]>
Commit | Line | Data |
---|---|---|
77241056 MM |
1 | /* |
2 | * | |
3 | * This file is provided under a dual BSD/GPLv2 license. When using or | |
4 | * redistributing this file, you may do so under either license. | |
5 | * | |
6 | * GPL LICENSE SUMMARY | |
7 | * | |
8 | * Copyright(c) 2015 Intel Corporation. | |
9 | * | |
10 | * This program is free software; you can redistribute it and/or modify | |
11 | * it under the terms of version 2 of the GNU General Public License as | |
12 | * published by the Free Software Foundation. | |
13 | * | |
14 | * This program is distributed in the hope that it will be useful, but | |
15 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
17 | * General Public License for more details. | |
18 | * | |
19 | * BSD LICENSE | |
20 | * | |
21 | * Copyright(c) 2015 Intel Corporation. | |
22 | * | |
23 | * Redistribution and use in source and binary forms, with or without | |
24 | * modification, are permitted provided that the following conditions | |
25 | * are met: | |
26 | * | |
27 | * - Redistributions of source code must retain the above copyright | |
28 | * notice, this list of conditions and the following disclaimer. | |
29 | * - Redistributions in binary form must reproduce the above copyright | |
30 | * notice, this list of conditions and the following disclaimer in | |
31 | * the documentation and/or other materials provided with the | |
32 | * distribution. | |
33 | * - Neither the name of Intel Corporation nor the names of its | |
34 | * contributors may be used to endorse or promote products derived | |
35 | * from this software without specific prior written permission. | |
36 | * | |
37 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
38 | * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
39 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR | |
40 | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT | |
41 | * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, | |
42 | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT | |
43 | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, | |
44 | * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY | |
45 | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT | |
46 | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | |
47 | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |
48 | * | |
49 | */ | |
50 | ||
51 | #include <linux/spinlock.h> | |
52 | #include <linux/seqlock.h> | |
53 | #include <linux/netdevice.h> | |
54 | #include <linux/moduleparam.h> | |
55 | #include <linux/bitops.h> | |
56 | #include <linux/timer.h> | |
57 | #include <linux/vmalloc.h> | |
58 | ||
59 | #include "hfi.h" | |
60 | #include "common.h" | |
61 | #include "qp.h" | |
62 | #include "sdma.h" | |
63 | #include "iowait.h" | |
64 | #include "trace.h" | |
65 | ||
66 | /* must be a power of 2 >= 64 <= 32768 */ | |
67 | #define SDMA_DESCQ_CNT 1024 | |
68 | #define INVALID_TAIL 0xffff | |
69 | ||
70 | static uint sdma_descq_cnt = SDMA_DESCQ_CNT; | |
71 | module_param(sdma_descq_cnt, uint, S_IRUGO); | |
72 | MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries"); | |
73 | ||
74 | static uint sdma_idle_cnt = 250; | |
75 | module_param(sdma_idle_cnt, uint, S_IRUGO); | |
76 | MODULE_PARM_DESC(sdma_idle_cnt, "sdma interrupt idle delay (ns,default 250)"); | |
77 | ||
78 | uint mod_num_sdma; | |
79 | module_param_named(num_sdma, mod_num_sdma, uint, S_IRUGO); | |
80 | MODULE_PARM_DESC(num_sdma, "Set max number SDMA engines to use"); | |
81 | ||
82 | #define SDMA_WAIT_BATCH_SIZE 20 | |
83 | /* max wait time for a SDMA engine to indicate it has halted */ | |
84 | #define SDMA_ERR_HALT_TIMEOUT 10 /* ms */ | |
85 | /* all SDMA engine errors that cause a halt */ | |
86 | ||
87 | #define SD(name) SEND_DMA_##name | |
88 | #define ALL_SDMA_ENG_HALT_ERRS \ | |
89 | (SD(ENG_ERR_STATUS_SDMA_WRONG_DW_ERR_SMASK) \ | |
90 | | SD(ENG_ERR_STATUS_SDMA_GEN_MISMATCH_ERR_SMASK) \ | |
91 | | SD(ENG_ERR_STATUS_SDMA_TOO_LONG_ERR_SMASK) \ | |
92 | | SD(ENG_ERR_STATUS_SDMA_TAIL_OUT_OF_BOUNDS_ERR_SMASK) \ | |
93 | | SD(ENG_ERR_STATUS_SDMA_FIRST_DESC_ERR_SMASK) \ | |
94 | | SD(ENG_ERR_STATUS_SDMA_MEM_READ_ERR_SMASK) \ | |
95 | | SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK) \ | |
96 | | SD(ENG_ERR_STATUS_SDMA_LENGTH_MISMATCH_ERR_SMASK) \ | |
97 | | SD(ENG_ERR_STATUS_SDMA_PACKET_DESC_OVERFLOW_ERR_SMASK) \ | |
98 | | SD(ENG_ERR_STATUS_SDMA_HEADER_SELECT_ERR_SMASK) \ | |
99 | | SD(ENG_ERR_STATUS_SDMA_HEADER_ADDRESS_ERR_SMASK) \ | |
100 | | SD(ENG_ERR_STATUS_SDMA_HEADER_LENGTH_ERR_SMASK) \ | |
101 | | SD(ENG_ERR_STATUS_SDMA_TIMEOUT_ERR_SMASK) \ | |
102 | | SD(ENG_ERR_STATUS_SDMA_DESC_TABLE_UNC_ERR_SMASK) \ | |
103 | | SD(ENG_ERR_STATUS_SDMA_ASSEMBLY_UNC_ERR_SMASK) \ | |
104 | | SD(ENG_ERR_STATUS_SDMA_PACKET_TRACKING_UNC_ERR_SMASK) \ | |
105 | | SD(ENG_ERR_STATUS_SDMA_HEADER_STORAGE_UNC_ERR_SMASK) \ | |
106 | | SD(ENG_ERR_STATUS_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SMASK)) | |
107 | ||
108 | /* sdma_sendctrl operations */ | |
109 | #define SDMA_SENDCTRL_OP_ENABLE (1U << 0) | |
110 | #define SDMA_SENDCTRL_OP_INTENABLE (1U << 1) | |
111 | #define SDMA_SENDCTRL_OP_HALT (1U << 2) | |
112 | #define SDMA_SENDCTRL_OP_CLEANUP (1U << 3) | |
113 | ||
114 | /* handle long defines */ | |
115 | #define SDMA_EGRESS_PACKET_OCCUPANCY_SMASK \ | |
116 | SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SMASK | |
117 | #define SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT \ | |
118 | SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT | |
119 | ||
120 | static const char * const sdma_state_names[] = { | |
121 | [sdma_state_s00_hw_down] = "s00_HwDown", | |
122 | [sdma_state_s10_hw_start_up_halt_wait] = "s10_HwStartUpHaltWait", | |
123 | [sdma_state_s15_hw_start_up_clean_wait] = "s15_HwStartUpCleanWait", | |
124 | [sdma_state_s20_idle] = "s20_Idle", | |
125 | [sdma_state_s30_sw_clean_up_wait] = "s30_SwCleanUpWait", | |
126 | [sdma_state_s40_hw_clean_up_wait] = "s40_HwCleanUpWait", | |
127 | [sdma_state_s50_hw_halt_wait] = "s50_HwHaltWait", | |
128 | [sdma_state_s60_idle_halt_wait] = "s60_IdleHaltWait", | |
129 | [sdma_state_s80_hw_freeze] = "s80_HwFreeze", | |
130 | [sdma_state_s82_freeze_sw_clean] = "s82_FreezeSwClean", | |
131 | [sdma_state_s99_running] = "s99_Running", | |
132 | }; | |
133 | ||
134 | static const char * const sdma_event_names[] = { | |
135 | [sdma_event_e00_go_hw_down] = "e00_GoHwDown", | |
136 | [sdma_event_e10_go_hw_start] = "e10_GoHwStart", | |
137 | [sdma_event_e15_hw_halt_done] = "e15_HwHaltDone", | |
138 | [sdma_event_e25_hw_clean_up_done] = "e25_HwCleanUpDone", | |
139 | [sdma_event_e30_go_running] = "e30_GoRunning", | |
140 | [sdma_event_e40_sw_cleaned] = "e40_SwCleaned", | |
141 | [sdma_event_e50_hw_cleaned] = "e50_HwCleaned", | |
142 | [sdma_event_e60_hw_halted] = "e60_HwHalted", | |
143 | [sdma_event_e70_go_idle] = "e70_GoIdle", | |
144 | [sdma_event_e80_hw_freeze] = "e80_HwFreeze", | |
145 | [sdma_event_e81_hw_frozen] = "e81_HwFrozen", | |
146 | [sdma_event_e82_hw_unfreeze] = "e82_HwUnfreeze", | |
147 | [sdma_event_e85_link_down] = "e85_LinkDown", | |
148 | [sdma_event_e90_sw_halted] = "e90_SwHalted", | |
149 | }; | |
150 | ||
151 | static const struct sdma_set_state_action sdma_action_table[] = { | |
152 | [sdma_state_s00_hw_down] = { | |
153 | .go_s99_running_tofalse = 1, | |
154 | .op_enable = 0, | |
155 | .op_intenable = 0, | |
156 | .op_halt = 0, | |
157 | .op_cleanup = 0, | |
158 | }, | |
159 | [sdma_state_s10_hw_start_up_halt_wait] = { | |
160 | .op_enable = 0, | |
161 | .op_intenable = 0, | |
162 | .op_halt = 1, | |
163 | .op_cleanup = 0, | |
164 | }, | |
165 | [sdma_state_s15_hw_start_up_clean_wait] = { | |
166 | .op_enable = 0, | |
167 | .op_intenable = 1, | |
168 | .op_halt = 0, | |
169 | .op_cleanup = 1, | |
170 | }, | |
171 | [sdma_state_s20_idle] = { | |
172 | .op_enable = 0, | |
173 | .op_intenable = 1, | |
174 | .op_halt = 0, | |
175 | .op_cleanup = 0, | |
176 | }, | |
177 | [sdma_state_s30_sw_clean_up_wait] = { | |
178 | .op_enable = 0, | |
179 | .op_intenable = 0, | |
180 | .op_halt = 0, | |
181 | .op_cleanup = 0, | |
182 | }, | |
183 | [sdma_state_s40_hw_clean_up_wait] = { | |
184 | .op_enable = 0, | |
185 | .op_intenable = 0, | |
186 | .op_halt = 0, | |
187 | .op_cleanup = 1, | |
188 | }, | |
189 | [sdma_state_s50_hw_halt_wait] = { | |
190 | .op_enable = 0, | |
191 | .op_intenable = 0, | |
192 | .op_halt = 0, | |
193 | .op_cleanup = 0, | |
194 | }, | |
195 | [sdma_state_s60_idle_halt_wait] = { | |
196 | .go_s99_running_tofalse = 1, | |
197 | .op_enable = 0, | |
198 | .op_intenable = 0, | |
199 | .op_halt = 1, | |
200 | .op_cleanup = 0, | |
201 | }, | |
202 | [sdma_state_s80_hw_freeze] = { | |
203 | .op_enable = 0, | |
204 | .op_intenable = 0, | |
205 | .op_halt = 0, | |
206 | .op_cleanup = 0, | |
207 | }, | |
208 | [sdma_state_s82_freeze_sw_clean] = { | |
209 | .op_enable = 0, | |
210 | .op_intenable = 0, | |
211 | .op_halt = 0, | |
212 | .op_cleanup = 0, | |
213 | }, | |
214 | [sdma_state_s99_running] = { | |
215 | .op_enable = 1, | |
216 | .op_intenable = 1, | |
217 | .op_halt = 0, | |
218 | .op_cleanup = 0, | |
219 | .go_s99_running_totrue = 1, | |
220 | }, | |
221 | }; | |
222 | ||
223 | #define SDMA_TAIL_UPDATE_THRESH 0x1F | |
224 | ||
225 | /* declare all statics here rather than keep sorting */ | |
226 | static void sdma_complete(struct kref *); | |
227 | static void sdma_finalput(struct sdma_state *); | |
228 | static void sdma_get(struct sdma_state *); | |
229 | static void sdma_hw_clean_up_task(unsigned long); | |
230 | static void sdma_put(struct sdma_state *); | |
231 | static void sdma_set_state(struct sdma_engine *, enum sdma_states); | |
232 | static void sdma_start_hw_clean_up(struct sdma_engine *); | |
233 | static void sdma_start_sw_clean_up(struct sdma_engine *); | |
234 | static void sdma_sw_clean_up_task(unsigned long); | |
235 | static void sdma_sendctrl(struct sdma_engine *, unsigned); | |
236 | static void init_sdma_regs(struct sdma_engine *, u32, uint); | |
237 | static void sdma_process_event( | |
238 | struct sdma_engine *sde, | |
239 | enum sdma_events event); | |
240 | static void __sdma_process_event( | |
241 | struct sdma_engine *sde, | |
242 | enum sdma_events event); | |
243 | static void dump_sdma_state(struct sdma_engine *sde); | |
244 | static void sdma_make_progress(struct sdma_engine *sde, u64 status); | |
245 | static void sdma_desc_avail(struct sdma_engine *sde, unsigned avail); | |
246 | static void sdma_flush_descq(struct sdma_engine *sde); | |
247 | ||
248 | /** | |
249 | * sdma_state_name() - return state string from enum | |
250 | * @state: state | |
251 | */ | |
252 | static const char *sdma_state_name(enum sdma_states state) | |
253 | { | |
254 | return sdma_state_names[state]; | |
255 | } | |
256 | ||
257 | static void sdma_get(struct sdma_state *ss) | |
258 | { | |
259 | kref_get(&ss->kref); | |
260 | } | |
261 | ||
262 | static void sdma_complete(struct kref *kref) | |
263 | { | |
264 | struct sdma_state *ss = | |
265 | container_of(kref, struct sdma_state, kref); | |
266 | ||
267 | complete(&ss->comp); | |
268 | } | |
269 | ||
270 | static void sdma_put(struct sdma_state *ss) | |
271 | { | |
272 | kref_put(&ss->kref, sdma_complete); | |
273 | } | |
274 | ||
275 | static void sdma_finalput(struct sdma_state *ss) | |
276 | { | |
277 | sdma_put(ss); | |
278 | wait_for_completion(&ss->comp); | |
279 | } | |
280 | ||
281 | static inline void write_sde_csr( | |
282 | struct sdma_engine *sde, | |
283 | u32 offset0, | |
284 | u64 value) | |
285 | { | |
286 | write_kctxt_csr(sde->dd, sde->this_idx, offset0, value); | |
287 | } | |
288 | ||
289 | static inline u64 read_sde_csr( | |
290 | struct sdma_engine *sde, | |
291 | u32 offset0) | |
292 | { | |
293 | return read_kctxt_csr(sde->dd, sde->this_idx, offset0); | |
294 | } | |
295 | ||
296 | /* | |
297 | * sdma_wait_for_packet_egress() - wait for the VL FIFO occupancy for | |
298 | * sdma engine 'sde' to drop to 0. | |
299 | */ | |
300 | static void sdma_wait_for_packet_egress(struct sdma_engine *sde, | |
301 | int pause) | |
302 | { | |
303 | u64 off = 8 * sde->this_idx; | |
304 | struct hfi1_devdata *dd = sde->dd; | |
305 | int lcnt = 0; | |
306 | ||
307 | while (1) { | |
308 | u64 reg = read_csr(dd, off + SEND_EGRESS_SEND_DMA_STATUS); | |
309 | ||
310 | reg &= SDMA_EGRESS_PACKET_OCCUPANCY_SMASK; | |
311 | reg >>= SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT; | |
312 | if (reg == 0) | |
313 | break; | |
314 | if (lcnt++ > 100) { | |
315 | dd_dev_err(dd, "%s: engine %u timeout waiting for packets to egress, remaining count %u\n", | |
316 | __func__, sde->this_idx, (u32)reg); | |
317 | break; | |
318 | } | |
319 | udelay(1); | |
320 | } | |
321 | } | |
322 | ||
323 | /* | |
