma_state_s50_hw_halt_wait] = "s50_HwHaltWait", | ||
62 | [qib_sdma_state_s99_running] = "s99_Running", | ||
63 | }; | ||
64 | |||
65 | char *qib_sdma_event_names[] = { | ||
66 | [qib_sdma_event_e00_go_hw_down] = "e00_GoHwDown", | ||
67 | [qib_sdma_event_e10_go_hw_start] = "e10_GoHwStart", | ||
68 | [qib_sdma_event_e20_hw_started] = "e20_HwStarted", | ||
69 | [qib_sdma_event_e30_go_running] = "e30_GoRunning", | ||
70 | [qib_sdma_event_e40_sw_cleaned] = "e40_SwCleaned", | ||
71 | [qib_sdma_event_e50_hw_cleaned] = "e50_HwCleaned", | ||
72 | [qib_sdma_event_e60_hw_halted] = "e60_HwHalted", | ||
73 | [qib_sdma_event_e70_go_idle] = "e70_GoIdle", | ||
74 | [qib_sdma_event_e7220_err_halted] = "e7220_ErrHalted", | ||
75 | [qib_sdma_event_e7322_err_halted] = "e7322_ErrHalted", | ||
76 | [qib_sdma_event_e90_timer_tick] = "e90_TimerTick", | ||
77 | }; | ||
78 | |||
79 | /* declare all statics here rather than keep sorting */ | ||
80 | static int alloc_sdma(struct qib_pportdata *); | ||
81 | static void sdma_complete(struct kref *); | ||
82 | static void sdma_finalput(struct qib_sdma_state *); | ||
83 | static void sdma_get(struct qib_sdma_state *); | ||
84 | static void sdma_put(struct qib_sdma_state *); | ||
85 | static void sdma_set_state(struct qib_pportdata *, enum qib_sdma_states); | ||
86 | static void sdma_start_sw_clean_up(struct qib_pportdata *); | ||
87 | static void sdma_sw_clean_up_task(unsigned long); | ||
88 | static void unmap_desc(struct qib_pportdata *, unsigned); | ||
89 | |||
90 | static void sdma_get(struct qib_sdma_state *ss) | ||
91 | { | ||
92 | kref_get(&ss->kref); | ||
93 | } | ||
94 | |||
95 | static void sdma_complete(struct kref *kref) | ||
96 | { | ||
97 | struct qib_sdma_state *ss = | ||
98 | container_of(kref, struct qib_sdma_state, kref); | ||
99 | |||
100 | complete(&ss->comp); | ||
101 | } | ||
102 | |||
103 | static void sdma_put(struct qib_sdma_state *ss) | ||
104 | { | ||
105 | kref_put(&ss->kref, sdma_complete); | ||
106 | } | ||
107 | |||
108 | static void sdma_finalput(struct qib_sdma_state *ss) | ||
109 | { | ||
110 | sdma_put(ss); | ||
111 | wait_for_completion(&ss->comp); | ||
112 | } | ||
113 | |||
114 | /* | ||
115 | * Complete all the sdma requests on the active list, in the correct | ||
116 | * order, and with appropriate processing. Called when cleaning up | ||
117 | * after sdma shutdown, and when new sdma requests are submitted for | ||
118 | * a link that is down. This matches what is done for requests | ||
119 | * that complete normally, it's just the full list. | ||
120 | * | ||
121 | * Must be called with sdma_lock held | ||
122 | */ | ||
123 | static void clear_sdma_activelist(struct qib_pportdata *ppd) | ||
124 | { | ||
125 | struct qib_sdma_txreq *txp, *txp_next; | ||
126 | |||
127 | list_for_each_entry_safe(txp, txp_next, &ppd->sdma_activelist, list) { | ||
128 | list_del_init(&txp->list); | ||
129 | if (txp->flags & QIB_SDMA_TXREQ_F_FREEDESC) { | ||
130 | unsigned idx; | ||
131 | |||
132 | idx = txp->start_idx; | ||
133 | while (idx != txp->next_descq_idx) { | ||
134 | unmap_desc(ppd, idx); | ||
135 | if (++idx == ppd->sdma_descq_cnt) | ||
136 | idx = 0; | ||
137 | } | ||
138 | } | ||
139 | if (txp->callback) | ||
140 | (*txp->callback)(txp, QIB_SDMA_TXREQ_S_ABORTED); | ||
141 | } | ||
142 | } | ||
143 | |||
144 | static void sdma_sw_clean_up_task(unsigned long opaque) | ||
145 | { | ||
146 | struct qib_pportdata *ppd = (struct qib_pportdata *) opaque; | ||
147 | unsigned long flags; | ||
148 | |||
149 | spin_lock_irqsave(&ppd->sdma_lock, flags); | ||
150 | |||
151 | /* | ||
152 | * At this point, the following should always be true: | ||
153 | * - We are halted, so no more descriptors are getting retired. | ||
154 | * - We are not running, so no one is submitting new work. | ||
155 | * - Only we can send the e40_sw_cleaned, so we can't start | ||
156 | * running again until we say so. So, the active list and | ||
157 | * descq are ours to play with. | ||
158 | */ | ||
159 | |||