324 | * sdma_wait() - wait for packet egress to complete for all SDMA engines, | |
325 | * and pause for credit return. | |
326 | */ | |
327 | void sdma_wait(struct hfi1_devdata *dd) | |
328 | { | |
329 | int i; | |
330 | ||
331 | for (i = 0; i < dd->num_sdma; i++) { | |
332 | struct sdma_engine *sde = &dd->per_sdma[i]; | |
333 | ||
334 | sdma_wait_for_packet_egress(sde, 0); | |
335 | } | |
336 | } | |
337 | ||
338 | static inline void sdma_set_desc_cnt(struct sdma_engine *sde, unsigned cnt) | |
339 | { | |
340 | u64 reg; | |
341 | ||
342 | if (!(sde->dd->flags & HFI1_HAS_SDMA_TIMEOUT)) | |
343 | return; | |
344 | reg = cnt; | |
345 | reg &= SD(DESC_CNT_CNT_MASK); | |
346 | reg <<= SD(DESC_CNT_CNT_SHIFT); | |
347 | write_sde_csr(sde, SD(DESC_CNT), reg); | |
348 | } | |
349 | ||
350 | /* | |
351 | * Complete all the sdma requests with a SDMA_TXREQ_S_ABORTED status | |
352 | * | |
353 | * Depending on timing there can be txreqs in two places: | |
354 | * - in the descq ring | |
355 | * - in the flush list | |
356 | * | |
357 | * To avoid ordering issues the descq ring needs to be flushed | |
358 | * first followed by the flush list. | |
359 | * | |
360 | * This routine is called from two places | |
361 | * - From a work queue item | |
362 | * - Directly from the state machine just before setting the | |
363 | * state to running | |
364 | * | |
365 | * Must be called with head_lock held | |
366 | * | |
367 | */ | |
368 | static void sdma_flush(struct sdma_engine *sde) | |
369 | { | |
370 | struct sdma_txreq *txp, *txp_next; | |
371 | LIST_HEAD(flushlist); | |
372 | ||
373 | /* flush from head to tail */ | |
374 | sdma_flush_descq(sde); | |
375 | spin_lock(&sde->flushlist_lock); | |
376 | /* copy flush list */ | |
377 | list_for_each_entry_safe(txp, txp_next, &sde->flushlist, list) { | |
378 | list_del_init(&txp->list); | |
379 | list_add_tail(&txp->list, &flushlist); | |
380 | } | |
381 | spin_unlock(&sde->flushlist_lock); | |
382 | /* flush from flush list */ | |
383 | list_for_each_entry_safe(txp, txp_next, &flushlist, list) { | |
384 | int drained = 0; | |
385 | /* protect against complete modifying */ | |
386 | struct iowait *wait = txp->wait; | |
387 | ||
388 | list_del_init(&txp->list); | |
389 | #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER | |
390 | trace_hfi1_sdma_out_sn(sde, txp->sn); | |
391 | if (WARN_ON_ONCE(sde->head_sn != txp->sn)) | |
392 | dd_dev_err(sde->dd, "expected %llu got %llu\n", | |
393 | sde->head_sn, txp->sn); | |
394 | sde->head_sn++; | |
395 | #endif | |
396 | sdma_txclean(sde->dd, txp); | |
397 | if (wait) | |
398 | drained = atomic_dec_and_test(&wait->sdma_busy); | |
399 | if (txp->complete) | |
400 | (*txp->complete)(txp, SDMA_TXREQ_S_ABORTED, drained); | |
401 | if (wait && drained) | |
402 | iowait_drain_wakeup(wait); | |
403 | } | |
404 | } | |
405 | ||
406 | /* | |
407 | * Fields a work request for flushing the descq ring | |
408 | * and the flush list | |
409 | * | |
410 | * If the engine has been brought to running during | |
411 | * the scheduling delay, the flush is ignored, assuming | |
412 | * that the process of bringing the engine to running | |
413 | * would have done this flush prior to going to running. | |
414 | * | |
415 | */ | |
416 | static void sdma_field_flush(struct work_struct *work) | |
417 | { | |
418 | unsigned long flags; | |
419 | struct sdma_engine *sde = | |
420 | container_of(work, struct sdma_engine, flush_worker); | |
421 | ||
422 | write_seqlock_irqsave(&sde->head_lock, flags); | |
423 | if (!__sdma_running(sde)) | |
424 | sdma_flush(sde); | |
425 | write_sequnlock_irqrestore(&sde->head_lock, flags); | |
426 | } | |
427 | ||
428 | static void sdma_err_halt_wait(struct work_struct *work) | |
429 | { | |
430 | struct sdma_engine *sde = container_of(work, struct sdma_engine, | |
431 | err_halt_worker); | |
432 | u64 statuscsr; | |
433 | unsigned long timeout; | |
434 | ||
435 | timeout = jiffies + msecs_to_jiffies(SDMA_ERR_HALT_TIMEOUT); | |
436 | while (1) { | |
437 | statuscsr = read_sde_csr(sde, SD(STATUS)); | |
438 | statuscsr &= SD(STATUS_ENG_HALTED_SMASK); | |
439 | if (statuscsr) | |
440 | break; | |
441 | if (time_after(jiffies, timeout)) { | |
442 | dd_dev_err(sde->dd, | |
443 | "SDMA engine %d - timeout waiting for engine to halt\n", | |
444 | sde->this_idx); | |
445 | /* | |
446 | * Continue anyway. This could happen if there was | |
447 | * an uncorrectable error in the wrong spot. | |
448 | */ | |
449 | break; | |
450 | } | |
451 | usleep_range(80, 120); | |
452 | } | |
453 | ||
454 | sdma_process_event(sde, sdma_event_e15_hw_halt_done); | |
455 | } | |
456 | ||
457 | static void sdma_start_err_halt_wait(struct sdma_engine *sde) | |
458 | { | |
459 | schedule_work(&sde->err_halt_worker); | |
460 | } | |
461 | ||
462 | ||
463 | static void sdma_err_progress_check_schedule(struct sdma_engine *sde) | |
464 | { | |
465 | if (!is_bx(sde->dd) && HFI1_CAP_IS_KSET(SDMA_AHG)) { | |
466 | ||
467 | unsigned index; | |
468 | struct hfi1_devdata *dd = sde->dd; | |
469 | ||
470 | for (index = 0; index < dd->num_sdma; index++) { | |
471 | struct sdma_engine *curr_sdma = &dd->per_sdma[index]; | |
472 | ||
473 | if (curr_sdma != sde) | |
474 | curr_sdma->progress_check_head = | |
475 | curr_sdma->descq_head; | |
476 | } | |
477 | dd_dev_err(sde->dd, | |
478 | "SDMA engine %d - check scheduled\n", | |
479 | sde->this_idx); | |
480 | mod_timer(&sde->err_progress_check_timer, jiffies + 10); | |
481 | } | |
482 | } | |
483 | ||
484 | static void sdma_err_progress_check(unsigned long data) | |
485 | { | |
486 | unsigned index; | |
487 | struct sdma_engine *sde = (struct sdma_engine *)data; | |
488 | ||
489 | dd_dev_err(sde->dd, "SDE progress check event\n"); | |
490 | for (index = 0; index < sde->dd->num_sdma; index++) { | |
491 | struct sdma_engine *curr_sde = &sde->dd->per_sdma[index]; | |
492 | unsigned long flags; | |
493 | ||
494 | /* check progress on each engine except the current one */ | |
495 | if (curr_sde == sde) | |
496 | continue; | |
497 | /* | |
498 | * We must lock interrupts when acquiring sde->lock, | |
499 | * to avoid a deadlock if interrupt triggers and spins on | |
500 | * the same lock on same CPU | |
501 | */ | |
502 | spin_lock_irqsave(&curr_sde->tail_lock, flags); | |
503 | write_seqlock(&curr_sde->head_lock); | |
504 | ||
505 | /* skip non-running queues */ | |
506 | if (curr_sde->state.current_state != sdma_state_s99_running) { | |
507 | write_sequnlock(&curr_sde->head_lock); | |
508 | spin_unlock_irqrestore(&curr_sde->tail_lock, flags); | |
509 | continue; | |
510 | } | |
511 | ||
512 | if ((curr_sde->descq_head != curr_sde->descq_tail) && | |
513 | (curr_sde->descq_head == | |
514 | curr_sde->progress_check_head)) | |
515 | __sdma_process_event(curr_sde, | |
516 | sdma_event_e90_sw_halted); | |
517 | write_sequnlock(&curr_sde->head_lock); | |
518 | spin_unlock_irqrestore(&curr_sde->tail_lock, flags); | |
519 | } | |
520 | schedule_work(&sde->err_halt_worker); | |
521 | } | |
522 | ||
523 | static void sdma_hw_clean_up_task(unsigned long opaque) | |
524 | { | |
525 | struct sdma_engine *sde = (struct sdma_engine *) opaque; | |
526 | u64 statuscsr; | |
527 | ||
528 | while (1) { | |
529 | #ifdef CONFIG_SDMA_VERBOSITY | |
530 | dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", | |
531 | sde->this_idx, slashstrip(__FILE__), __LINE__, | |
532 | __func__); | |
533 | #endif | |
534 | statuscsr = read_sde_csr(sde, SD(STATUS)); | |
535 | statuscsr &= SD(STATUS_ENG_CLEANED_UP_SMASK); | |
536 | if (statuscsr) | |
537 | break; | |
538 | udelay(10); | |
539 | } | |
540 | ||
541 | sdma_process_event(sde, sdma_event_e25_hw_clean_up_done); | |
542 | } | |
543 | ||
544 | static inline struct sdma_txreq *get_txhead(struct sdma_engine *sde) | |
545 | { | |
546 | smp_read_barrier_depends(); /* see sdma_update_tail() */ | |
547 | return sde->tx_ring[sde->tx_head & sde->sdma_mask]; | |
548 | } | |
549 | ||
550 | /* | |
551 | * flush ring for recovery | |
552 | */ | |
553 | static void sdma_flush_descq(struct sdma_engine *sde) | |
554 | { | |
555 | u16 head, tail; | |
556 | int progress = 0; | |
557 | struct sdma_txreq *txp = get_txhead(sde); | |
558 | ||
559 | /* The reason for some of the complexity of this code is that | |
560 | * not all descriptors have corresponding txps. So, we have to | |
561 | * be able to skip over descs until we wander into the range of | |
562 | * the next txp on the list. | |
563 | */ | |
564 | head = sde->descq_head & sde->sdma_mask; | |
565 | tail = sde->descq_tail & sde->sdma_mask; | |
566 | while (head != tail) { | |
567 | /* advance head, wrap if needed */ | |
568 | head = ++sde->descq_head & sde->sdma_mask; | |
569 | /* if now past this txp's descs, do the callback */ | |
570 | if (txp && txp->next_descq_idx == head) { | |
571 | int drained = 0; | |
572 | /* protect against complete modifying */ | |
573 | struct iowait *wait = txp->wait; | |
574 | ||
575 | /* remove from list */ | |
576 | sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL; | |
577 | if (wait) | |
578 | drained = atomic_dec_and_test(&wait->sdma_busy); | |
579 | #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER | |
580 | trace_hfi1_sdma_out_sn(sde, txp->sn); | |
581 | if (WARN_ON_ONCE(sde->head_sn != txp->sn)) | |
582 | dd_dev_err(sde->dd, "expected %llu got %llu\n", | |
583 | sde->head_sn, txp->sn); | |
584 | sde->head_sn++; | |
585 | #endif | |
586 | sdma_txclean(sde->dd, txp); | |
587 | trace_hfi1_sdma_progress(sde, head, tail, txp); | |
588 | if (txp->complete) | |
589 | (*txp->complete)( | |
590 | txp, | |
591 | SDMA_TXREQ_S_ABORTED, | |
592 | drained); | |
593 | if (wait && drained) | |
594 | iowait_drain_wakeup(wait); | |
595 | /* see if there is another txp */ | |
596 | txp = get_txhead(sde); | |
597 | } | |
598 | progress++; | |
599 | } | |
600 | if (progress) | |
601 | sdma_desc_avail(sde, sdma_descq_freecnt(sde)); | |
602 | } | |
603 | ||
604 | static void sdma_sw_clean_up_task(unsigned long opaque) | |
605 | { | |
606 | struct sdma_engine *sde = (struct sdma_engine *) opaque; | |
607 | unsigned long flags; | |
608 | ||
609 | spin_lock_irqsave(&sde->tail_lock, flags); | |
610 | write_seqlock(&sde->head_lock); | |
611 | ||
612 | /* | |
613 | * At this point, the following should always be true: | |
614 | * - We are halted, so no more descriptors are getting retired. | |
615 | * - We are not running, so no one is submitting new work. | |
616 | * - Only we can send the e40_sw_cleaned, so we can't start | |
617 | * running again until we say so. So, the active list and | |
618 | * descq are ours to play with. | |
619 | */ | |
620 | ||
621 | ||
622 | /* | |
623 | * In the error clean up sequence, software clean must be called | |
624 | * before the hardware clean so we can use the hardware head in | |
625 | * the progress routine. A hardware clean or SPC unfreeze will | |
626 | * reset the hardware head. | |
627 | * | |
628 | * Process all retired requests. The progress routine will use the | |
629 | * latest physical hardware head - we are not running so speed does | |
630 | * not matter. | |
631 | */ | |
632 | sdma_make_progress(sde, 0); | |
633 | ||
634 | sdma_flush(sde); | |
635 | ||
636 | /* | |
637 | * Reset our notion of head and tail. | |
638 | * Note that the HW registers have been reset via an earlier | |
639 | * clean up. | |
640 | */ | |
641 | sde->descq_tail = 0; | |
642 | sde->descq_head = 0; | |
643 | sde->desc_avail = sdma_descq_freecnt(sde); | |
644 | *sde->head_dma = 0; | |
645 | ||
646 | __sdma_process_event(sde, sdma_event_e40_sw_cleaned); | |
647 | ||
648 | write_sequnlock(&sde->head_lock); | |
649 | spin_unlock_irqrestore(&sde->tail_lock, flags); | |
650 | } | |
651 | ||
652 | static void sdma_sw_tear_down(struct sdma_engine *sde) | |
653 | { | |
654 | struct sdma_state *ss = &sde->state; | |
655 | ||
656 | /* Releasing this reference means the state machine has stopped. */ | |
657 | sdma_put(ss); | |
658 | ||
659 | /* stop waiting for all unfreeze events to complete */ | |
660 | atomic_set(&sde->dd->sdma_unfreeze_count, -1); | |
661 | wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); | |
662 | } | |
663 | ||
664 | static void sdma_start_hw_clean_up(struct sdma_engine *sde) | |
665 | { | |
666 | tasklet_hi_schedule(&sde->sdma_hw_clean_up_task); | |
667 | } | |
668 | ||
669 | static void sdma_start_sw_clean_up(struct sdma_engine *sde) | |
670 | { | |
671 | tasklet_hi_schedule(&sde->sdma_sw_clean_up_task); | |
672 | } | |
673 | ||
674 | static void sdma_set_state(struct sdma_engine *sde, | |
675 | enum sdma_states next_state) | |
676 | { | |
677 | struct sdma_state *ss = &sde->state; | |
678 | const struct sdma_set_state_action *action = sdma_action_table; | |
679 | unsigned op = 0; | |