160 | /* Process all retired requests. */ | ||
161 | qib_sdma_make_progress(ppd); | ||
162 | |||
163 | clear_sdma_activelist(ppd); | ||
164 | |||
165 | /* | ||
166 | * Resync count of added and removed. It is VERY important that | ||
167 | * sdma_descq_removed NEVER decrement - user_sdma depends on it. | ||
168 | */ | ||
169 | ppd->sdma_descq_removed = ppd->sdma_descq_added; | ||
170 | |||
171 | /* | ||
172 | * Reset our notion of head and tail. | ||
173 | * Note that the HW registers will be reset when switching states | ||
174 | * due to calling __qib_sdma_process_event() below. | ||
175 | */ | ||
176 | ppd->sdma_descq_tail = 0; | ||
177 | ppd->sdma_descq_head = 0; | ||
178 | ppd->sdma_head_dma[0] = 0; | ||
179 | ppd->sdma_generation = 0; | ||
180 | |||
181 | __qib_sdma_process_event(ppd, qib_sdma_event_e40_sw_cleaned); | ||
182 | |||
183 | spin_unlock_irqrestore(&ppd->sdma_lock, flags); | ||
184 | } | ||
185 | |||
186 | /* | ||
187 | * This is called when changing to state qib_sdma_state_s10_hw_start_up_wait | ||
188 | * as a result of send buffer errors or send DMA descriptor errors. | ||
189 | * We want to disarm the buffers in these cases. | ||
190 | */ | ||
191 | static void sdma_hw_start_up(struct qib_pportdata *ppd) | ||
192 | { | ||
193 | struct qib_sdma_state *ss = &ppd->sdma_state; | ||
194 | unsigned bufno; | ||
195 | |||
196 | for (bufno = ss->first_sendbuf; bufno < ss->last_sendbuf; ++bufno) | ||
197 | ppd->dd->f_sendctrl(ppd, QIB_SENDCTRL_DISARM_BUF(bufno)); | ||
198 | |||
199 | ppd->dd->f_sdma_hw_start_up(ppd); | ||
200 | } | ||
201 | |||
202 | static void sdma_sw_tear_down(struct qib_pportdata *ppd) | ||
203 | { | ||
204 | struct qib_sdma_state *ss = &ppd->sdma_state; | ||
205 | |||
206 | /* Releasing this reference means the state machine has stopped. */ | ||
207 | sdma_put(ss); | ||
208 | } | ||
209 | |||
210 | static void sdma_start_sw_clean_up(struct qib_pportdata *ppd) | ||
211 | { | ||
212 | tasklet_hi_schedule(&ppd->sdma_sw_clean_up_task); | ||
213 | } | ||
214 | |||
215 | static void sdma_set_state(struct qib_pportdata *ppd, | ||
216 | enum qib_sdma_states next_state) | ||
217 | { | ||
218 | struct qib_sdma_state *ss = &ppd->sdma_state; | ||
219 | struct sdma_set_state_action *action = ss->set_state_action; | ||
220 | unsigned op = 0; | ||
221 | |||
222 | /* debugging bookkeeping */ | ||
223 | ss->previous_state = ss->current_state; | ||
224 | ss->previous_op = ss->current_op; | ||
225 | |||
226 | ss->current_state = next_state; | ||
227 | |||
228 | if (action[next_state].op_enable) | ||
229 | op |= QIB_SDMA_SENDCTRL_OP_ENABLE; | ||
230 | |||
231 | if (action[next_state].op_intenable) | ||
232 | op |= QIB_SDMA_SENDCTRL_OP_INTENABLE; | ||
233 | |||
234 | if (action[next_state].op_halt) | ||
235 | op |= QIB_SDMA_SENDCTRL_OP_HALT; | ||
236 | |||
237 | if (action[next_state].op_drain) | ||
238 | op |= QIB_SDMA_SENDCTRL_OP_DRAIN; | ||
239 | |||
240 | if (action[next_state].go_s99_running_tofalse) | ||
241 | ss->go_s99_running = 0; | ||
242 | |||
243 | if (action[next_state].go_s99_running_totrue) | ||
244 | ss->go_s99_running = 1; | ||
245 | |||
246 | ss->current_op = op; | ||
247 | |||
248 | ppd->dd->f_sdma_sendctrl(ppd, ss->current_op); | ||
249 | } | ||
250 | |||
251 | static void unmap_desc(struct qib_pportdata *ppd, unsigned head) | ||
252 | { | ||
253 | __le64 *descqp = &ppd->sdma_descq[head].qw[0]; | ||
254 | u64 desc[2]; | ||
255 | dma_addr_t addr; | ||
256 | size_t len; | ||
257 | |||
258 | desc[0] = le64_to_cpu(descqp[0]); | ||
259 | desc[1] = le64_to_cpu(descqp[1]); | ||
260 | |||
261 | addr = (desc[1] << 32) | (desc[0] >> 32); | ||
262 | len = (desc[0] >> 14) & (0x7ffULL << 2); | ||
263 | dma_unmap_single(&ppd->dd->pcidev->dev, addr, len, DMA_TO_DEVICE); | ||
264 | } | ||
265 | |||
266 | static int alloc_sdma(struct qib_pportdata *ppd) | ||
267 | { | ||
268 | ppd->sdma_descq_cnt = sdma_descq_cnt; | ||
269 | if (!ppd->sdma_descq_cnt) | ||
270 | ppd->sdma_descq_cnt = 256; | ||
271 | |||
272 | /* Allocate memory for SendDMA descriptor FIFO */ | ||