680 | ||
681 | trace_hfi1_sdma_state( | |
682 | sde, | |
683 | sdma_state_names[ss->current_state], | |
684 | sdma_state_names[next_state]); | |
685 | ||
686 | /* debugging bookkeeping */ | |
687 | ss->previous_state = ss->current_state; | |
688 | ss->previous_op = ss->current_op; | |
689 | ss->current_state = next_state; | |
690 | ||
691 | if (ss->previous_state != sdma_state_s99_running | |
692 | && next_state == sdma_state_s99_running) | |
693 | sdma_flush(sde); | |
694 | ||
695 | if (action[next_state].op_enable) | |
696 | op |= SDMA_SENDCTRL_OP_ENABLE; | |
697 | ||
698 | if (action[next_state].op_intenable) | |
699 | op |= SDMA_SENDCTRL_OP_INTENABLE; | |
700 | ||
701 | if (action[next_state].op_halt) | |
702 | op |= SDMA_SENDCTRL_OP_HALT; | |
703 | ||
704 | if (action[next_state].op_cleanup) | |
705 | op |= SDMA_SENDCTRL_OP_CLEANUP; | |
706 | ||
707 | if (action[next_state].go_s99_running_tofalse) | |
708 | ss->go_s99_running = 0; | |
709 | ||
710 | if (action[next_state].go_s99_running_totrue) | |
711 | ss->go_s99_running = 1; | |
712 | ||
713 | ss->current_op = op; | |
714 | sdma_sendctrl(sde, ss->current_op); | |
715 | } | |
716 | ||
717 | /** | |
718 | * sdma_get_descq_cnt() - called when device probed | |
719 | * | |
720 | * Return a validated descq count. | |
721 | * | |
722 | * This is currently only used in the verbs initialization to build the tx | |
723 | * list. | |
724 | * | |
725 | * This will probably be deleted in favor of a more scalable approach to | |
726 | * alloc tx's. | |
727 | * | |
728 | */ | |
729 | u16 sdma_get_descq_cnt(void) | |
730 | { | |
731 | u16 count = sdma_descq_cnt; | |
732 | ||
733 | if (!count) | |
734 | return SDMA_DESCQ_CNT; | |
735 | /* count must be a power of 2 greater than 64 and less than | |
736 | * 32768. Otherwise return default. | |
737 | */ | |
738 | if (!is_power_of_2(count)) | |
739 | return SDMA_DESCQ_CNT; | |
aeef010a | 740 | if (count < 64 || count > 32768) |
77241056 MM |
741 | return SDMA_DESCQ_CNT; |
742 | return count; | |
743 | } | |
744 | /** | |
745 | * sdma_select_engine_vl() - select sdma engine | |
746 | * @dd: devdata | |
747 | * @selector: a spreading factor | |
748 | * @vl: this vl | |
749 | * | |
750 | * | |
751 | * This function returns an engine based on the selector and a vl. The | |
752 | * mapping fields are protected by RCU. | |
753 | */ | |
754 | struct sdma_engine *sdma_select_engine_vl( | |
755 | struct hfi1_devdata *dd, | |
756 | u32 selector, | |
757 | u8 vl) | |
758 | { | |
759 | struct sdma_vl_map *m; | |
760 | struct sdma_map_elem *e; | |
761 | struct sdma_engine *rval; | |
762 | ||
763 | if (WARN_ON(vl > 8)) | |
764 | return NULL; | |
765 | ||
766 | rcu_read_lock(); | |
767 | m = rcu_dereference(dd->sdma_map); | |
768 | if (unlikely(!m)) { | |
769 | rcu_read_unlock(); | |
770 | return NULL; | |
771 | } | |
772 | e = m->map[vl & m->mask]; | |
773 | rval = e->sde[selector & e->mask]; | |
774 | rcu_read_unlock(); | |
775 | ||
776 | trace_hfi1_sdma_engine_select(dd, selector, vl, rval->this_idx); | |
777 | return rval; | |
778 | } | |
779 | ||
780 | /** | |
781 | * sdma_select_engine_sc() - select sdma engine | |
782 | * @dd: devdata | |
783 | * @selector: a spreading factor | |
784 | * @sc5: the 5 bit sc | |
785 | * | |
786 | * | |
787 | * This function returns an engine based on the selector and an sc. | |
788 | */ | |
789 | struct sdma_engine *sdma_select_engine_sc( | |
790 | struct hfi1_devdata *dd, | |
791 | u32 selector, | |
792 | u8 sc5) | |
793 | { | |
794 | u8 vl = sc_to_vlt(dd, sc5); | |
795 | ||
796 | return sdma_select_engine_vl(dd, selector, vl); | |
797 | } | |
798 | ||
799 | /* | |
800 | * Free the indicated map struct | |
801 | */ | |
802 | static void sdma_map_free(struct sdma_vl_map *m) | |
803 | { | |
804 | int i; | |
805 | ||
806 | for (i = 0; m && i < m->actual_vls; i++) | |
807 | kfree(m->map[i]); | |
808 | kfree(m); | |
809 | } | |
810 | ||
811 | /* | |
812 | * Handle RCU callback | |
813 | */ | |
814 | static void sdma_map_rcu_callback(struct rcu_head *list) | |
815 | { | |
816 | struct sdma_vl_map *m = container_of(list, struct sdma_vl_map, list); | |
817 | ||
818 | sdma_map_free(m); | |
819 | } | |
820 | ||
821 | /** | |
822 | * sdma_map_init - called when # vls change | |
823 | * @dd: hfi1_devdata | |
824 | * @port: port number | |
825 | * @num_vls: number of vls | |
826 | * @vl_engines: per vl engine mapping (optional) | |
827 | * | |
828 | * This routine changes the mapping based on the number of vls. | |
829 | * | |
830 | * vl_engines is used to specify a non-uniform vl/engine loading. NULL | |
831 | * implies auto computing the loading and giving each VLs a uniform | |
832 | * distribution of engines per VL. | |
833 | * | |
834 | * The auto algorithm computes the sde_per_vl and the number of extra | |
835 | * engines. Any extra engines are added from the last VL on down. | |
836 | * | |
837 | * rcu locking is used here to control access to the mapping fields. | |
838 | * | |
839 | * If either the num_vls or num_sdma are non-power of 2, the array sizes | |
840 | * in the struct sdma_vl_map and the struct sdma_map_elem are rounded | |
841 | * up to the next highest power of 2 and the first entry is reused | |
842 | * in a round robin fashion. | |
843 | * | |
844 | * If an error occurs the map change is not done and the mapping is | |
845 | * not changed. | |
846 | * | |
847 | */ | |
848 | int sdma_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_engines) | |
849 | { | |
850 | int i, j; | |
851 | int extra, sde_per_vl; | |
852 | int engine = 0; | |
853 | u8 lvl_engines[OPA_MAX_VLS]; | |
854 | struct sdma_vl_map *oldmap, *newmap; | |
855 | ||
856 | if (!(dd->flags & HFI1_HAS_SEND_DMA)) | |
857 | return 0; | |
858 | ||
859 | if (!vl_engines) { | |
860 | /* truncate divide */ | |
861 | sde_per_vl = dd->num_sdma / num_vls; | |
862 | /* extras */ | |
863 | extra = dd->num_sdma % num_vls; | |
864 | vl_engines = lvl_engines; | |
865 | /* add extras from last vl down */ | |
866 | for (i = num_vls - 1; i >= 0; i--, extra--) | |
867 | vl_engines[i] = sde_per_vl + (extra > 0 ? 1 : 0); | |
868 | } | |
869 | /* build new map */ | |
870 | newmap = kzalloc( | |
871 | sizeof(struct sdma_vl_map) + | |
872 | roundup_pow_of_two(num_vls) * | |
873 | sizeof(struct sdma_map_elem *), | |
874 | GFP_KERNEL); | |
875 | if (!newmap) | |
876 | goto bail; | |
877 | newmap->actual_vls = num_vls; | |
878 | newmap->vls = roundup_pow_of_two(num_vls); | |
879 | newmap->mask = (1 << ilog2(newmap->vls)) - 1; | |
880 | for (i = 0; i < newmap->vls; i++) { | |
881 | /* save for wrap around */ | |
882 | int first_engine = engine; | |
883 | ||
884 | if (i < newmap->actual_vls) { | |
885 | int sz = roundup_pow_of_two(vl_engines[i]); | |
886 | ||
887 | /* only allocate once */ | |
888 | newmap->map[i] = kzalloc( | |
889 | sizeof(struct sdma_map_elem) + | |
890 | sz * sizeof(struct sdma_engine *), | |
891 | GFP_KERNEL); | |
892 | if (!newmap->map[i]) | |
893 | goto bail; | |
894 | newmap->map[i]->mask = (1 << ilog2(sz)) - 1; | |
895 | /* assign engines */ | |
896 | for (j = 0; j < sz; j++) { | |
897 | newmap->map[i]->sde[j] = | |
898 | &dd->per_sdma[engine]; | |
899 | if (++engine >= first_engine + vl_engines[i]) | |
900 | /* wrap back to first engine */ | |
901 | engine = first_engine; | |
902 | } | |
903 | } else { | |
904 | /* just re-use entry without allocating */ | |
905 | newmap->map[i] = newmap->map[i % num_vls]; | |
906 | } | |
907 | engine = first_engine + vl_engines[i]; | |
908 | } | |
909 | /* newmap in hand, save old map */ | |
910 | spin_lock_irq(&dd->sde_map_lock); | |
911 | oldmap = rcu_dereference_protected(dd->sdma_map, | |
912 | lockdep_is_held(&dd->sde_map_lock)); | |
913 | ||
914 | /* publish newmap */ | |
915 | rcu_assign_pointer(dd->sdma_map, newmap); | |
916 | ||
917 | spin_unlock_irq(&dd->sde_map_lock); | |
918 | /* success, free any old map after grace period */ | |
919 | if (oldmap) | |
920 | call_rcu(&oldmap->list, sdma_map_rcu_callback); | |
921 | return 0; | |
922 | bail: | |
923 | /* free any partial allocation */ | |
924 | sdma_map_free(newmap); | |
925 | return -ENOMEM; | |
926 | } | |
927 | ||
928 | /* | |
929 | * Clean up allocated memory. | |
930 | * | |
931 | * This routine is can be called regardless of the success of sdma_init() | |
932 | * | |
933 | */ | |
934 | static void sdma_clean(struct hfi1_devdata *dd, size_t num_engines) | |
935 | { | |
936 | size_t i; | |
937 | struct sdma_engine *sde; | |
938 | ||
939 | if (dd->sdma_pad_dma) { | |
940 | dma_free_coherent(&dd->pcidev->dev, 4, | |
941 | (void *)dd->sdma_pad_dma, | |
942 | dd->sdma_pad_phys); | |
943 | dd->sdma_pad_dma = NULL; | |
944 | dd->sdma_pad_phys = 0; | |
945 | } | |
946 | if (dd->sdma_heads_dma) { | |
947 | dma_free_coherent(&dd->pcidev->dev, dd->sdma_heads_size, | |
948 | (void *)dd->sdma_heads_dma, | |
949 | dd->sdma_heads_phys); | |
950 | dd->sdma_heads_dma = NULL; | |
951 | dd->sdma_heads_phys = 0; | |
952 | } | |
953 | for (i = 0; dd->per_sdma && i < num_engines; ++i) { | |
954 | sde = &dd->per_sdma[i]; | |
955 | ||
956 | sde->head_dma = NULL; | |
957 | sde->head_phys = 0; | |
958 | ||
959 | if (sde->descq) { | |
960 | dma_free_coherent( | |
961 | &dd->pcidev->dev, | |
962 | sde->descq_cnt * sizeof(u64[2]), | |
963 | sde->descq, | |
964 | sde->descq_phys | |
965 | ); | |
966 | sde->descq = NULL; | |
967 | sde->descq_phys = 0; | |
968 | } | |
969 | if (is_vmalloc_addr(sde->tx_ring)) | |
970 | vfree(sde->tx_ring); | |
971 | else | |
972 | kfree(sde->tx_ring); | |
973 | sde->tx_ring = NULL; | |
974 | } | |
975 | spin_lock_irq(&dd->sde_map_lock); | |
976 | kfree(rcu_access_pointer(dd->sdma_map)); | |
977 | RCU_INIT_POINTER(dd->sdma_map, NULL); | |
978 | spin_unlock_irq(&dd->sde_map_lock); | |
979 | synchronize_rcu(); | |
980 | kfree(dd->per_sdma); | |
981 | dd->per_sdma = NULL; | |
982 | } | |
983 | ||
984 | /** | |
985 | * sdma_init() - called when device probed | |
986 | * @dd: hfi1_devdata | |
987 | * @port: port number (currently only zero) | |
988 | * | |
989 | * sdma_init initializes the specified number of engines. | |
990 | * | |
991 | * The code initializes each sde, its csrs. Interrupts | |
992 | * are not required to be enabled. | |
993 | * | |
994 | * Returns: | |
995 | * 0 - success, -errno on failure | |
996 | */ | |
997 | int sdma_init(struct hfi1_devdata *dd, u8 port) | |
998 | { | |
999 | unsigned this_idx; | |
1000 | struct sdma_engine *sde; | |
1001 | u16 descq_cnt; | |
1002 | void *curr_head; | |
1003 | struct hfi1_pportdata *ppd = dd->pport + port; | |
1004 | u32 per_sdma_credits; | |
1005 | uint idle_cnt = sdma_idle_cnt; | |
1006 | size_t num_engines = dd->chip_sdma_engines; | |
1007 | ||
1008 | if (!HFI1_CAP_IS_KSET(SDMA)) { | |
1009 | HFI1_CAP_CLEAR(SDMA_AHG); | |
1010 | return 0; | |
1011 | } | |
1012 | if (mod_num_sdma && | |
1013 | /* can't exceed chip support */ | |
1014 | mod_num_sdma <= dd->chip_sdma_engines && | |
1015 | /* count must be >= vls */ | |
1016 | mod_num_sdma >= num_vls) | |
1017 | num_engines = mod_num_sdma; | |
1018 | ||
1019 | dd_dev_info(dd, "SDMA mod_num_sdma: %u\n", mod_num_sdma); | |
1020 | dd_dev_info(dd, "SDMA chip_sdma_engines: %u\n", dd->chip_sdma_engines); | |
1021 | dd_dev_info(dd, "SDMA chip_sdma_mem_size: %u\n", | |
1022 | dd->chip_sdma_mem_size); | |
1023 | ||
1024 | per_sdma_credits = | |
1025 | dd->chip_sdma_mem_size/(num_engines * SDMA_BLOCK_SIZE); | |
1026 | ||
1027 | /* set up freeze waitqueue */ | |
1028 | init_waitqueue_head(&dd->sdma_unfreeze_wq); | |
1029 | atomic_set(&dd->sdma_unfreeze_count, 0); | |
1030 | ||
1031 | descq_cnt = sdma_get_descq_cnt(); | |
1032 | dd_dev_info(dd, "SDMA engines %zu descq_cnt %u\n", | |
1033 | num_engines, descq_cnt); | |
1034 | ||
1035 | /* alloc memory for array of send engines */ | |
1036 | dd->per_sdma = kcalloc(num_engines, sizeof(*dd->per_sdma), GFP_KERNEL); | |
1037 | if (!dd->per_sdma) | |
1038 | return -ENOMEM; | |
1039 | ||
1040 | idle_cnt = ns_to_cclock(dd, idle_cnt); | |
1041 | /* Allocate memory for SendDMA descriptor FIFOs */ | |
1042 | for (this_idx = 0; this_idx < num_engines; ++this_idx) { | |
1043 | sde = &dd->per_sdma[this_idx]; | |
1044 | sde->dd = dd; | |
1045 | sde->ppd = ppd; | |
1046 | sde->this_idx = this_idx; | |
1047 | sde->descq_cnt = descq_cnt; | |
1048 | sde->desc_avail = sdma_descq_freecnt(sde); | |
1049 | sde->sdma_shift = ilog2(descq_cnt); | |
1050 | sde->sdma_mask = (1 << sde->sdma_shift) - 1; | |
1051 | sde->descq_full_count = 0; | |
1052 | ||
1053 | /* Create a mask for all 3 chip interrupt sources */ | |
1054 | sde->imask = (u64)1 << (0*TXE_NUM_SDMA_ENGINES + this_idx) | |
1055 | | (u64)1 << (1*TXE_NUM_SDMA_ENGINES + this_idx) | |
1056 | | (u64)1 << (2*TXE_NUM_SDMA_ENGINES + this_idx); | |
1057 | /* Create a mask specifically for sdma_idle */ | |
1058 | sde->idle_mask = | |
1059 | (u64)1 << (2*TXE_NUM_SDMA_ENGINES + this_idx); | |
1060 | /* Create a mask specifically for sdma_progress */ | |
1061 | sde->progress_mask = | |
1062 | (u64)1 << (TXE_NUM_SDMA_ENGINES + this_idx); | |
1063 | spin_lock_init(&sde->tail_lock); | |
1064 | seqlock_init(&sde->head_lock); | |
1065 | spin_lock_init(&sde->senddmactrl_lock); | |
1066 | spin_lock_init(&sde->flushlist_lock); | |
1067 | /* insure there is always a zero bit */ | |
1068 | sde->ahg_bits = 0xfffffffe00000000ULL; | |
1069 | ||
1070 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
1071 | ||
1072 | /* set up reference counting */ | |
1073 | kref_init(&sde->state.kref); | |
1074 | init_completion(&sde->state.comp); | |
1075 | ||
1076 | INIT_LIST_HEAD(&sde->flushlist); | |
1077 | INIT_LIST_HEAD(&sde->dmawait); | |
1078 | ||
1079 | sde->tail_csr = | |
1080 | get_kctxt_csr_addr(dd, this_idx, SD(TAIL)); | |
1081 | ||
1082 | if (idle_cnt) | |
1083 | dd->default_desc1 = | |
1084 | SDMA_DESC1_HEAD_TO_HOST_FLAG; | |
1085 | else | |
1086 | dd->default_desc1 = | |
1087 | SDMA_DESC1_INT_REQ_FLAG; | |
1088 | ||
1089 | tasklet_init(&sde->sdma_hw_clean_up_task, sdma_hw_clean_up_task, | |
1090 | (unsigned long)sde); | |
1091 | ||
1092 | tasklet_init(&sde->sdma_sw_clean_up_task, sdma_sw_clean_up_task, | |
1093 | (unsigned long)sde); | |
1094 | INIT_WORK(&sde->err_halt_worker, sdma_err_halt_wait); | |
1095 | INIT_WORK(&sde->flush_worker, sdma_field_flush); | |
1096 | ||
1097 | sde->progress_check_head = 0; | |
1098 | ||
1099 | init_timer(&sde->err_progress_check_timer); | |
1100 | sde->err_progress_check_timer.function = | |
1101 | sdma_err_progress_check; | |
1102 | sde->err_progress_check_timer.data = (unsigned long)sde; | |
1103 | ||
1104 | sde->descq = dma_zalloc_coherent( | |
1105 | &dd->pcidev->dev, | |
1106 | descq_cnt * sizeof(u64[2]), | |
1107 | &sde->descq_phys, | |
1108 | GFP_KERNEL | |
1109 | ); | |
1110 | if (!sde->descq) | |
1111 | goto bail; | |
1112 | sde->tx_ring = | |
1113 | kcalloc(descq_cnt, sizeof(struct sdma_txreq *), | |
1114 | GFP_KERNEL); | |
1115 | if (!sde->tx_ring) | |
1116 | sde->tx_ring = | |
1117 | vzalloc( | |
1118 | sizeof(struct sdma_txreq *) * | |
1119 | descq_cnt); | |
1120 | if (!sde->tx_ring) | |
1121 | goto bail; | |
1122 | } | |
1123 | ||
1124 | dd->sdma_heads_size = L1_CACHE_BYTES * num_engines; | |
1125 | /* Allocate memory for DMA of head registers to memory */ | |
1126 | dd->sdma_heads_dma = dma_zalloc_coherent( | |
1127 | &dd->pcidev->dev, | |
1128 | dd->sdma_heads_size, | |
1129 | &dd->sdma_heads_phys, | |
1130 | GFP_KERNEL | |
1131 | ); | |
1132 | if (!dd->sdma_heads_dma) { | |
1133 | dd_dev_err(dd, "failed to allocate SendDMA head memory\n"); | |
1134 | goto bail; | |
1135 | } | |
1136 | ||
1137 | /* Allocate memory for pad */ | |
1138 | dd->sdma_pad_dma = dma_zalloc_coherent( | |
1139 | &dd->pcidev->dev, | |
1140 | sizeof(u32), | |
1141 | &dd->sdma_pad_phys, | |
1142 | GFP_KERNEL | |
1143 | ); | |
1144 | if (!dd->sdma_pad_dma) { | |
1145 | dd_dev_err(dd, "failed to allocate SendDMA pad memory\n"); | |
1146 | goto bail; | |
1147 | } | |
1148 | ||
1149 | /* assign each engine to different cacheline and init registers */ | |
1150 | curr_head = (void *)dd->sdma_heads_dma; | |
1151 | for (this_idx = 0; this_idx < num_engines; ++this_idx) { | |
1152 | unsigned long phys_offset; | |
1153 | ||
1154 | sde = &dd->per_sdma[this_idx]; | |
1155 | ||
1156 | sde->head_dma = curr_head; | |
1157 | curr_head += L1_CACHE_BYTES; | |
1158 | phys_offset = (unsigned long)sde->head_dma - | |
1159 | (unsigned long)dd->sdma_heads_dma; | |
1160 | sde->head_phys = dd->sdma_heads_phys + phys_offset; | |
1161 | init_sdma_regs(sde, per_sdma_credits, idle_cnt); | |
1162 | } | |
1163 | dd->flags |= HFI1_HAS_SEND_DMA; | |
1164 | dd->flags |= idle_cnt ? HFI1_HAS_SDMA_TIMEOUT : 0; | |
1165 | dd->num_sdma = num_engines; | |
1166 | if (sdma_map_init(dd, port, ppd->vls_operational, NULL)) | |
1167 | goto bail; | |
1168 | dd_dev_info(dd, "SDMA num_sdma: %u\n", dd->num_sdma); | |
1169 | return 0; | |
1170 | ||
1171 | bail: | |
1172 | sdma_clean(dd, num_engines); | |
1173 | return -ENOMEM; | |
1174 | } | |
1175 | ||
1176 | /** | |
1177 | * sdma_all_running() - called when the link goes up | |
1178 | * @dd: hfi1_devdata | |
1179 | * | |
1180 | * This routine moves all engines to the running state. | |
1181 | */ | |
1182 | void sdma_all_running(struct hfi1_devdata *dd) | |
1183 | { | |
1184 | struct sdma_engine *sde; | |
1185 | unsigned int i; | |
1186 | ||
1187 | /* move all engines to running */ | |
1188 | for (i = 0; i < dd->num_sdma; ++i) { | |
1189 | sde = &dd->per_sdma[i]; | |
1190 | sdma_process_event(sde, sdma_event_e30_go_running); | |
1191 | } | |
1192 | } | |
1193 | ||
1194 | /** | |
1195 | * sdma_all_idle() - called when the link goes down | |
1196 | * @dd: hfi1_devdata | |
1197 | * | |
1198 | * This routine moves all engines to the idle state. | |
1199 | */ | |
1200 | void sdma_all_idle(struct hfi1_devdata *dd) | |
1201 | { | |
1202 | struct sdma_engine *sde; | |
1203 | unsigned int i; | |
1204 | ||
1205 | /* idle all engines */ | |
1206 | for (i = 0; i < dd->num_sdma; ++i) { | |
1207 | sde = &dd->per_sdma[i]; | |
1208 | sdma_process_event(sde, sdma_event_e70_go_idle); | |
1209 | } | |
1210 | } | |
1211 | ||
1212 | /** | |
1213 | * sdma_start() - called to kick off state processing for all engines | |
1214 | * @dd: hfi1_devdata | |
1215 | * | |
1216 | * This routine is for kicking off the state processing for all required | |
1217 | * sdma engines. Interrupts need to be working at this point. | |
1218 | * | |
1219 | */ | |
1220 | void sdma_start(struct hfi1_devdata *dd) | |
1221 | { | |
1222 | unsigned i; | |
1223 | struct sdma_engine *sde; | |
1224 | ||
1225 | /* kick off the engines state processing */ | |
1226 | for (i = 0; i < dd->num_sdma; ++i) { | |
1227 | sde = &dd->per_sdma[i]; | |
1228 | sdma_process_event(sde, sdma_event_e10_go_hw_start); | |
1229 | } | |
1230 | } | |
1231 | ||
1232 | /** | |
1233 | * sdma_exit() - used when module is removed | |
1234 | * @dd: hfi1_devdata | |
1235 | */ | |
1236 | void sdma_exit(struct hfi1_devdata *dd) | |
1237 | { | |
1238 | unsigned this_idx; | |
1239 | struct sdma_engine *sde; | |
1240 | ||
1241 | for (this_idx = 0; dd->per_sdma && this_idx < dd->num_sdma; | |
1242 | ++this_idx) { | |
1243 | ||
1244 | sde = &dd->per_sdma[this_idx]; | |
1245 | if (!list_empty(&sde->dmawait)) | |
1246 | dd_dev_err(dd, "sde %u: dmawait list not empty!\n", | |
1247 | sde->this_idx); | |
1248 | sdma_process_event(sde, sdma_event_e00_go_hw_down); | |
1249 | ||
1250 | del_timer_sync(&sde->err_progress_check_timer); | |
1251 | ||
1252 | /* | |
1253 | * This waits for the state machine to exit so it is not | |
1254 | * necessary to kill the sdma_sw_clean_up_task to make sure | |
1255 | * it is not running. | |
1256 | */ | |
1257 | sdma_finalput(&sde->state); | |
1258 | } | |
1259 | sdma_clean(dd, dd->num_sdma); | |
1260 | } | |
1261 | ||
1262 | /* | |
1263 | * unmap the indicated descriptor | |
1264 | */ | |
1265 | static inline void sdma_unmap_desc( | |
1266 | struct hfi1_devdata *dd, | |
1267 | struct sdma_desc *descp) | |
1268 | { | |
1269 | switch (sdma_mapping_type(descp)) { | |
1270 | case SDMA_MAP_SINGLE: | |
1271 | dma_unmap_single( | |
1272 | &dd->pcidev->dev, | |
1273 | sdma_mapping_addr(descp), | |
1274 | sdma_mapping_len(descp), | |
1275 | DMA_TO_DEVICE); | |
1276 | break; | |
1277 | case SDMA_MAP_PAGE: | |
1278 | dma_unmap_page( | |
1279 | &dd->pcidev->dev, | |
1280 | sdma_mapping_addr(descp), | |
1281 | sdma_mapping_len(descp), | |
1282 | DMA_TO_DEVICE); | |
1283 | break; | |
1284 | } | |
1285 | } | |
1286 | ||
1287 | /* | |
1288 | * return the mode as indicated by the first | |
1289 | * descriptor in the tx. | |
1290 | */ | |
1291 | static inline u8 ahg_mode(struct sdma_txreq *tx) | |
1292 | { | |
1293 | return (tx->descp[0].qw[1] & SDMA_DESC1_HEADER_MODE_SMASK) | |
1294 | >> SDMA_DESC1_HEADER_MODE_SHIFT; | |
1295 | } | |
1296 | ||
1297 | /** | |
1298 | * sdma_txclean() - clean tx of mappings, descp *kmalloc's | |
1299 | * @dd: hfi1_devdata for unmapping | |
1300 | * @tx: tx request to clean | |
1301 | * | |
1302 | * This is used in the progress routine to clean the tx or | |
1303 | * by the ULP to toss an in-process tx build. | |
1304 | * | |
1305 | * The code can be called multiple times without issue. | |
1306 | * | |
1307 | */ | |
1308 | void sdma_txclean( | |
1309 | struct hfi1_devdata *dd, | |
1310 | struct sdma_txreq *tx) | |
1311 | { | |
1312 | u16 i; | |
1313 | ||
1314 | if (tx->num_desc) { | |
1315 | u8 skip = 0, mode = ahg_mode(tx); | |
1316 | ||
1317 | /* unmap first */ | |
1318 | sdma_unmap_desc(dd, &tx->descp[0]); | |
1319 | /* determine number of AHG descriptors to skip */ | |
1320 | if (mode > SDMA_AHG_APPLY_UPDATE1) | |
1321 | skip = mode >> 1; | |
1322 | for (i = 1 + skip; i < tx->num_desc; i++) | |
1323 | sdma_unmap_desc(dd, &tx->descp[i]); | |
1324 | tx->num_desc = 0; | |
1325 | } | |
1326 | kfree(tx->coalesce_buf); | |
1327 | tx->coalesce_buf = NULL; | |
1328 | /* kmalloc'ed descp */ | |
1329 | if (unlikely(tx->desc_limit > ARRAY_SIZE(tx->descs))) { | |
1330 | tx->desc_limit = ARRAY_SIZE(tx->descs); | |
1331 | kfree(tx->descp); | |
1332 | } | |
1333 | } | |
1334 | ||
1335 | static inline u16 sdma_gethead(struct sdma_engine *sde) | |
1336 | { | |
1337 | struct hfi1_devdata *dd = sde->dd; | |
1338 | int use_dmahead; | |
1339 | u16 hwhead; | |
1340 | ||
1341 | #ifdef CONFIG_SDMA_VERBOSITY | |
1342 | dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", | |
1343 | sde->this_idx, slashstrip(__FILE__), __LINE__, __func__); | |
1344 | #endif | |
1345 | ||
1346 | retry: | |
1347 | use_dmahead = HFI1_CAP_IS_KSET(USE_SDMA_HEAD) && __sdma_running(sde) && | |
1348 | (dd->flags & HFI1_HAS_SDMA_TIMEOUT); | |
1349 | hwhead = use_dmahead ? | |
1350 | (u16) le64_to_cpu(*sde->head_dma) : | |
1351 | (u16) read_sde_csr(sde, SD(HEAD)); | |
1352 | ||
1353 | if (unlikely(HFI1_CAP_IS_KSET(SDMA_HEAD_CHECK))) { | |
1354 | u16 cnt; | |
1355 | u16 swtail; | |
1356 | u16 swhead; | |
1357 | int sane; | |
1358 | ||
1359 | swhead = sde->descq_head & sde->sdma_mask; | |
1360 | /* this code is really bad for cache line trading */ | |
1361 | swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask; | |
1362 | cnt = sde->descq_cnt; | |
1363 | ||
1364 | if (swhead < swtail) | |
1365 | /* not wrapped */ | |
1366 | sane = (hwhead >= swhead) & (hwhead <= swtail); | |
1367 | else if (swhead > swtail) | |
1368 | /* wrapped around */ | |
1369 | sane = ((hwhead >= swhead) && (hwhead < cnt)) || | |
1370 | (hwhead <= swtail); | |
1371 | else | |
1372 | /* empty */ | |
1373 | sane = (hwhead == swhead); | |
1374 | ||
1375 | if (unlikely(!sane)) { | |
1376 | dd_dev_err(dd, "SDMA(%u) bad head (%s) hwhd=%hu swhd=%hu swtl=%hu cnt=%hu\n", | |
1377 | sde->this_idx, | |
1378 | use_dmahead ? "dma" : "kreg", | |
1379 | hwhead, swhead, swtail, cnt); | |
1380 | if (use_dmahead) { | |
1381 | /* try one more time, using csr */ | |
1382 | use_dmahead = 0; | |
1383 | goto retry; | |
1384 | } | |
1385 | /* proceed as if no progress */ | |
1386 | hwhead = swhead; | |
1387 | } | |
1388 | } | |
1389 | return hwhead; | |
1390 | } | |
1391 | ||
1392 | /* | |
1393 | * This is called when there are send DMA descriptors that might be | |
1394 | * available. | |
1395 | * | |
1396 | * This is called with head_lock held. | |
1397 | */ | |
1398 | static void sdma_desc_avail(struct sdma_engine *sde, unsigned avail) | |
1399 | { | |
1400 | struct iowait *wait, *nw; | |
1401 | struct iowait *waits[SDMA_WAIT_BATCH_SIZE]; | |
1402 | unsigned i, n = 0, seq; | |
1403 | struct sdma_txreq *stx; | |
1404 | struct hfi1_ibdev *dev = &sde->dd->verbs_dev; | |
1405 | ||
1406 | #ifdef CONFIG_SDMA_VERBOSITY | |
1407 | dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", sde->this_idx, | |
1408 | slashstrip(__FILE__), __LINE__, __func__); | |
1409 | dd_dev_err(sde->dd, "avail: %u\n", avail); | |
1410 | #endif | |
1411 | ||
1412 | do { | |
1413 | seq = read_seqbegin(&dev->iowait_lock); | |
1414 | if (!list_empty(&sde->dmawait)) { | |
1415 | /* at least one item */ | |
1416 | write_seqlock(&dev->iowait_lock); | |
1417 | /* Harvest waiters wanting DMA descriptors */ | |
1418 | list_for_each_entry_safe( | |
1419 | wait, | |
1420 | nw, | |
1421 | &sde->dmawait, | |
1422 | list) { | |
1423 | u16 num_desc = 0; | |
1424 | ||
1425 | if (!wait->wakeup) | |
1426 | continue; | |
1427 | if (n == ARRAY_SIZE(waits)) | |
1428 | break; | |
1429 | if (!list_empty(&wait->tx_head)) { | |
1430 | stx = list_first_entry( | |
1431 | &wait->tx_head, | |
1432 | struct sdma_txreq, | |
1433 | list); | |