273 | ppd->sdma_descq = dma_alloc_coherent(&ppd->dd->pcidev->dev, | ||
274 | ppd->sdma_descq_cnt * sizeof(u64[2]), &ppd->sdma_descq_phys, | ||
275 | GFP_KERNEL); | ||
276 | |||
277 | if (!ppd->sdma_descq) { | ||
278 | qib_dev_err(ppd->dd, "failed to allocate SendDMA descriptor " | ||
279 | "FIFO memory\n"); | ||
280 | goto bail; | ||
281 | } | ||
282 | |||
283 | /* Allocate memory for DMA of head register to memory */ | ||
284 | ppd->sdma_head_dma = dma_alloc_coherent(&ppd->dd->pcidev->dev, | ||
285 | PAGE_SIZE, &ppd->sdma_head_phys, GFP_KERNEL); | ||
286 | if (!ppd->sdma_head_dma) { | ||
287 | qib_dev_err(ppd->dd, "failed to allocate SendDMA " | ||
288 | "head memory\n"); | ||
289 | goto cleanup_descq; | ||
290 | } | ||
291 | ppd->sdma_head_dma[0] = 0; | ||
292 | return 0; | ||
293 | |||
294 | cleanup_descq: | ||
295 | dma_free_coherent(&ppd->dd->pcidev->dev, | ||
296 | ppd->sdma_descq_cnt * sizeof(u64[2]), (void *)ppd->sdma_descq, | ||
297 | ppd->sdma_descq_phys); | ||
298 | ppd->sdma_descq = NULL; | ||
299 | ppd->sdma_descq_phys = 0; | ||
300 | bail: | ||
301 | ppd->sdma_descq_cnt = 0; | ||
302 | return -ENOMEM; | ||
303 | } | ||
304 | |||
305 | static void free_sdma(struct qib_pportdata *ppd) | ||
306 | { | ||
307 | struct qib_devdata *dd = ppd->dd; | ||
308 | |||
309 | if (ppd->sdma_head_dma) { | ||
310 | dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE, | ||
311 | (void *)ppd->sdma_head_dma, | ||
312 | ppd->sdma_head_phys); | ||
313 | ppd->sdma_head_dma = NULL; | ||
314 | ppd->sdma_head_phys = 0; | ||
315 | } | ||
316 | |||
317 | if (ppd->sdma_descq) { | ||
318 | dma_free_coherent(&dd->pcidev->dev, | ||
319 | ppd->sdma_descq_cnt * sizeof(u64[2]), | ||
320 | ppd->sdma_descq, ppd->sdma_descq_phys); | ||
321 | ppd->sdma_descq = NULL; | ||
322 | ppd->sdma_descq_phys = 0; | ||
323 | } | ||
324 | } | ||
325 | |||
326 | static inline void make_sdma_desc(struct qib_pportdata *ppd, | ||
327 | u64 *sdmadesc, u64 addr, u64 dwlen, | ||
328 | u64 dwoffset) | ||
329 | { | ||
330 | |||
331 | WARN_ON(addr & 3); | ||
332 | /* SDmaPhyAddr[47:32] */ | ||
333 | sdmadesc[1] = addr >> 32; | ||
334 | /* SDmaPhyAddr[31:0] */ | ||
335 | sdmadesc[0] = (addr & 0xfffffffcULL) << 32; | ||
336 | /* SDmaGeneration[1:0] */ | ||
337 | sdmadesc[0] |= (ppd->sdma_generation & 3ULL) << | ||
338 | SDMA_DESC_GEN_LSB; | ||
339 | /* SDmaDwordCount[10:0] */ | ||
340 | sdmadesc[0] |= (dwlen & 0x7ffULL) << SDMA_DESC_COUNT_LSB; | ||
341 | /* SDmaBufOffset[12:2] */ | ||
342 | sdmadesc[0] |= dwoffset & 0x7ffULL; | ||
343 | } | ||
344 | |||
345 | /* sdma_lock must be held */ | ||
346 | int qib_sdma_make_progress(struct qib_pportdata *ppd) | ||
347 | { | ||
348 | struct list_head *lp = NULL; | ||
349 | struct qib_sdma_txreq *txp = NULL; | ||
350 | struct qib_devdata *dd = ppd->dd; | ||
351 | int progress = 0; | ||
352 | u16 hwhead; | ||
353 | u16 idx = 0; | ||
354 | |||
355 | hwhead = dd->f_sdma_gethead(ppd); | ||
356 | |||
357 | /* The reason for some of the complexity of this code is that | ||
358 | * not all descriptors have corresponding txps. So, we have to | ||
359 | * be able to skip over descs until we wander into the range of | ||
360 | * the next txp on the list. | ||
361 | */ | ||
362 | |||
363 | if (!list_empty(&ppd->sdma_activelist)) { | ||
364 | lp = ppd->sdma_activelist.next; | ||
365 | txp = list_entry(lp, struct qib_sdma_txreq, list); | ||
366 | idx = txp->start_idx; | ||
367 | } | ||
368 | |||
369 | while (ppd->sdma_descq_head != hwhead) { | ||
370 | /* if desc is part of this txp, unmap if needed */ | ||
371 | if (txp && (txp->flags & QIB_SDMA_TXREQ_F_FREEDESC) && | ||
372 | (idx == ppd->sdma_descq_head)) { | ||
373 | unmap_desc(ppd, ppd->sdma_descq_head); | ||
374 | if (++idx == ppd->sdma_descq_cnt) | ||
375 | idx = 0; | ||
376 | } | ||
377 | |||
378 | /* increment dequed desc count */ | ||
379 | ppd->sdma_descq_removed++; | ||
380 | |||
381 | /* advance head, wrap if needed */ | ||
382 | if (++ppd->sdma_descq_head == ppd->sdma_descq_cnt) | ||