1434 | num_desc = stx->num_desc; | |
1435 | } | |
1436 | if (num_desc > avail) | |
1437 | break; | |
1438 | avail -= num_desc; | |
1439 | list_del_init(&wait->list); | |
1440 | waits[n++] = wait; | |
1441 | } | |
1442 | write_sequnlock(&dev->iowait_lock); | |
1443 | break; | |
1444 | } | |
1445 | } while (read_seqretry(&dev->iowait_lock, seq)); | |
1446 | ||
1447 | for (i = 0; i < n; i++) | |
1448 | waits[i]->wakeup(waits[i], SDMA_AVAIL_REASON); | |
1449 | } | |
1450 | ||
1451 | /* head_lock must be held */ | |
1452 | static void sdma_make_progress(struct sdma_engine *sde, u64 status) | |
1453 | { | |
1454 | struct sdma_txreq *txp = NULL; | |
1455 | int progress = 0; | |
1456 | u16 hwhead, swhead, swtail; | |
1457 | int idle_check_done = 0; | |
1458 | ||
1459 | hwhead = sdma_gethead(sde); | |
1460 | ||
1461 | /* The reason for some of the complexity of this code is that | |
1462 | * not all descriptors have corresponding txps. So, we have to | |
1463 | * be able to skip over descs until we wander into the range of | |
1464 | * the next txp on the list. | |
1465 | */ | |
1466 | ||
1467 | retry: | |
1468 | txp = get_txhead(sde); | |
1469 | swhead = sde->descq_head & sde->sdma_mask; | |
1470 | trace_hfi1_sdma_progress(sde, hwhead, swhead, txp); | |
1471 | while (swhead != hwhead) { | |
1472 | /* advance head, wrap if needed */ | |
1473 | swhead = ++sde->descq_head & sde->sdma_mask; | |
1474 | ||
1475 | /* if now past this txp's descs, do the callback */ | |
1476 | if (txp && txp->next_descq_idx == swhead) { | |
1477 | int drained = 0; | |
1478 | /* protect against complete modifying */ | |
1479 | struct iowait *wait = txp->wait; | |
1480 | ||
1481 | /* remove from list */ | |
1482 | sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL; | |
1483 | if (wait) | |
1484 | drained = atomic_dec_and_test(&wait->sdma_busy); | |
1485 | #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER | |
1486 | trace_hfi1_sdma_out_sn(sde, txp->sn); | |
1487 | if (WARN_ON_ONCE(sde->head_sn != txp->sn)) | |
1488 | dd_dev_err(sde->dd, "expected %llu got %llu\n", | |
1489 | sde->head_sn, txp->sn); | |
1490 | sde->head_sn++; | |
1491 | #endif | |
1492 | sdma_txclean(sde->dd, txp); | |
1493 | if (txp->complete) | |
1494 | (*txp->complete)( | |
1495 | txp, | |
1496 | SDMA_TXREQ_S_OK, | |
1497 | drained); | |
1498 | if (wait && drained) | |
1499 | iowait_drain_wakeup(wait); | |
1500 | /* see if there is another txp */ | |
1501 | txp = get_txhead(sde); | |
1502 | } | |
1503 | trace_hfi1_sdma_progress(sde, hwhead, swhead, txp); | |
1504 | progress++; | |
1505 | } | |
1506 | ||
1507 | /* | |
1508 | * The SDMA idle interrupt is not guaranteed to be ordered with respect | |
1509 | * to updates to the the dma_head location in host memory. The head | |
1510 | * value read might not be fully up to date. If there are pending | |
1511 | * descriptors and the SDMA idle interrupt fired then read from the | |
1512 | * CSR SDMA head instead to get the latest value from the hardware. | |
1513 | * The hardware SDMA head should be read at most once in this invocation | |
1514 | * of sdma_make_progress(..) which is ensured by idle_check_done flag | |
1515 | */ | |
1516 | if ((status & sde->idle_mask) && !idle_check_done) { | |
1517 | swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask; | |
1518 | if (swtail != hwhead) { | |
1519 | hwhead = (u16)read_sde_csr(sde, SD(HEAD)); | |
1520 | idle_check_done = 1; | |
1521 | goto retry; | |
1522 | } | |
1523 | } | |
1524 | ||
1525 | sde->last_status = status; | |
1526 | if (progress) | |
1527 | sdma_desc_avail(sde, sdma_descq_freecnt(sde)); | |
1528 | } | |
1529 | ||
1530 | /* | |
1531 | * sdma_engine_interrupt() - interrupt handler for engine | |
1532 | * @sde: sdma engine | |
1533 | * @status: sdma interrupt reason | |
1534 | * | |
1535 | * Status is a mask of the 3 possible interrupts for this engine. It will | |
1536 | * contain bits _only_ for this SDMA engine. It will contain at least one | |
1537 | * bit, it may contain more. | |
1538 | */ | |
1539 | void sdma_engine_interrupt(struct sdma_engine *sde, u64 status) | |
1540 | { | |
1541 | trace_hfi1_sdma_engine_interrupt(sde, status); | |
1542 | write_seqlock(&sde->head_lock); | |
1543 | sdma_set_desc_cnt(sde, sde->descq_cnt / 2); | |
1544 | sdma_make_progress(sde, status); | |
1545 | write_sequnlock(&sde->head_lock); | |
1546 | } | |
1547 | ||
1548 | /** | |
1549 | * sdma_engine_error() - error handler for engine | |
1550 | * @sde: sdma engine | |
1551 | * @status: sdma interrupt reason | |
1552 | */ | |
1553 | void sdma_engine_error(struct sdma_engine *sde, u64 status) | |
1554 | { | |
1555 | unsigned long flags; | |
1556 | ||
1557 | #ifdef CONFIG_SDMA_VERBOSITY | |
1558 | dd_dev_err(sde->dd, "CONFIG SDMA(%u) error status 0x%llx state %s\n", | |
1559 | sde->this_idx, | |
1560 | (unsigned long long)status, | |
1561 | sdma_state_names[sde->state.current_state]); | |
1562 | #endif | |
1563 | spin_lock_irqsave(&sde->tail_lock, flags); | |
1564 | write_seqlock(&sde->head_lock); | |
1565 | if (status & ALL_SDMA_ENG_HALT_ERRS) | |
1566 | __sdma_process_event(sde, sdma_event_e60_hw_halted); | |
1567 | if (status & ~SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK)) { | |
1568 | dd_dev_err(sde->dd, | |
1569 | "SDMA (%u) engine error: 0x%llx state %s\n", | |
1570 | sde->this_idx, | |
1571 | (unsigned long long)status, | |
1572 | sdma_state_names[sde->state.current_state]); | |
1573 | dump_sdma_state(sde); | |
1574 | } | |
1575 | write_sequnlock(&sde->head_lock); | |
1576 | spin_unlock_irqrestore(&sde->tail_lock, flags); | |
1577 | } | |
1578 | ||
1579 | static void sdma_sendctrl(struct sdma_engine *sde, unsigned op) | |
1580 | { | |
1581 | u64 set_senddmactrl = 0; | |
1582 | u64 clr_senddmactrl = 0; | |
1583 | unsigned long flags; | |
1584 | ||
1585 | #ifdef CONFIG_SDMA_VERBOSITY | |
1586 | dd_dev_err(sde->dd, "CONFIG SDMA(%u) senddmactrl E=%d I=%d H=%d C=%d\n", | |
1587 | sde->this_idx, | |
1588 | (op & SDMA_SENDCTRL_OP_ENABLE) ? 1 : 0, | |
1589 | (op & SDMA_SENDCTRL_OP_INTENABLE) ? 1 : 0, | |
1590 | (op & SDMA_SENDCTRL_OP_HALT) ? 1 : 0, | |
1591 | (op & SDMA_SENDCTRL_OP_CLEANUP) ? 1 : 0); | |
1592 | #endif | |
1593 | ||
1594 | if (op & SDMA_SENDCTRL_OP_ENABLE) | |
1595 | set_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK); | |
1596 | else | |
1597 | clr_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK); | |
1598 | ||
1599 | if (op & SDMA_SENDCTRL_OP_INTENABLE) | |
1600 | set_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK); | |
1601 | else | |
1602 | clr_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK); | |
1603 | ||
1604 | if (op & SDMA_SENDCTRL_OP_HALT) | |
1605 | set_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK); | |
1606 | else | |
1607 | clr_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK); | |
1608 | ||
1609 | spin_lock_irqsave(&sde->senddmactrl_lock, flags); | |
1610 | ||
1611 | sde->p_senddmactrl |= set_senddmactrl; | |
1612 | sde->p_senddmactrl &= ~clr_senddmactrl; | |
1613 | ||
1614 | if (op & SDMA_SENDCTRL_OP_CLEANUP) | |
1615 | write_sde_csr(sde, SD(CTRL), | |
1616 | sde->p_senddmactrl | | |
1617 | SD(CTRL_SDMA_CLEANUP_SMASK)); | |
1618 | else | |
1619 | write_sde_csr(sde, SD(CTRL), sde->p_senddmactrl); | |
1620 | ||
1621 | spin_unlock_irqrestore(&sde->senddmactrl_lock, flags); | |
1622 | ||
1623 | #ifdef CONFIG_SDMA_VERBOSITY | |
1624 | sdma_dumpstate(sde); | |
1625 | #endif | |
1626 | } | |
1627 | ||
1628 | static void sdma_setlengen(struct sdma_engine *sde) | |
1629 | { | |
1630 | #ifdef CONFIG_SDMA_VERBOSITY | |
1631 | dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", | |
1632 | sde->this_idx, slashstrip(__FILE__), __LINE__, __func__); | |
1633 | #endif | |
1634 | ||
1635 | /* | |
1636 | * Set SendDmaLenGen and clear-then-set the MSB of the generation | |
1637 | * count to enable generation checking and load the internal | |
1638 | * generation counter. | |
1639 | */ | |
1640 | write_sde_csr(sde, SD(LEN_GEN), | |
1641 | (sde->descq_cnt/64) << SD(LEN_GEN_LENGTH_SHIFT) | |
1642 | ); | |
1643 | write_sde_csr(sde, SD(LEN_GEN), | |
1644 | ((sde->descq_cnt/64) << SD(LEN_GEN_LENGTH_SHIFT)) | |
1645 | | (4ULL << SD(LEN_GEN_GENERATION_SHIFT)) | |
1646 | ); | |
1647 | } | |
1648 | ||
1649 | static inline void sdma_update_tail(struct sdma_engine *sde, u16 tail) | |
1650 | { | |
1651 | /* Commit writes to memory and advance the tail on the chip */ | |
1652 | smp_wmb(); /* see get_txhead() */ | |
1653 | writeq(tail, sde->tail_csr); | |
1654 | } | |
1655 | ||
1656 | /* | |
1657 | * This is called when changing to state s10_hw_start_up_halt_wait as | |
1658 | * a result of send buffer errors or send DMA descriptor errors. | |
1659 | */ | |
1660 | static void sdma_hw_start_up(struct sdma_engine *sde) | |
1661 | { | |
1662 | u64 reg; | |
1663 | ||
1664 | #ifdef CONFIG_SDMA_VERBOSITY | |
1665 | dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", | |
1666 | sde->this_idx, slashstrip(__FILE__), __LINE__, __func__); | |
1667 | #endif | |
1668 | ||
1669 | sdma_setlengen(sde); | |
1670 | sdma_update_tail(sde, 0); /* Set SendDmaTail */ | |
1671 | *sde->head_dma = 0; | |
1672 | ||
1673 | reg = SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_MASK) << | |
1674 | SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SHIFT); | |
1675 | write_sde_csr(sde, SD(ENG_ERR_CLEAR), reg); | |
1676 | } | |
1677 | ||
1678 | #define CLEAR_STATIC_RATE_CONTROL_SMASK(r) \ | |
1679 | (r &= ~SEND_DMA_CHECK_ENABLE_DISALLOW_PBC_STATIC_RATE_CONTROL_SMASK) | |
1680 | ||
1681 | #define SET_STATIC_RATE_CONTROL_SMASK(r) \ | |
1682 | (r |= SEND_DMA_CHECK_ENABLE_DISALLOW_PBC_STATIC_RATE_CONTROL_SMASK) | |
1683 | /* | |
1684 | * set_sdma_integrity | |
1685 | * | |
1686 | * Set the SEND_DMA_CHECK_ENABLE register for send DMA engine 'sde'. | |
1687 | */ | |
1688 | static void set_sdma_integrity(struct sdma_engine *sde) | |
1689 | { | |
1690 | struct hfi1_devdata *dd = sde->dd; | |
1691 | u64 reg; | |
1692 | ||
1693 | if (unlikely(HFI1_CAP_IS_KSET(NO_INTEGRITY))) | |
1694 | return; | |
1695 | ||
1696 | reg = hfi1_pkt_base_sdma_integrity(dd); | |
1697 | ||
1698 | if (HFI1_CAP_IS_KSET(STATIC_RATE_CTRL)) | |
1699 | CLEAR_STATIC_RATE_CONTROL_SMASK(reg); | |
1700 | else | |
1701 | SET_STATIC_RATE_CONTROL_SMASK(reg); | |
1702 | ||
1703 | write_sde_csr(sde, SD(CHECK_ENABLE), reg); | |
1704 | } | |
1705 | ||
1706 | ||
1707 | static void init_sdma_regs( | |
1708 | struct sdma_engine *sde, | |
1709 | u32 credits, | |
1710 | uint idle_cnt) | |
1711 | { | |
1712 | u8 opval, opmask; | |
1713 | #ifdef CONFIG_SDMA_VERBOSITY | |
1714 | struct hfi1_devdata *dd = sde->dd; | |
1715 | ||
1716 | dd_dev_err(dd, "CONFIG SDMA(%u) %s:%d %s()\n", | |
1717 | sde->this_idx, slashstrip(__FILE__), __LINE__, __func__); | |
1718 | #endif | |
1719 | ||
1720 | write_sde_csr(sde, SD(BASE_ADDR), sde->descq_phys); | |
1721 | sdma_setlengen(sde); | |
1722 | sdma_update_tail(sde, 0); /* Set SendDmaTail */ | |
1723 | write_sde_csr(sde, SD(RELOAD_CNT), idle_cnt); | |
1724 | write_sde_csr(sde, SD(DESC_CNT), 0); | |
1725 | write_sde_csr(sde, SD(HEAD_ADDR), sde->head_phys); | |
1726 | write_sde_csr(sde, SD(MEMORY), | |
1727 | ((u64)credits << | |
1728 | SD(MEMORY_SDMA_MEMORY_CNT_SHIFT)) | | |
1729 | ((u64)(credits * sde->this_idx) << | |
1730 | SD(MEMORY_SDMA_MEMORY_INDEX_SHIFT))); | |
1731 | write_sde_csr(sde, SD(ENG_ERR_MASK), ~0ull); | |
1732 | set_sdma_integrity(sde); | |
1733 | opmask = OPCODE_CHECK_MASK_DISABLED; | |
1734 | opval = OPCODE_CHECK_VAL_DISABLED; | |
1735 | write_sde_csr(sde, SD(CHECK_OPCODE), | |
1736 | (opmask << SEND_CTXT_CHECK_OPCODE_MASK_SHIFT) | | |
1737 | (opval << SEND_CTXT_CHECK_OPCODE_VALUE_SHIFT)); | |
1738 | } | |
1739 | ||
1740 | #ifdef CONFIG_SDMA_VERBOSITY | |
1741 | ||
1742 | #define sdma_dumpstate_helper0(reg) do { \ | |
1743 | csr = read_csr(sde->dd, reg); \ | |
1744 | dd_dev_err(sde->dd, "%36s 0x%016llx\n", #reg, csr); \ | |
1745 | } while (0) | |
1746 | ||
1747 | #define sdma_dumpstate_helper(reg) do { \ | |
1748 | csr = read_sde_csr(sde, reg); \ | |
1749 | dd_dev_err(sde->dd, "%36s[%02u] 0x%016llx\n", \ | |
1750 | #reg, sde->this_idx, csr); \ | |
1751 | } while (0) | |
1752 | ||
1753 | #define sdma_dumpstate_helper2(reg) do { \ | |
1754 | csr = read_csr(sde->dd, reg + (8 * i)); \ | |
1755 | dd_dev_err(sde->dd, "%33s_%02u 0x%016llx\n", \ | |
1756 | #reg, i, csr); \ | |
1757 | } while (0) | |
1758 | ||
1759 | void sdma_dumpstate(struct sdma_engine *sde) | |
1760 | { | |
1761 | u64 csr; | |
1762 | unsigned i; | |
1763 | ||
1764 | sdma_dumpstate_helper(SD(CTRL)); | |
1765 | sdma_dumpstate_helper(SD(STATUS)); | |
1766 | sdma_dumpstate_helper0(SD(ERR_STATUS)); | |
1767 | sdma_dumpstate_helper0(SD(ERR_MASK)); | |
1768 | sdma_dumpstate_helper(SD(ENG_ERR_STATUS)); | |
1769 | sdma_dumpstate_helper(SD(ENG_ERR_MASK)); | |
1770 | ||
1771 | for (i = 0; i < CCE_NUM_INT_CSRS; ++i) { | |
6fd8edab | 1772 | sdma_dumpstate_helper2(CCE_INT_STATUS); |
77241056 MM |
1773 | sdma_dumpstate_helper2(CCE_INT_MASK); |
1774 | sdma_dumpstate_helper2(CCE_INT_BLOCKED); | |
1775 | } | |
1776 | ||
1777 | sdma_dumpstate_helper(SD(TAIL)); | |
1778 | sdma_dumpstate_helper(SD(HEAD)); | |
1779 | sdma_dumpstate_helper(SD(PRIORITY_THLD)); | |
6fd8edab | 1780 | sdma_dumpstate_helper(SD(IDLE_CNT)); |
77241056 MM |
1781 | sdma_dumpstate_helper(SD(RELOAD_CNT)); |
1782 | sdma_dumpstate_helper(SD(DESC_CNT)); | |
1783 | sdma_dumpstate_helper(SD(DESC_FETCHED_CNT)); | |
1784 | sdma_dumpstate_helper(SD(MEMORY)); | |
1785 | sdma_dumpstate_helper0(SD(ENGINES)); | |
1786 | sdma_dumpstate_helper0(SD(MEM_SIZE)); | |
1787 | /* sdma_dumpstate_helper(SEND_EGRESS_SEND_DMA_STATUS); */ | |
1788 | sdma_dumpstate_helper(SD(BASE_ADDR)); | |
1789 | sdma_dumpstate_helper(SD(LEN_GEN)); | |
1790 | sdma_dumpstate_helper(SD(HEAD_ADDR)); | |
1791 | sdma_dumpstate_helper(SD(CHECK_ENABLE)); | |
1792 | sdma_dumpstate_helper(SD(CHECK_VL)); | |
1793 | sdma_dumpstate_helper(SD(CHECK_JOB_KEY)); | |
1794 | sdma_dumpstate_helper(SD(CHECK_PARTITION_KEY)); | |
1795 | sdma_dumpstate_helper(SD(CHECK_SLID)); | |
1796 | sdma_dumpstate_helper(SD(CHECK_OPCODE)); | |
1797 | } | |
1798 | #endif | |
1799 | ||
1800 | static void dump_sdma_state(struct sdma_engine *sde) | |
1801 | { | |
1802 | struct hw_sdma_desc *descq; | |
1803 | struct hw_sdma_desc *descqp; | |
1804 | u64 desc[2]; | |
1805 | u64 addr; | |
1806 | u8 gen; | |
1807 | u16 len; | |
1808 | u16 head, tail, cnt; | |
1809 | ||
1810 | head = sde->descq_head & sde->sdma_mask; | |
1811 | tail = sde->descq_tail & sde->sdma_mask; | |
1812 | cnt = sdma_descq_freecnt(sde); | |
1813 | descq = sde->descq; | |
1814 | ||
1815 | dd_dev_err(sde->dd, | |
1816 | "SDMA (%u) descq_head: %u descq_tail: %u freecnt: %u FLE %d\n", | |
1817 | sde->this_idx, | |
1818 | head, | |
1819 | tail, | |
1820 | cnt, | |
1821 | !list_empty(&sde->flushlist)); | |
1822 | ||
1823 | /* print info for each entry in the descriptor queue */ | |
1824 | while (head != tail) { | |
1825 | char flags[6] = { 'x', 'x', 'x', 'x', 0 }; | |
1826 | ||
1827 | descqp = &sde->descq[head]; | |
1828 | desc[0] = le64_to_cpu(descqp->qw[0]); | |
1829 | desc[1] = le64_to_cpu(descqp->qw[1]); | |
1830 | flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-'; | |
1831 | flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ? | |
1832 | 'H' : '-'; | |
1833 | flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-'; | |
1834 | flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-'; | |
1835 | addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT) | |
1836 | & SDMA_DESC0_PHY_ADDR_MASK; | |
1837 | gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT) | |
1838 | & SDMA_DESC1_GENERATION_MASK; | |
1839 | len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT) | |
1840 | & SDMA_DESC0_BYTE_COUNT_MASK; | |
1841 | dd_dev_err(sde->dd, | |
1842 | "SDMA sdmadesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n", | |
1843 | head, flags, addr, gen, len); | |
1844 | dd_dev_err(sde->dd, | |
1845 | "\tdesc0:0x%016llx desc1 0x%016llx\n", | |
1846 | desc[0], desc[1]); | |
1847 | if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) | |
1848 | dd_dev_err(sde->dd, | |
1849 | "\taidx: %u amode: %u alen: %u\n", | |
1850 | (u8)((desc[1] & SDMA_DESC1_HEADER_INDEX_SMASK) | |
7d630467 | 1851 | >> SDMA_DESC1_HEADER_INDEX_SHIFT), |
77241056 MM |
1852 | (u8)((desc[1] & SDMA_DESC1_HEADER_MODE_SMASK) |
1853 | >> SDMA_DESC1_HEADER_MODE_SHIFT), | |
1854 | (u8)((desc[1] & SDMA_DESC1_HEADER_DWS_SMASK) | |
1855 | >> SDMA_DESC1_HEADER_DWS_SHIFT)); | |
1856 | head++; | |
1857 | head &= sde->sdma_mask; | |
1858 | } | |
1859 | } | |
1860 | ||
1861 | #define SDE_FMT \ | |
1862 | "SDE %u STE %s C 0x%llx S 0x%016llx E 0x%llx T(HW) 0x%llx T(SW) 0x%x H(HW) 0x%llx H(SW) 0x%x H(D) 0x%llx DM 0x%llx GL 0x%llx R 0x%llx LIS 0x%llx AHGI 0x%llx TXT %u TXH %u DT %u DH %u FLNE %d DQF %u SLC 0x%llx\n" | |
1863 | /** | |
1864 | * sdma_seqfile_dump_sde() - debugfs dump of sde | |
1865 | * @s: seq file | |
1866 | * @sde: send dma engine to dump | |
1867 | * | |
1868 | * This routine dumps the sde to the indicated seq file. | |
1869 | */ | |
1870 | void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *sde) | |
1871 | { | |
1872 | u16 head, tail; | |
1873 | struct hw_sdma_desc *descqp; | |
1874 | u64 desc[2]; | |
1875 | u64 addr; | |
1876 | u8 gen; | |
1877 | u16 len; | |
1878 | ||
1879 | head = sde->descq_head & sde->sdma_mask; | |
1880 | tail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask; | |
1881 | seq_printf(s, SDE_FMT, sde->this_idx, | |
1882 | sdma_state_name(sde->state.current_state), | |
1883 | (unsigned long long)read_sde_csr(sde, SD(CTRL)), | |
1884 | (unsigned long long)read_sde_csr(sde, SD(STATUS)), | |
1885 | (unsigned long long)read_sde_csr(sde, | |
1886 | SD(ENG_ERR_STATUS)), | |
1887 | (unsigned long long)read_sde_csr(sde, SD(TAIL)), | |
1888 | tail, | |
1889 | (unsigned long long)read_sde_csr(sde, SD(HEAD)), | |
1890 | head, | |
1891 | (unsigned long long)le64_to_cpu(*sde->head_dma), | |
1892 | (unsigned long long)read_sde_csr(sde, SD(MEMORY)), | |
1893 | (unsigned long long)read_sde_csr(sde, SD(LEN_GEN)), | |
1894 | (unsigned long long)read_sde_csr(sde, SD(RELOAD_CNT)), | |
1895 | (unsigned long long)sde->last_status, | |
1896 | (unsigned long long)sde->ahg_bits, | |
1897 | sde->tx_tail, | |
1898 | sde->tx_head, | |
1899 | sde->descq_tail, | |
1900 | sde->descq_head, | |
1901 | !list_empty(&sde->flushlist), | |
1902 | sde->descq_full_count, | |
1903 | (unsigned long long)read_sde_csr(sde, SEND_DMA_CHECK_SLID)); | |
1904 | ||
1905 | /* print info for each entry in the descriptor queue */ | |
1906 | while (head != tail) { | |
1907 | char flags[6] = { 'x', 'x', 'x', 'x', 0 }; | |
1908 | ||
1909 | descqp = &sde->descq[head]; | |
1910 | desc[0] = le64_to_cpu(descqp->qw[0]); | |
1911 | desc[1] = le64_to_cpu(descqp->qw[1]); | |
1912 | flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-'; | |
1913 | flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ? | |
1914 | 'H' : '-'; | |
1915 | flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-'; | |
1916 | flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-'; | |
1917 | addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT) | |
1918 | & SDMA_DESC0_PHY_ADDR_MASK; | |
1919 | gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT) | |
1920 | & SDMA_DESC1_GENERATION_MASK; | |
1921 | len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT) | |
1922 | & SDMA_DESC0_BYTE_COUNT_MASK; | |
1923 | seq_printf(s, | |
1924 | "\tdesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n", | |
1925 | head, flags, addr, gen, len); | |
1926 | if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) | |
1927 | seq_printf(s, "\t\tahgidx: %u ahgmode: %u\n", | |
1928 | (u8)((desc[1] & SDMA_DESC1_HEADER_INDEX_SMASK) | |
7d630467 | 1929 | >> SDMA_DESC1_HEADER_INDEX_SHIFT), |
77241056 MM |
1930 | (u8)((desc[1] & SDMA_DESC1_HEADER_MODE_SMASK) |
1931 | >> SDMA_DESC1_HEADER_MODE_SHIFT)); | |
1932 | head = (head + 1) & sde->sdma_mask; | |
1933 | } | |
1934 | } | |
1935 | ||
1936 | /* | |
1937 | * add the generation number into | |
1938 | * the qw1 and return | |
1939 | */ | |
1940 | static inline u64 add_gen(struct sdma_engine *sde, u64 qw1) | |
1941 | { | |
1942 | u8 generation = (sde->descq_tail >> sde->sdma_shift) & 3; | |
1943 | ||
1944 | qw1 &= ~SDMA_DESC1_GENERATION_SMASK; | |
1945 | qw1 |= ((u64)generation & SDMA_DESC1_GENERATION_MASK) | |
1946 | << SDMA_DESC1_GENERATION_SHIFT; | |
1947 | return qw1; | |
1948 | } | |
1949 | ||
1950 | /* | |
1951 | * This routine submits the indicated tx | |
1952 | * | |
1953 | * Space has already been guaranteed and | |
1954 | * tail side of ring is locked. | |
1955 | * | |
1956 | * The hardware tail update is done | |
1957 | * in the caller and that is facilitated | |
1958 | * by returning the new tail. | |
1959 | * | |
1960 | * There is special case logic for ahg | |
1961 | * to not add the generation number for | |
1962 | * up to 2 descriptors that follow the | |
1963 | * first descriptor. | |
1964 | * | |
1965 | */ | |
1966 | static inline u16 submit_tx(struct sdma_engine *sde, struct sdma_txreq *tx) | |
1967 | { | |
1968 | int i; | |
1969 | u16 tail; | |
1970 | struct sdma_desc *descp = tx->descp; | |
1971 | u8 skip = 0, mode = ahg_mode(tx); | |
1972 | ||
1973 | tail = sde->descq_tail & sde->sdma_mask; | |
1974 | sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]); | |
1975 | sde->descq[tail].qw[1] = cpu_to_le64(add_gen(sde, descp->qw[1])); | |
1976 | trace_hfi1_sdma_descriptor(sde, descp->qw[0], descp->qw[1], | |
1977 | tail, &sde->descq[tail]); | |
1978 | tail = ++sde->descq_tail & sde->sdma_mask; | |
1979 | descp++; | |
1980 | if (mode > SDMA_AHG_APPLY_UPDATE1) | |
1981 | skip = mode >> 1; | |
1982 | for (i = 1; i < tx->num_desc; i++, descp++) { | |
1983 | u64 qw1; | |
1984 | ||
1985 | sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]); | |
1986 | if (skip) { | |
1987 | /* edits don't have generation */ | |
1988 | qw1 = descp->qw[1]; | |
1989 | skip--; | |
1990 | } else { | |
1991 | /* replace generation with real one for non-edits */ | |
1992 | qw1 = add_gen(sde, descp->qw[1]); | |
1993 | } | |
1994 | sde->descq[tail].qw[1] = cpu_to_le64(qw1); | |
1995 | trace_hfi1_sdma_descriptor(sde, descp->qw[0], qw1, | |
1996 | tail, &sde->descq[tail]); | |
1997 | tail = ++sde->descq_tail & sde->sdma_mask; | |
1998 | } | |
1999 | tx->next_descq_idx = tail; | |
2000 | #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER | |
2001 | tx->sn = sde->tail_sn++; | |
2002 | trace_hfi1_sdma_in_sn(sde, tx->sn); | |
2003 | WARN_ON_ONCE(sde->tx_ring[sde->tx_tail & sde->sdma_mask]); | |
2004 | #endif | |
2005 | sde->tx_ring[sde->tx_tail++ & sde->sdma_mask] = tx; | |
2006 | sde->desc_avail -= tx->num_desc; | |
2007 | return tail; | |
2008 | } | |
2009 | ||
2010 | /* | |
2011 | * Check for progress | |
2012 | */ | |
2013 | static int sdma_check_progress( | |
2014 | struct sdma_engine *sde, | |
2015 | struct iowait *wait, | |
2016 | struct sdma_txreq *tx) | |
2017 | { | |
2018 | int ret; | |
2019 | ||
2020 | sde->desc_avail = sdma_descq_freecnt(sde); | |
2021 | if (tx->num_desc <= sde->desc_avail) | |
2022 | return -EAGAIN; | |
2023 | /* pulse the head_lock */ | |
2024 | if (wait && wait->sleep) { | |
2025 | unsigned seq; | |
2026 | ||
2027 | seq = raw_seqcount_begin( | |
2028 | (const seqcount_t *)&sde->head_lock.seqcount); | |
2029 | ret = wait->sleep(sde, wait, tx, seq); | |
2030 | if (ret == -EAGAIN) | |
2031 | sde->desc_avail = sdma_descq_freecnt(sde); | |
2032 | } else | |
2033 | ret = -EBUSY; | |
2034 | return ret; | |
2035 | } | |
2036 | ||
2037 | /** | |
2038 | * sdma_send_txreq() - submit a tx req to ring | |
2039 | * @sde: sdma engine to use | |
2040 | * @wait: wait structure to use when full (may be NULL) | |
2041 | * @tx: sdma_txreq to submit | |
2042 | * | |
2043 | * The call submits the tx into the ring. If a iowait structure is non-NULL | |
2044 | * the packet will be queued to the list in wait. | |
2045 | * | |
2046 | * Return: | |
2047 | * 0 - Success, -EINVAL - sdma_txreq incomplete, -EBUSY - no space in | |
2048 | * ring (wait == NULL) | |
2049 | * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state | |
2050 | */ | |
2051 | int sdma_send_txreq(struct sdma_engine *sde, | |
2052 | struct iowait *wait, | |
2053 | struct sdma_txreq *tx) | |
2054 | { | |
2055 | int ret = 0; | |
2056 | u16 tail; | |
2057 | unsigned long flags; | |
2058 | ||
2059 | /* user should have supplied entire packet */ | |
2060 | if (unlikely(tx->tlen)) | |
2061 | return -EINVAL; | |
2062 | tx->wait = wait; | |
2063 | spin_lock_irqsave(&sde->tail_lock, flags); | |
2064 | retry: | |
2065 | if (unlikely(!__sdma_running(sde))) | |
2066 | goto unlock_noconn; | |
2067 | if (unlikely(tx->num_desc > sde->desc_avail)) | |
2068 | goto nodesc; | |
2069 | tail = submit_tx(sde, tx); | |
2070 | if (wait) | |
2071 | atomic_inc(&wait->sdma_busy); | |
2072 | sdma_update_tail(sde, tail); | |
2073 | unlock: | |
2074 | spin_unlock_irqrestore(&sde->tail_lock, flags); | |
2075 | return ret; | |
2076 | unlock_noconn: | |
2077 | if (wait) | |
2078 | atomic_inc(&wait->sdma_busy); | |
2079 | tx->next_descq_idx = 0; | |
2080 | #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER | |
2081 | tx->sn = sde->tail_sn++; | |
2082 | trace_hfi1_sdma_in_sn(sde, tx->sn); | |
2083 | #endif | |
2084 | spin_lock(&sde->flushlist_lock); | |
2085 | list_add_tail(&tx->list, &sde->flushlist); | |
2086 | spin_unlock(&sde->flushlist_lock); | |