383 | ppd->sdma_descq_head = 0; | ||
384 | |||
385 | /* if now past this txp's descs, do the callback */ | ||
386 | if (txp && txp->next_descq_idx == ppd->sdma_descq_head) { | ||
387 | /* remove from active list */ | ||
388 | list_del_init(&txp->list); | ||
389 | if (txp->callback) | ||
390 | (*txp->callback)(txp, QIB_SDMA_TXREQ_S_OK); | ||
391 | /* see if there is another txp */ | ||
392 | if (list_empty(&ppd->sdma_activelist)) | ||
393 | txp = NULL; | ||
394 | else { | ||
395 | lp = ppd->sdma_activelist.next; | ||
396 | txp = list_entry(lp, struct qib_sdma_txreq, | ||
397 | list); | ||
398 | idx = txp->start_idx; | ||
399 | } | ||
400 | } | ||
401 | progress = 1; | ||
402 | } | ||
403 | if (progress) | ||
404 | qib_verbs_sdma_desc_avail(ppd, qib_sdma_descq_freecnt(ppd)); | ||
405 | return progress; | ||
406 | } | ||
407 | |||
408 | /* | ||
409 | * This is called from interrupt context. | ||
410 | */ | ||
411 | void qib_sdma_intr(struct qib_pportdata *ppd) | ||
412 | { | ||
413 | unsigned long flags; | ||
414 | |||
415 | spin_lock_irqsave(&ppd->sdma_lock, flags); | ||
416 | |||
417 | __qib_sdma_intr(ppd); | ||
418 | |||
419 | spin_unlock_irqrestore(&ppd->sdma_lock, flags); | ||
420 | } | ||
421 | |||
422 | void __qib_sdma_intr(struct qib_pportdata *ppd) | ||
423 | { | ||
424 | if (__qib_sdma_running(ppd)) | ||
425 | qib_sdma_make_progress(ppd); | ||
426 | } | ||
427 | |||
428 | int qib_setup_sdma(struct qib_pportdata *ppd) | ||
429 | { | ||
430 | struct qib_devdata *dd = ppd->dd; | ||
431 | unsigned long flags; | ||
432 | int ret = 0; | ||
433 | |||
434 | ret = alloc_sdma(ppd); | ||
435 | if (ret) | ||
436 | goto bail; | ||
437 | |||
438 | /* set consistent sdma state */ | ||
439 | ppd->dd->f_sdma_init_early(ppd); | ||
440 | spin_lock_irqsave(&ppd->sdma_lock, flags); | ||
441 | sdma_set_state(ppd, qib_sdma_state_s00_hw_down); | ||
442 | spin_unlock_irqrestore(&ppd->sdma_lock, flags); | ||
443 | |||
444 | /* set up reference counting */ | ||
445 | kref_init(&ppd->sdma_state.kref); | ||
446 | init_completion(&ppd->sdma_state.comp); | ||
447 | |||
448 | ppd->sdma_generation = 0; | ||
449 | ppd->sdma_descq_head = 0; | ||
450 | ppd->sdma_descq_removed = 0; | ||
451 | ppd->sdma_descq_added = 0; | ||
452 | |||
453 | INIT_LIST_HEAD(&ppd->sdma_activelist); | ||
454 | |||
455 | tasklet_init(&ppd->sdma_sw_clean_up_task, sdma_sw_clean_up_task, | ||
456 | (unsigned long)ppd); | ||
457 | |||
458 | ret = dd->f_init_sdma_regs(ppd); | ||
459 | if (ret) | ||
460 | goto bail_alloc; | ||
461 | |||
462 | qib_sdma_process_event(ppd, qib_sdma_event_e10_go_hw_start); | ||
463 | |||
464 | return 0; | ||
465 | |||
466 | bail_alloc: | ||
467 | qib_teardown_sdma(ppd); | ||
468 | bail: | ||
469 | return ret; | ||
470 | } | ||
471 | |||
472 | void qib_teardown_sdma(struct qib_pportdata *ppd) | ||
473 | { | ||
474 | qib_sdma_process_event(ppd, qib_sdma_event_e00_go_hw_down); | ||
475 | |||
476 | /* | ||
477 | * This waits for the state machine to exit so it is not | ||
478 | * necessary to kill the sdma_sw_clean_up_task to make sure | ||
479 | * it is not running. | ||
480 | */ | ||
481 | sdma_finalput(&ppd->sdma_state); | ||
482 | |||
483 | free_sdma(ppd); | ||
484 | } | ||
485 | |||
486 | int qib_sdma_running(struct qib_pportdata *ppd) | ||
487 | { | ||
488 | unsigned long flags; | ||
489 | int ret; | ||
490 | |||
491 | spin_lock_irqsave(&ppd->sdma_lock, flags); | ||
492 | ret = __qib_sdma_running(ppd); | ||
493 | spin_unlock_irqrestore(&ppd->sdma_lock, flags); | ||
494 | |||
495 | return ret; | ||
496 | } | ||
497 | |||
498 | /* | ||
499 | * Complete a request when sdma not running; likely only request | ||
500 | * but to simplify the code, always queue it, then process the full | ||
501 | * activelist. We process the entire list to ensure that this particular | ||
502 | * request does get it's callback, but in the correct order. | ||
503 | * Must be called with sdma_lock held | ||
504 | */ | ||
505 | static void complete_sdma_err_req(struct qib_pportdata *ppd, | ||
506 | struct qib_verbs_txreq *tx) | ||
507 | { | ||
508 | atomic_inc(&tx->qp->s_dma_busy); | ||
509 | /* no sdma descriptors, so no unmap_desc */ | ||
510 | tx->txreq.start_idx = 0; | ||
511 | tx->txreq.next_descq_idx = 0; | ||
512 | list_add_tail(&tx->txreq.list, &ppd->sdma_activelist); | ||
513 | clear_sdma_activelist(ppd); | ||
514 | } | ||
515 | |||
516 | /* | ||
517 | * This function queues one IB packet onto the send DMA queue per call. | ||
518 | * The caller is responsible for checking: | ||
519 | * 1) The number of send DMA descriptor entries is less than the size of | ||
520 | * the descriptor queue. | ||
521 | * 2) The IB SGE addresses and lengths are 32-bit aligned | ||
522 | * (except possibly the last SGE's length) | ||
523 | * 3) The SGE addresses are suitable for passing to dma_map_single(). | ||
524 | */ | ||
525 | int qib_sdma_verbs_send(struct qib_pportdata *ppd, | ||
526 | struct qib_sge_state *ss, u32 dwords, | ||
527 | struct qib_verbs_txreq *tx) | ||
528 | { | ||
529 | unsigned long flags; | ||
530 | struct qib_sge *sge; | ||
531 | struct qib_qp *qp; | ||
532 | int ret = 0; | ||
533 | u16 tail; | ||
534 | __le64 *descqp; | ||
535 | u64 sdmadesc[2]; | ||
536 | u32 dwoffset; | ||
537 | dma_addr_t addr; | ||
538 | |||
539 | spin_lock_irqsave(&ppd->sdma_lock, flags); | ||
540 | |||
541 | retry: | ||
542 | if (unlikely(!__qib_sdma_running(ppd))) { | ||
543 | complete_sdma_err_req(ppd, tx); | ||
544 | goto unlock; | ||
545 | } | ||
546 | |||
547 | if (tx->txreq.sg_count > qib_sdma_descq_freecnt(ppd)) { | ||
548 | if (qib_sdma_make_progress(ppd)) | ||
549 | goto retry; | ||
550 | if (ppd->dd->flags & QIB_HAS_SDMA_TIMEOUT) | ||
551 | ppd->dd->f_sdma_set_desc_cnt(ppd, | ||
552 | ppd->sdma_descq_cnt / 2); | ||
553 | goto busy; | ||
554 | } | ||
555 | |||
556 | dwoffset = tx->hdr_dwords; | ||
557 | make_sdma_desc(ppd, sdmadesc, (u64) tx->txreq.addr, dwoffset, 0); | ||
558 | |||
559 | sdmadesc[0] |= SDMA_DESC_FIRST; | ||
560 | if (tx->txreq.flags & QIB_SDMA_TXREQ_F_USELARGEBUF) | ||
561 | sdmadesc[0] |= SDMA_DESC_USE_LARGE_BUF; | ||
562 | |||
563 | /* write to the descq */ | ||
564 | tail = ppd->sdma_descq_tail; | ||
565 | descqp = &ppd->sdma_descq[tail].qw[0]; | ||
566 | *descqp++ = cpu_to_le64(sdmadesc[0]); | ||
567 | *descqp++ = cpu_to_le64(sdmadesc[1]); | ||
568 | |||
569 | /* increment the tail */ | ||
570 | if (++tail == ppd->sdma_descq_cnt) { | ||
571 | tail = 0; | ||
572 | descqp = &ppd->sdma_descq[0].qw[0]; | ||
573 | ++ppd->sdma_generation; | ||
574 | } | ||
575 | |||
576 | tx->txreq.start_idx = tail; | ||
577 | |||
578 | sge = &ss->sge; | ||
579 | while (dwords) { | ||
580 | u32 dw; | ||
581 | u32 len; | ||
582 | |||
583 | len = dwords << 2; | ||
584 | if (len > sge->length) | ||
585 | len = sge->length; | ||
586 | if (len > sge->sge_length) | ||
587 | len = sge->sge_length; | ||
588 | BUG_ON(len == 0); | ||
589 | dw = (len + 3) >> 2; | ||
590 | addr = dma_map_single(&ppd->dd->pcidev->dev, sge->vaddr, | ||
591 | dw << 2, DMA_TO_DEVICE); | ||
592 | if (dma_mapping_error(&ppd->dd->pcidev->dev, addr)) | ||
593 | goto unmap; | ||
594 | sdmadesc[0] = 0; | ||
595 | make_sdma_desc(ppd, sdmadesc, (u64) addr, dw, dwoffset); | ||
596 | /* SDmaUseLargeBuf has to be set in every descriptor */ | ||
597 | if (tx->txreq.flags & QIB_SDMA_TXREQ_F_USELARGEBUF) | ||
598 | sdmadesc[0] |= SDMA_DESC_USE_LARGE_BUF; | ||
599 | /* write to the descq */ | ||
600 | *descqp++ = cpu_to_le64(sdmadesc[0]); | ||
601 | *descqp++ = cpu_to_le64(sdmadesc[1]); | ||
602 | |||
603 | /* increment the tail */ | ||
604 | if (++tail == ppd->sdma_descq_cnt) { | ||
605 | tail = 0; | ||
606 | descqp = &ppd->sdma_descq[0].qw[0]; | ||
607 | ++ppd->sdma_generation; | ||
608 | } | ||
609 | sge->vaddr += len; | ||
610 | sge->length -= len; | ||
611 | sge->sge_length -= len; | ||
612 | if (sge->sge_length == 0) { | ||
613 | if (--ss->num_sge) | ||
614 | *sge = *ss->sg_list++; | ||
615 | } else if (sge->length == 0 && sge->mr->lkey) { | ||
616 | if (++sge->n >= QIB_SEGSZ) { | ||