2087 | if (wait) { | |
2088 | wait->tx_count++; | |
2089 | wait->count += tx->num_desc; | |
2090 | } | |
2091 | schedule_work(&sde->flush_worker); | |
2092 | ret = -ECOMM; | |
2093 | goto unlock; | |
2094 | nodesc: | |
2095 | ret = sdma_check_progress(sde, wait, tx); | |
2096 | if (ret == -EAGAIN) { | |
2097 | ret = 0; | |
2098 | goto retry; | |
2099 | } | |
2100 | sde->descq_full_count++; | |
2101 | goto unlock; | |
2102 | } | |
2103 | ||
2104 | /** | |
2105 | * sdma_send_txlist() - submit a list of tx req to ring | |
2106 | * @sde: sdma engine to use | |
2107 | * @wait: wait structure to use when full (may be NULL) | |
2108 | * @tx_list: list of sdma_txreqs to submit | |
2109 | * | |
2110 | * The call submits the list into the ring. | |
2111 | * | |
2112 | * If the iowait structure is non-NULL and not equal to the iowait list | |
2113 | * the unprocessed part of the list will be appended to the list in wait. | |
2114 | * | |
2115 | * In all cases, the tx_list will be updated so the head of the tx_list is | |
2116 | * the list of descriptors that have yet to be transmitted. | |
2117 | * | |
2118 | * The intent of this call is to provide a more efficient | |
2119 | * way of submitting multiple packets to SDMA while holding the tail | |
2120 | * side locking. | |
2121 | * | |
2122 | * Return: | |
2123 | * 0 - Success, -EINVAL - sdma_txreq incomplete, -EBUSY - no space in ring | |
2124 | * (wait == NULL) | |
2125 | * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state | |
2126 | */ | |
2127 | int sdma_send_txlist(struct sdma_engine *sde, | |
2128 | struct iowait *wait, | |
2129 | struct list_head *tx_list) | |
2130 | { | |
2131 | struct sdma_txreq *tx, *tx_next; | |
2132 | int ret = 0; | |
2133 | unsigned long flags; | |
2134 | u16 tail = INVALID_TAIL; | |
2135 | int count = 0; | |
2136 | ||
2137 | spin_lock_irqsave(&sde->tail_lock, flags); | |
2138 | retry: | |
2139 | list_for_each_entry_safe(tx, tx_next, tx_list, list) { | |
2140 | tx->wait = wait; | |
2141 | if (unlikely(!__sdma_running(sde))) | |
2142 | goto unlock_noconn; | |
2143 | if (unlikely(tx->num_desc > sde->desc_avail)) | |
2144 | goto nodesc; | |
2145 | if (unlikely(tx->tlen)) { | |
2146 | ret = -EINVAL; | |
2147 | goto update_tail; | |
2148 | } | |
2149 | list_del_init(&tx->list); | |
2150 | tail = submit_tx(sde, tx); | |
2151 | count++; | |
2152 | if (tail != INVALID_TAIL && | |
2153 | (count & SDMA_TAIL_UPDATE_THRESH) == 0) { | |
2154 | sdma_update_tail(sde, tail); | |
2155 | tail = INVALID_TAIL; | |
2156 | } | |
2157 | } | |
2158 | update_tail: | |
2159 | if (wait) | |
2160 | atomic_add(count, &wait->sdma_busy); | |
2161 | if (tail != INVALID_TAIL) | |
2162 | sdma_update_tail(sde, tail); | |
2163 | spin_unlock_irqrestore(&sde->tail_lock, flags); | |
2164 | return ret; | |
2165 | unlock_noconn: | |
2166 | spin_lock(&sde->flushlist_lock); | |
2167 | list_for_each_entry_safe(tx, tx_next, tx_list, list) { | |
2168 | tx->wait = wait; | |
2169 | list_del_init(&tx->list); | |
2170 | if (wait) | |
2171 | atomic_inc(&wait->sdma_busy); | |
2172 | tx->next_descq_idx = 0; | |
2173 | #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER | |
2174 | tx->sn = sde->tail_sn++; | |
2175 | trace_hfi1_sdma_in_sn(sde, tx->sn); | |
2176 | #endif | |
2177 | list_add_tail(&tx->list, &sde->flushlist); | |
2178 | if (wait) { | |
2179 | wait->tx_count++; | |
2180 | wait->count += tx->num_desc; | |
2181 | } | |
2182 | } | |
2183 | spin_unlock(&sde->flushlist_lock); | |
2184 | schedule_work(&sde->flush_worker); | |
2185 | ret = -ECOMM; | |
2186 | goto update_tail; | |
2187 | nodesc: | |
2188 | ret = sdma_check_progress(sde, wait, tx); | |
2189 | if (ret == -EAGAIN) { | |
2190 | ret = 0; | |
2191 | goto retry; | |
2192 | } | |
2193 | sde->descq_full_count++; | |
2194 | goto update_tail; | |
2195 | } | |
2196 | ||
2197 | static void sdma_process_event(struct sdma_engine *sde, | |
2198 | enum sdma_events event) | |
2199 | { | |
2200 | unsigned long flags; | |
2201 | ||
2202 | spin_lock_irqsave(&sde->tail_lock, flags); | |
2203 | write_seqlock(&sde->head_lock); | |
2204 | ||
2205 | __sdma_process_event(sde, event); | |
2206 | ||
2207 | if (sde->state.current_state == sdma_state_s99_running) | |
2208 | sdma_desc_avail(sde, sdma_descq_freecnt(sde)); | |
2209 | ||
2210 | write_sequnlock(&sde->head_lock); | |
2211 | spin_unlock_irqrestore(&sde->tail_lock, flags); | |
2212 | } | |
2213 | ||
2214 | static void __sdma_process_event(struct sdma_engine *sde, | |
2215 | enum sdma_events event) | |
2216 | { | |
2217 | struct sdma_state *ss = &sde->state; | |
2218 | int need_progress = 0; | |
2219 | ||
2220 | /* CONFIG SDMA temporary */ | |
2221 | #ifdef CONFIG_SDMA_VERBOSITY | |
2222 | dd_dev_err(sde->dd, "CONFIG SDMA(%u) [%s] %s\n", sde->this_idx, | |
2223 | sdma_state_names[ss->current_state], | |
2224 | sdma_event_names[event]); | |
2225 | #endif | |
2226 | ||
2227 | switch (ss->current_state) { | |
2228 | case sdma_state_s00_hw_down: | |
2229 | switch (event) { | |
2230 | case sdma_event_e00_go_hw_down: | |
2231 | break; | |
2232 | case sdma_event_e30_go_running: | |
2233 | /* | |
2234 | * If down, but running requested (usually result | |
2235 | * of link up, then we need to start up. | |
2236 | * This can happen when hw down is requested while | |
2237 | * bringing the link up with traffic active on | |
2238 | * 7220, e.g. */ | |
2239 | ss->go_s99_running = 1; | |
2240 | /* fall through and start dma engine */ | |
2241 | case sdma_event_e10_go_hw_start: | |
2242 | /* This reference means the state machine is started */ | |
2243 | sdma_get(&sde->state); | |
2244 | sdma_set_state(sde, | |
2245 | sdma_state_s10_hw_start_up_halt_wait); | |
2246 | break; | |
2247 | case sdma_event_e15_hw_halt_done: | |
2248 | break; | |
2249 | case sdma_event_e25_hw_clean_up_done: | |
2250 | break; | |
2251 | case sdma_event_e40_sw_cleaned: | |
2252 | sdma_sw_tear_down(sde); | |
2253 | break; | |
2254 | case sdma_event_e50_hw_cleaned: | |
2255 | break; | |
2256 | case sdma_event_e60_hw_halted: | |
2257 | break; | |
2258 | case sdma_event_e70_go_idle: | |
2259 | break; | |
2260 | case sdma_event_e80_hw_freeze: | |
2261 | break; | |
2262 | case sdma_event_e81_hw_frozen: | |
2263 | break; | |
2264 | case sdma_event_e82_hw_unfreeze: | |
2265 | break; | |
2266 | case sdma_event_e85_link_down: | |
2267 | break; | |
2268 | case sdma_event_e90_sw_halted: | |
2269 | break; | |
2270 | } | |
2271 | break; | |
2272 | ||
2273 | case sdma_state_s10_hw_start_up_halt_wait: | |
2274 | switch (event) { | |
2275 | case sdma_event_e00_go_hw_down: | |
2276 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2277 | sdma_sw_tear_down(sde); | |
2278 | break; | |
2279 | case sdma_event_e10_go_hw_start: | |
2280 | break; | |
2281 | case sdma_event_e15_hw_halt_done: | |
2282 | sdma_set_state(sde, | |
2283 | sdma_state_s15_hw_start_up_clean_wait); | |
2284 | sdma_start_hw_clean_up(sde); | |
2285 | break; | |
2286 | case sdma_event_e25_hw_clean_up_done: | |
2287 | break; | |
2288 | case sdma_event_e30_go_running: | |
2289 | ss->go_s99_running = 1; | |
2290 | break; | |
2291 | case sdma_event_e40_sw_cleaned: | |
2292 | break; | |
2293 | case sdma_event_e50_hw_cleaned: | |
2294 | break; | |
2295 | case sdma_event_e60_hw_halted: | |
2296 | sdma_start_err_halt_wait(sde); | |
2297 | break; | |
2298 | case sdma_event_e70_go_idle: | |
2299 | ss->go_s99_running = 0; | |
2300 | break; | |
2301 | case sdma_event_e80_hw_freeze: | |
2302 | break; | |
2303 | case sdma_event_e81_hw_frozen: | |
2304 | break; | |
2305 | case sdma_event_e82_hw_unfreeze: | |
2306 | break; | |
2307 | case sdma_event_e85_link_down: | |
2308 | break; | |
2309 | case sdma_event_e90_sw_halted: | |
2310 | break; | |
2311 | } | |
2312 | break; | |
2313 | ||
2314 | case sdma_state_s15_hw_start_up_clean_wait: | |
2315 | switch (event) { | |
2316 | case sdma_event_e00_go_hw_down: | |
2317 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2318 | sdma_sw_tear_down(sde); | |
2319 | break; | |
2320 | case sdma_event_e10_go_hw_start: | |
2321 | break; | |
2322 | case sdma_event_e15_hw_halt_done: | |
2323 | break; | |
2324 | case sdma_event_e25_hw_clean_up_done: | |
2325 | sdma_hw_start_up(sde); | |
2326 | sdma_set_state(sde, ss->go_s99_running ? | |
2327 | sdma_state_s99_running : | |
2328 | sdma_state_s20_idle); | |
2329 | break; | |
2330 | case sdma_event_e30_go_running: | |
2331 | ss->go_s99_running = 1; | |
2332 | break; | |
2333 | case sdma_event_e40_sw_cleaned: | |
2334 | break; | |
2335 | case sdma_event_e50_hw_cleaned: | |
2336 | break; | |
2337 | case sdma_event_e60_hw_halted: | |
2338 | break; | |
2339 | case sdma_event_e70_go_idle: | |
2340 | ss->go_s99_running = 0; | |
2341 | break; | |
2342 | case sdma_event_e80_hw_freeze: | |
2343 | break; | |
2344 | case sdma_event_e81_hw_frozen: | |
2345 | break; | |
2346 | case sdma_event_e82_hw_unfreeze: | |
2347 | break; | |
2348 | case sdma_event_e85_link_down: | |
2349 | break; | |
2350 | case sdma_event_e90_sw_halted: | |
2351 | break; | |
2352 | } | |
2353 | break; | |
2354 | ||
2355 | case sdma_state_s20_idle: | |
2356 | switch (event) { | |
2357 | case sdma_event_e00_go_hw_down: | |
2358 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2359 | sdma_sw_tear_down(sde); | |
2360 | break; | |
2361 | case sdma_event_e10_go_hw_start: | |
2362 | break; | |
2363 | case sdma_event_e15_hw_halt_done: | |
2364 | break; | |
2365 | case sdma_event_e25_hw_clean_up_done: | |
2366 | break; | |
2367 | case sdma_event_e30_go_running: | |
2368 | sdma_set_state(sde, sdma_state_s99_running); | |
2369 | ss->go_s99_running = 1; | |
2370 | break; | |
2371 | case sdma_event_e40_sw_cleaned: | |
2372 | break; | |
2373 | case sdma_event_e50_hw_cleaned: | |
2374 | break; | |
2375 | case sdma_event_e60_hw_halted: | |
2376 | sdma_set_state(sde, sdma_state_s50_hw_halt_wait); | |
2377 | sdma_start_err_halt_wait(sde); | |
2378 | break; | |
2379 | case sdma_event_e70_go_idle: | |
2380 | break; | |
2381 | case sdma_event_e85_link_down: | |
2382 | /* fall through */ | |
2383 | case sdma_event_e80_hw_freeze: | |
2384 | sdma_set_state(sde, sdma_state_s80_hw_freeze); | |
2385 | atomic_dec(&sde->dd->sdma_unfreeze_count); | |
2386 | wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); | |
2387 | break; | |
2388 | case sdma_event_e81_hw_frozen: | |
2389 | break; | |
2390 | case sdma_event_e82_hw_unfreeze: | |
2391 | break; | |
2392 | case sdma_event_e90_sw_halted: | |
2393 | break; | |
2394 | } | |
2395 | break; | |
2396 | ||
2397 | case sdma_state_s30_sw_clean_up_wait: | |
2398 | switch (event) { | |
2399 | case sdma_event_e00_go_hw_down: | |
2400 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2401 | break; | |
2402 | case sdma_event_e10_go_hw_start: | |
2403 | break; | |
2404 | case sdma_event_e15_hw_halt_done: | |
2405 | break; | |
2406 | case sdma_event_e25_hw_clean_up_done: | |
2407 | break; | |
2408 | case sdma_event_e30_go_running: | |
2409 | ss->go_s99_running = 1; | |
2410 | break; | |
2411 | case sdma_event_e40_sw_cleaned: | |
2412 | sdma_set_state(sde, sdma_state_s40_hw_clean_up_wait); | |
2413 | sdma_start_hw_clean_up(sde); | |
2414 | break; | |
2415 | case sdma_event_e50_hw_cleaned: | |
2416 | break; | |
2417 | case sdma_event_e60_hw_halted: | |
2418 | break; | |
2419 | case sdma_event_e70_go_idle: | |
2420 | ss->go_s99_running = 0; | |
2421 | break; | |
2422 | case sdma_event_e80_hw_freeze: | |
2423 | break; | |
2424 | case sdma_event_e81_hw_frozen: | |
2425 | break; | |
2426 | case sdma_event_e82_hw_unfreeze: | |
2427 | break; | |
2428 | case sdma_event_e85_link_down: | |
2429 | ss->go_s99_running = 0; | |
2430 | break; | |
2431 | case sdma_event_e90_sw_halted: | |
2432 | break; | |
2433 | } | |
2434 | break; | |
2435 | ||
2436 | case sdma_state_s40_hw_clean_up_wait: | |
2437 | switch (event) { | |
2438 | case sdma_event_e00_go_hw_down: | |
2439 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2440 | sdma_start_sw_clean_up(sde); | |
2441 | break; | |
2442 | case sdma_event_e10_go_hw_start: | |
2443 | break; | |
2444 | case sdma_event_e15_hw_halt_done: | |
2445 | break; | |
2446 | case sdma_event_e25_hw_clean_up_done: | |
2447 | sdma_hw_start_up(sde); | |
2448 | sdma_set_state(sde, ss->go_s99_running ? | |
2449 | sdma_state_s99_running : | |
2450 | sdma_state_s20_idle); | |
2451 | break; | |
2452 | case sdma_event_e30_go_running: | |
2453 | ss->go_s99_running = 1; | |
2454 | break; | |
2455 | case sdma_event_e40_sw_cleaned: | |
2456 | break; | |
2457 | case sdma_event_e50_hw_cleaned: | |
2458 | break; | |
2459 | case sdma_event_e60_hw_halted: | |
2460 | break; | |
2461 | case sdma_event_e70_go_idle: | |
2462 | ss->go_s99_running = 0; | |
2463 | break; | |
2464 | case sdma_event_e80_hw_freeze: | |