617 | if (++sge->m >= sge->mr->mapsz) | ||
618 | break; | ||
619 | sge->n = 0; | ||
620 | } | ||
621 | sge->vaddr = | ||
622 | sge->mr->map[sge->m]->segs[sge->n].vaddr; | ||
623 | sge->length = | ||
624 | sge->mr->map[sge->m]->segs[sge->n].length; | ||
625 | } | ||
626 | |||
627 | dwoffset += dw; | ||
628 | dwords -= dw; | ||
629 | } | ||
630 | |||
631 | if (!tail) | ||
632 | descqp = &ppd->sdma_descq[ppd->sdma_descq_cnt].qw[0]; | ||
633 | descqp -= 2; | ||
634 | descqp[0] |= cpu_to_le64(SDMA_DESC_LAST); | ||
635 | if (tx->txreq.flags & QIB_SDMA_TXREQ_F_HEADTOHOST) | ||
636 | descqp[0] |= cpu_to_le64(SDMA_DESC_DMA_HEAD); | ||
637 | if (tx->txreq.flags & QIB_SDMA_TXREQ_F_INTREQ) | ||
638 | descqp[0] |= cpu_to_le64(SDMA_DESC_INTR); | ||
639 | |||
640 | atomic_inc(&tx->qp->s_dma_busy); | ||
641 | tx->txreq.next_descq_idx = tail; | ||
642 | ppd->dd->f_sdma_update_tail(ppd, tail); | ||
643 | ppd->sdma_descq_added += tx->txreq.sg_count; | ||
644 | list_add_tail(&tx->txreq.list, &ppd->sdma_activelist); | ||
645 | goto unlock; | ||
646 | |||
647 | unmap: | ||
648 | for (;;) { | ||
649 | if (!tail) | ||
650 | tail = ppd->sdma_descq_cnt - 1; | ||
651 | else | ||
652 | tail--; | ||
653 | if (tail == ppd->sdma_descq_tail) | ||
654 | break; | ||
655 | unmap_desc(ppd, tail); | ||
656 | } | ||
657 | qp = tx->qp; | ||
658 | qib_put_txreq(tx); | ||
659 | spin_lock(&qp->s_lock); | ||
660 | if (qp->ibqp.qp_type == IB_QPT_RC) { | ||
661 | /* XXX what about error sending RDMA read responses? */ | ||
662 | if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) | ||
663 | qib_error_qp(qp, IB_WC_GENERAL_ERR); | ||
664 | } else if (qp->s_wqe) | ||
665 | qib_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR); | ||
666 | spin_unlock(&qp->s_lock); | ||
667 | /* return zero to process the next send work request */ | ||
668 | goto unlock; | ||
669 | |||
670 | busy: | ||
671 | qp = tx->qp; | ||
672 | spin_lock(&qp->s_lock); | ||
673 | if (ib_qib_state_ops[qp->state] & QIB_PROCESS_RECV_OK) { | ||
674 | struct qib_ibdev *dev; | ||
675 | |||
676 | /* | ||
677 | * If we couldn't queue the DMA request, save the info | ||
678 | * and try again later rather than destroying the | ||
679 | * buffer and undoing the side effects of the copy. | ||
680 | */ | ||
681 | tx->ss = ss; | ||
682 | tx->dwords = dwords; | ||
683 | qp->s_tx = tx; | ||
684 | dev = &ppd->dd->verbs_dev; | ||
685 | spin_lock(&dev->pending_lock); | ||
686 | if (list_empty(&qp->iowait)) { | ||
687 | struct qib_ibport *ibp; | ||
688 | |||
689 | ibp = &ppd->ibport_data; | ||
690 | ibp->n_dmawait++; | ||
691 | qp->s_flags |= QIB_S_WAIT_DMA_DESC; | ||
692 | list_add_tail(&qp->iowait, &dev->dmawait); | ||
693 | } | ||
694 | spin_unlock(&dev->pending_lock); | ||
695 | qp->s_flags &= ~QIB_S_BUSY; | ||
696 | spin_unlock(&qp->s_lock); | ||
697 | ret = -EBUSY; | ||
698 | } else { | ||
699 | spin_unlock(&qp->s_lock); | ||
700 | qib_put_txreq(tx); | ||
701 | } | ||
702 | unlock: | ||
703 | spin_unlock_irqrestore(&ppd->sdma_lock, flags); | ||
704 | return ret; | ||
705 | } | ||
706 | |||
707 | void qib_sdma_process_event(struct qib_pportdata *ppd, | ||
708 | enum qib_sdma_events event) | ||
709 | { | ||
710 | unsigned long flags; | ||
711 | |||
712 | spin_lock_irqsave(&ppd->sdma_lock, flags); | ||
713 | |||
714 | __qib_sdma_process_event(ppd, event); | ||
715 | |||
716 | if (ppd->sdma_state.current_state == qib_sdma_state_s99_running) | ||
717 | qib_verbs_sdma_desc_avail(ppd, qib_sdma_descq_freecnt(ppd)); | ||
718 | |||
719 | spin_unlock_irqrestore(&ppd->sdma_lock, flags); | ||
720 | } | ||
721 | |||
722 | void __qib_sdma_process_event(struct qib_pportdata *ppd, | ||
723 | enum qib_sdma_events event) | ||
724 | { | ||
725 | struct qib_sdma_state *ss = &ppd->sdma_state; | ||
726 | |||
727 | switch (ss->current_state) { | ||
728 | case qib_sdma_state_s00_hw_down: | ||
729 | switch (event) { | ||
730 | case qib_sdma_event_e00_go_hw_down: | ||
731 | break; | ||
732 | case qib_sdma_event_e30_go_running: | ||
733 | /* | ||
734 | * If down, but running requested (usually result | ||
735 | * of link up, then we need to start up. | ||
736 | * This can happen when hw down is requested while | ||
737 | * bringing the link up with traffic active on | ||
738 | * 7220, e.g. */ | ||
739 | ss->go_s99_running = 1; | ||
740 | /* fall through and start dma engine */ | ||
741 | case qib_sdma_event_e10_go_hw_start: | ||
742 | /* This reference means the state machine is started */ | ||
743 | sdma_get(&ppd->sdma_state); | ||
744 | sdma_set_state(ppd, | ||
745 | qib_sdma_state_s10_hw_start_up_wait); | ||
746 | break; | ||
747 | case qib_sdma_event_e20_hw_started: | ||
748 | break; | ||
749 | case qib_sdma_event_e40_sw_cleaned: | ||
750 | sdma_sw_tear_down(ppd); | ||
751 | break; | ||
752 | case qib_sdma_event_e50_hw_cleaned: | ||
753 | break; | ||
754 | case qib_sdma_event_e60_hw_halted: | ||
755 | break; | ||
756 | case qib_sdma_event_e70_go_idle: | ||
757 | break; | ||
758 | case qib_sdma_event_e7220_err_halted: | ||
759 | break; | ||
760 | case qib_sdma_event_e7322_err_halted: | ||
761 | break; | ||
762 | case qib_sdma_event_e90_timer_tick: | ||
763 | break; | ||
764 | } | ||
765 | break; | ||
766 | |||
767 | case qib_sdma_state_s10_hw_start_up_wait: | ||
768 | switch (event) { | ||
769 | case qib_sdma_event_e00_go_hw_down: | ||
770 | sdma_set_state(ppd, qib_sdma_state_s00_hw_down); | ||
771 | sdma_sw_tear_down(ppd); | ||
772 | break; | ||
773 | case qib_sdma_event_e10_go_hw_start: | ||
774 | break; | ||
775 | case qib_sdma_event_e20_hw_started: | ||
776 | sdma_set_state(ppd, ss->go_s99_running ? | ||
777 | qib_sdma_state_s99_running : | ||
778 | qib_sdma_state_s20_idle); | ||
779 | break; | ||
780 | case qib_sdma_event_e30_go_running: | ||
781 | ss->go_s99_running = 1; | ||
782 | break; | ||
783 | case qib_sdma_event_e40_sw_cleaned: | ||
784 | break; | ||
785 | case qib_sdma_event_e50_hw_cleaned: | ||
786 | break; | ||
787 | case qib_sdma_event_e60_hw_halted: | ||
788 | break; | ||
789 | case qib_sdma_event_e70_go_idle: | ||
790 | ss->go_s99_running = 0; | ||
791 | break; | ||
792 | case qib_sdma_event_e7220_err_halted: | ||
793 | break; | ||
794 | case qib_sdma_event_e7322_err_halted: | ||
795 | break; | ||
796 | case qib_sdma_event_e90_timer_tick: | ||
797 | break; | ||
798 | } | ||
799 | break; | ||
800 | |||
801 | case qib_sdma_state_s20_idle: | ||
802 | switch (event) { | ||
803 | case qib_sdma_event_e00_go_hw_down: | ||
804 | sdma_set_state(ppd, qib_sdma_state_s00_hw_down); | ||
805 | sdma_sw_tear_down(ppd); | ||
806 | break; | ||
807 | case qib_sdma_event_e10_go_hw_start: | ||
808 | break; | ||
809 | case qib_sdma_event_e20_hw_started: | ||
810 | break; | ||
811 | case qib_sdma_event_e30_go_running: | ||
812 | sdma_set_state(ppd, qib_sdma_state_s99_running); | ||
813 | ss->go_s99_running = 1; | ||
814 | break; | ||
815 | case qib_sdma_event_e40_sw_cleaned: | ||
816 | break; | ||
817 | case qib_sdma_event_e50_hw_cleaned: | ||
818 | break; | ||
819 | case qib_sdma_event_e60_hw_halted: | ||
820 | break; | ||
821 | case qib_sdma_event_e70_go_idle: | ||
822 | break; | ||
823 | case qib_sdma_event_e7220_err_halted: | ||
824 | break; | ||
825 | case qib_sdma_event_e7322_err_halted: | ||
826 | break; | ||
827 | case qib_sdma_event_e90_timer_tick: | ||
828 | break; | ||
829 | } | ||
830 | break; | ||
831 | |||
832 | case qib_sdma_state_s30_sw_clean_up_wait: | ||
833 | switch (event) { | ||
834 | case qib_sdma_event_e00_go_hw_down: | ||
835 | sdma_set_state(ppd, qib_sdma_state_s00_hw_down); | ||
836 | break; | ||
837 | case qib_sdma_event_e10_go_hw_start: | ||
838 | break; | ||
839 | case qib_sdma_event_e20_hw_started: | ||
840 | break; | ||
841 | case qib_sdma_event_e30_go_running: | ||
842 | ss->go_s99_running = 1; | ||
843 | break; | ||
844 | case qib_sdma_event_e40_sw_cleaned: | ||
845 | sdma_set_state(ppd, | ||
846 | qib_sdma_state_s10_hw_start_up_wait); | ||
847 | sdma_hw_start_up(ppd); | ||
848 | break; | ||
849 | case qib_sdma_event_e50_hw_cleaned: | ||
850 | break; | ||
851 | case qib_sdma_event_e60_hw_halted: | ||
852 | break; | ||
853 | case qib_sdma_event_e70_go_idle: | ||
854 | ss->go_s99_running = 0; | ||
855 | break; | ||
856 | case qib_sdma_event_e7220_err_halted: | ||
857 | break; | ||
858 | case qib_sdma_event_e7322_err_halted: | ||
859 | break; | ||
860 | case qib_sdma_event_e90_timer_tick: | ||
861 | break; | ||
862 | } | ||
863 | break; | ||
864 | |||
865 | case qib_sdma_state_s40_hw_clean_up_wait: | ||
866 | switch (event) { | ||
867 | case qib_sdma_event_e00_go_hw_down: | ||
868 | sdma_set_state(ppd, qib_sdma_state_s00_hw_down); | ||
869 | sdma_start_sw_clean_up(ppd); | ||
870 | break; | ||
871 | case qib_sdma_event_e10_go_hw_start: | ||
872 | break; | ||
873 | case qib_sdma_event_e20_hw_started: | ||
874 | break; | ||
875 | case qib_sdma_event_e30_go_running: | ||
876 | ss->go_s99_running = 1; | ||
877 | break; | ||
878 | case qib_sdma_event_e40_sw_cleaned: | ||
879 | break; | ||
880 | case qib_sdma_event_e50_hw_cleaned: | ||
881 | sdma_set_state(ppd, | ||
882 | qib_sdma_state_s30_sw_clean_up_wait); | ||
883 | sdma_start_sw_clean_up(ppd); | ||
884 | break; | ||
885 | case qib_sdma_event_e60_hw_halted: | ||
886 | break; | ||
887 | case qib_sdma_event_e70_go_idle: | ||
888 | ss->go_s99_running = 0; | ||
889 | break; | ||
890 | case qib_sdma_event_e7220_err_halted: | ||
891 | break; | ||
892 | case qib_sdma_event_e7322_err_halted: | ||
893 | break; | ||
894 | case qib_sdma_event_e90_timer_tick: | ||
895 | break; | ||
896 | } | ||
897 | break; | ||
898 | |||
899 | case qib_sdma_state_s50_hw_halt_wait: | ||
900 | switch (event) { | ||
901 | case qib_sdma_event_e00_go_hw_down: | ||
902 | sdma_set_state(ppd, qib_sdma_state_s00_hw_down); | ||
903 | sdma_start_sw_clean_up(ppd); | ||
904 | break; | ||
905 | case qib_sdma_event_e10_go_hw_start: | ||
906 | break; | ||
907 | case qib_sdma_event_e20_hw_started: | ||
908 | break; | ||
909 | case qib_sdma_event_e30_go_running: | ||
910 | ss->go_s99_running = 1; | ||
911 | break; | ||
912 | case qib_sdma_event_e40_sw_cleaned: | ||
913 | break; | ||
914 | case qib_sdma_event_e50_hw_cleaned: | ||
915 | break; | ||
916 | case qib_sdma_event_e60_hw_halted: | ||
917 | sdma_set_state(ppd, | ||
918 | qib_sdma_state_s40_hw_clean_up_wait); | ||
919 | ppd->dd->f_sdma_hw_clean_up(ppd); | ||
920 | break; | ||
921 | case qib_sdma_event_e70_go_idle: | ||
922 | ss->go_s99_running = 0; | ||
923 | break; | ||
924 | case qib_sdma_event_e7220_err_halted: | ||
925 | break; | ||
926 | case qib_sdma_event_e7322_err_halted: | ||
927 | break; | ||
928 | case qib_sdma_event_e90_timer_tick: | ||
929 | break; | ||
930 | } | ||
931 | break; | ||
932 | |||
933 | case qib_sdma_state_s99_running: | ||
934 | switch (event) { | ||
935 | case qib_sdma_event_e00_go_hw_down: | ||
936 | sdma_set_state(ppd, qib_sdma_state_s00_hw_down); | ||
937 | sdma_start_sw_clean_up(ppd); | ||
938 | break; | ||
939 | case qib_sdma_event_e10_go_hw_start: | ||
940 | break; | ||
941 | case qib_sdma_event_e20_hw_started: | ||
942 | break; | ||
943 | case qib_sdma_event_e30_go_running: | ||
944 | break; | ||
945 | case qib_sdma_event_e40_sw_cleaned: | ||
946 | break; | ||
947 | case qib_sdma_event_e50_hw_cleaned: | ||
948 | break; | ||
949 | case qib_sdma_event_e60_hw_halted: | ||
950 | sdma_set_state(ppd, | ||
951 | qib_sdma_state_s30_sw_clean_up_wait); | ||
952 | sdma_start_sw_clean_up(ppd); | ||
953 | break; | ||
954 | case qib_sdma_event_e70_go_idle: | ||
955 | sdma_set_state(ppd, qib_sdma_state_s50_hw_halt_wait); | ||
956 | ss->go_s99_running = 0; | ||
957 | break; | ||
958 | case qib_sdma_event_e7220_err_halted: | ||
959 | sdma_set_state(ppd, | ||
960 | qib_sdma_state_s30_sw_clean_up_wait); | ||
961 | sdma_start_sw_clean_up(ppd); | ||
962 | break; | ||
963 | case qib_sdma_event_e7322_err_halted: | ||
964 | sdma_set_state(ppd, qib_sdma_state_s50_hw_halt_wait); | ||
965 | break; | ||
966 | case qib_sdma_event_e90_timer_tick: | ||
967 | break; | ||
968 | } | ||
969 | break; | ||
970 | } | ||
971 | |||
972 | ss->last_event = event; | ||
973 | } | ||