2465 | break; | |
2466 | case sdma_event_e81_hw_frozen: | |
2467 | break; | |
2468 | case sdma_event_e82_hw_unfreeze: | |
2469 | break; | |
2470 | case sdma_event_e85_link_down: | |
2471 | ss->go_s99_running = 0; | |
2472 | break; | |
2473 | case sdma_event_e90_sw_halted: | |
2474 | break; | |
2475 | } | |
2476 | break; | |
2477 | ||
2478 | case sdma_state_s50_hw_halt_wait: | |
2479 | switch (event) { | |
2480 | case sdma_event_e00_go_hw_down: | |
2481 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2482 | sdma_start_sw_clean_up(sde); | |
2483 | break; | |
2484 | case sdma_event_e10_go_hw_start: | |
2485 | break; | |
2486 | case sdma_event_e15_hw_halt_done: | |
2487 | sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait); | |
2488 | sdma_start_sw_clean_up(sde); | |
2489 | break; | |
2490 | case sdma_event_e25_hw_clean_up_done: | |
2491 | break; | |
2492 | case sdma_event_e30_go_running: | |
2493 | ss->go_s99_running = 1; | |
2494 | break; | |
2495 | case sdma_event_e40_sw_cleaned: | |
2496 | break; | |
2497 | case sdma_event_e50_hw_cleaned: | |
2498 | break; | |
2499 | case sdma_event_e60_hw_halted: | |
2500 | sdma_start_err_halt_wait(sde); | |
2501 | break; | |
2502 | case sdma_event_e70_go_idle: | |
2503 | ss->go_s99_running = 0; | |
2504 | break; | |
2505 | case sdma_event_e80_hw_freeze: | |
2506 | break; | |
2507 | case sdma_event_e81_hw_frozen: | |
2508 | break; | |
2509 | case sdma_event_e82_hw_unfreeze: | |
2510 | break; | |
2511 | case sdma_event_e85_link_down: | |
2512 | ss->go_s99_running = 0; | |
2513 | break; | |
2514 | case sdma_event_e90_sw_halted: | |
2515 | break; | |
2516 | } | |
2517 | break; | |
2518 | ||
2519 | case sdma_state_s60_idle_halt_wait: | |
2520 | switch (event) { | |
2521 | case sdma_event_e00_go_hw_down: | |
2522 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2523 | sdma_start_sw_clean_up(sde); | |
2524 | break; | |
2525 | case sdma_event_e10_go_hw_start: | |
2526 | break; | |
2527 | case sdma_event_e15_hw_halt_done: | |
2528 | sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait); | |
2529 | sdma_start_sw_clean_up(sde); | |
2530 | break; | |
2531 | case sdma_event_e25_hw_clean_up_done: | |
2532 | break; | |
2533 | case sdma_event_e30_go_running: | |
2534 | ss->go_s99_running = 1; | |
2535 | break; | |
2536 | case sdma_event_e40_sw_cleaned: | |
2537 | break; | |
2538 | case sdma_event_e50_hw_cleaned: | |
2539 | break; | |
2540 | case sdma_event_e60_hw_halted: | |
2541 | sdma_start_err_halt_wait(sde); | |
2542 | break; | |
2543 | case sdma_event_e70_go_idle: | |
2544 | ss->go_s99_running = 0; | |
2545 | break; | |
2546 | case sdma_event_e80_hw_freeze: | |
2547 | break; | |
2548 | case sdma_event_e81_hw_frozen: | |
2549 | break; | |
2550 | case sdma_event_e82_hw_unfreeze: | |
2551 | break; | |
2552 | case sdma_event_e85_link_down: | |
2553 | break; | |
2554 | case sdma_event_e90_sw_halted: | |
2555 | break; | |
2556 | } | |
2557 | break; | |
2558 | ||
2559 | case sdma_state_s80_hw_freeze: | |
2560 | switch (event) { | |
2561 | case sdma_event_e00_go_hw_down: | |
2562 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2563 | sdma_start_sw_clean_up(sde); | |
2564 | break; | |
2565 | case sdma_event_e10_go_hw_start: | |
2566 | break; | |
2567 | case sdma_event_e15_hw_halt_done: | |
2568 | break; | |
2569 | case sdma_event_e25_hw_clean_up_done: | |
2570 | break; | |
2571 | case sdma_event_e30_go_running: | |
2572 | ss->go_s99_running = 1; | |
2573 | break; | |
2574 | case sdma_event_e40_sw_cleaned: | |
2575 | break; | |
2576 | case sdma_event_e50_hw_cleaned: | |
2577 | break; | |
2578 | case sdma_event_e60_hw_halted: | |
2579 | break; | |
2580 | case sdma_event_e70_go_idle: | |
2581 | ss->go_s99_running = 0; | |
2582 | break; | |
2583 | case sdma_event_e80_hw_freeze: | |
2584 | break; | |
2585 | case sdma_event_e81_hw_frozen: | |
2586 | sdma_set_state(sde, sdma_state_s82_freeze_sw_clean); | |
2587 | sdma_start_sw_clean_up(sde); | |
2588 | break; | |
2589 | case sdma_event_e82_hw_unfreeze: | |
2590 | break; | |
2591 | case sdma_event_e85_link_down: | |
2592 | break; | |
2593 | case sdma_event_e90_sw_halted: | |
2594 | break; | |
2595 | } | |
2596 | break; | |
2597 | ||
2598 | case sdma_state_s82_freeze_sw_clean: | |
2599 | switch (event) { | |
2600 | case sdma_event_e00_go_hw_down: | |
2601 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2602 | sdma_start_sw_clean_up(sde); | |
2603 | break; | |
2604 | case sdma_event_e10_go_hw_start: | |
2605 | break; | |
2606 | case sdma_event_e15_hw_halt_done: | |
2607 | break; | |
2608 | case sdma_event_e25_hw_clean_up_done: | |
2609 | break; | |
2610 | case sdma_event_e30_go_running: | |
2611 | ss->go_s99_running = 1; | |
2612 | break; | |
2613 | case sdma_event_e40_sw_cleaned: | |
2614 | /* notify caller this engine is done cleaning */ | |
2615 | atomic_dec(&sde->dd->sdma_unfreeze_count); | |
2616 | wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); | |
2617 | break; | |
2618 | case sdma_event_e50_hw_cleaned: | |
2619 | break; | |
2620 | case sdma_event_e60_hw_halted: | |
2621 | break; | |
2622 | case sdma_event_e70_go_idle: | |
2623 | ss->go_s99_running = 0; | |
2624 | break; | |
2625 | case sdma_event_e80_hw_freeze: | |
2626 | break; | |
2627 | case sdma_event_e81_hw_frozen: | |
2628 | break; | |
2629 | case sdma_event_e82_hw_unfreeze: | |
2630 | sdma_hw_start_up(sde); | |
2631 | sdma_set_state(sde, ss->go_s99_running ? | |
2632 | sdma_state_s99_running : | |
2633 | sdma_state_s20_idle); | |
2634 | break; | |
2635 | case sdma_event_e85_link_down: | |
2636 | break; | |
2637 | case sdma_event_e90_sw_halted: | |
2638 | break; | |
2639 | } | |
2640 | break; | |
2641 | ||
2642 | case sdma_state_s99_running: | |
2643 | switch (event) { | |
2644 | case sdma_event_e00_go_hw_down: | |
2645 | sdma_set_state(sde, sdma_state_s00_hw_down); | |
2646 | sdma_start_sw_clean_up(sde); | |
2647 | break; | |
2648 | case sdma_event_e10_go_hw_start: | |
2649 | break; | |
2650 | case sdma_event_e15_hw_halt_done: | |
2651 | break; | |
2652 | case sdma_event_e25_hw_clean_up_done: | |
2653 | break; | |
2654 | case sdma_event_e30_go_running: | |
2655 | break; | |
2656 | case sdma_event_e40_sw_cleaned: | |
2657 | break; | |
2658 | case sdma_event_e50_hw_cleaned: | |
2659 | break; | |
2660 | case sdma_event_e60_hw_halted: | |
2661 | need_progress = 1; | |
2662 | sdma_err_progress_check_schedule(sde); | |
2663 | case sdma_event_e90_sw_halted: | |
2664 | /* | |
2665 | * SW initiated halt does not perform engines | |
2666 | * progress check | |
2667 | */ | |
2668 | sdma_set_state(sde, sdma_state_s50_hw_halt_wait); | |
2669 | sdma_start_err_halt_wait(sde); | |
2670 | break; | |
2671 | case sdma_event_e70_go_idle: | |
2672 | sdma_set_state(sde, sdma_state_s60_idle_halt_wait); | |
2673 | break; | |
2674 | case sdma_event_e85_link_down: | |
2675 | ss->go_s99_running = 0; | |
2676 | /* fall through */ | |
2677 | case sdma_event_e80_hw_freeze: | |
2678 | sdma_set_state(sde, sdma_state_s80_hw_freeze); | |
2679 | atomic_dec(&sde->dd->sdma_unfreeze_count); | |
2680 | wake_up_interruptible(&sde->dd->sdma_unfreeze_wq); | |
2681 | break; | |
2682 | case sdma_event_e81_hw_frozen: | |
2683 | break; | |
2684 | case sdma_event_e82_hw_unfreeze: | |
2685 | break; | |
2686 | } | |
2687 | break; | |
2688 | } | |
2689 | ||
2690 | ss->last_event = event; | |
2691 | if (need_progress) | |
2692 | sdma_make_progress(sde, 0); | |
2693 | } | |
2694 | ||
2695 | /* | |
2696 | * _extend_sdma_tx_descs() - helper to extend txreq | |
2697 | * | |
2698 | * This is called once the initial nominal allocation | |
2699 | * of descriptors in the sdma_txreq is exhausted. | |
2700 | * | |
2701 | * The code will bump the allocation up to the max | |
2702 | * of MAX_DESC (64) descriptors. There doesn't seem | |
2703 | * much point in an interim step. | |
2704 | * | |
2705 | */ | |
2706 | int _extend_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx) | |
2707 | { | |
2708 | int i; | |
2709 | ||
2710 | tx->descp = kmalloc_array( | |
2711 | MAX_DESC, | |
2712 | sizeof(struct sdma_desc), | |
2713 | GFP_ATOMIC); | |
2714 | if (!tx->descp) | |
2715 | return -ENOMEM; | |
2716 | tx->desc_limit = MAX_DESC; | |
2717 | /* copy ones already built */ | |
2718 | for (i = 0; i < tx->num_desc; i++) | |
2719 | tx->descp[i] = tx->descs[i]; | |
2720 | return 0; | |
2721 | } | |
2722 | ||
2723 | /* Update sdes when the lmc changes */ | |
2724 | void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid) | |
2725 | { | |
2726 | struct sdma_engine *sde; | |
2727 | int i; | |
2728 | u64 sreg; | |
2729 | ||
2730 | sreg = ((mask & SD(CHECK_SLID_MASK_MASK)) << | |
2731 | SD(CHECK_SLID_MASK_SHIFT)) | | |
2732 | (((lid & mask) & SD(CHECK_SLID_VALUE_MASK)) << | |
2733 | SD(CHECK_SLID_VALUE_SHIFT)); | |
2734 | ||
2735 | for (i = 0; i < dd->num_sdma; i++) { | |
2736 | hfi1_cdbg(LINKVERB, "SendDmaEngine[%d].SLID_CHECK = 0x%x", | |
2737 | i, (u32)sreg); | |
2738 | sde = &dd->per_sdma[i]; | |
2739 | write_sde_csr(sde, SD(CHECK_SLID), sreg); | |
2740 | } | |
2741 | } | |
2742 | ||
2743 | /* tx not dword sized - pad */ | |
2744 | int _pad_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx) | |
2745 | { | |
2746 | int rval = 0; | |
2747 | ||
2748 | if ((unlikely(tx->num_desc == tx->desc_limit))) { | |
2749 | rval = _extend_sdma_tx_descs(dd, tx); | |
2750 | if (rval) | |
2751 | return rval; | |
2752 | } | |
2753 | /* finish the one just added */ | |
2754 | tx->num_desc++; | |
2755 | make_tx_sdma_desc( | |
2756 | tx, | |
2757 | SDMA_MAP_NONE, | |
2758 | dd->sdma_pad_phys, | |
2759 | sizeof(u32) - (tx->packet_len & (sizeof(u32) - 1))); | |
2760 | _sdma_close_tx(dd, tx); | |
2761 | return rval; | |
2762 | } | |
2763 | ||
2764 | /* | |
2765 | * Add ahg to the sdma_txreq | |
2766 | * | |
2767 | * The logic will consume up to 3 | |
2768 | * descriptors at the beginning of | |
2769 | * sdma_txreq. | |
2770 | */ | |
2771 | void _sdma_txreq_ahgadd( | |
2772 | struct sdma_txreq *tx, | |
2773 | u8 num_ahg, | |
2774 | u8 ahg_entry, | |
2775 | u32 *ahg, | |
2776 | u8 ahg_hlen) | |
2777 | { | |
2778 | u32 i, shift = 0, desc = 0; | |
2779 | u8 mode; | |
2780 | ||
2781 | WARN_ON_ONCE(num_ahg > 9 || (ahg_hlen & 3) || ahg_hlen == 4); | |
2782 | /* compute mode */ | |
2783 | if (num_ahg == 1) | |
2784 | mode = SDMA_AHG_APPLY_UPDATE1; | |
2785 | else if (num_ahg <= 5) | |
2786 | mode = SDMA_AHG_APPLY_UPDATE2; | |
2787 | else | |
2788 | mode = SDMA_AHG_APPLY_UPDATE3; | |
2789 | tx->num_desc++; | |
2790 | /* initialize to consumed descriptors to zero */ | |
2791 | switch (mode) { | |
2792 | case SDMA_AHG_APPLY_UPDATE3: | |
2793 | tx->num_desc++; | |
2794 | tx->descs[2].qw[0] = 0; | |
2795 | tx->descs[2].qw[1] = 0; | |
2796 | /* FALLTHROUGH */ | |
2797 | case SDMA_AHG_APPLY_UPDATE2: | |
2798 | tx->num_desc++; | |
2799 | tx->descs[1].qw[0] = 0; | |
2800 | tx->descs[1].qw[1] = 0; | |
2801 | break; | |
2802 | } | |
2803 | ahg_hlen >>= 2; | |
2804 | tx->descs[0].qw[1] |= | |
2805 | (((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK) | |
2806 | << SDMA_DESC1_HEADER_INDEX_SHIFT) | | |
2807 | (((u64)ahg_hlen & SDMA_DESC1_HEADER_DWS_MASK) | |
2808 | << SDMA_DESC1_HEADER_DWS_SHIFT) | | |
2809 | (((u64)mode & SDMA_DESC1_HEADER_MODE_MASK) | |
2810 | << SDMA_DESC1_HEADER_MODE_SHIFT) | | |
2811 | (((u64)ahg[0] & SDMA_DESC1_HEADER_UPDATE1_MASK) | |
2812 | << SDMA_DESC1_HEADER_UPDATE1_SHIFT); | |
2813 | for (i = 0; i < (num_ahg - 1); i++) { | |
2814 | if (!shift && !(i & 2)) | |
2815 | desc++; | |
2816 | tx->descs[desc].qw[!!(i & 2)] |= | |
2817 | (((u64)ahg[i + 1]) | |
2818 | << shift); | |
2819 | shift = (shift + 32) & 63; | |
2820 | } | |
2821 | } | |
2822 | ||
2823 | /** | |
2824 | * sdma_ahg_alloc - allocate an AHG entry | |
2825 | * @sde: engine to allocate from | |
2826 | * | |
2827 | * Return: | |
2828 | * 0-31 when successful, -EOPNOTSUPP if AHG is not enabled, | |
2829 | * -ENOSPC if an entry is not available | |
2830 | */ | |
2831 | int sdma_ahg_alloc(struct sdma_engine *sde) | |
2832 | { | |
2833 | int nr; | |
2834 | int oldbit; | |
2835 | ||
2836 | if (!sde) { | |
2837 | trace_hfi1_ahg_allocate(sde, -EINVAL); | |
2838 | return -EINVAL; | |
2839 | } | |
2840 | while (1) { | |
2841 | nr = ffz(ACCESS_ONCE(sde->ahg_bits)); | |
2842 | if (nr > 31) { | |
2843 | trace_hfi1_ahg_allocate(sde, -ENOSPC); | |
2844 | return -ENOSPC; | |
2845 | } | |
2846 | oldbit = test_and_set_bit(nr, &sde->ahg_bits); | |
2847 | if (!oldbit) | |
2848 | break; | |
2849 | cpu_relax(); | |
2850 | } | |
2851 | trace_hfi1_ahg_allocate(sde, nr); | |
2852 | return nr; | |
2853 | } | |
2854 | ||
2855 | /** | |
2856 | * sdma_ahg_free - free an AHG entry | |
2857 | * @sde: engine to return AHG entry | |
2858 | * @ahg_index: index to free | |
2859 | * | |
2860 | * This routine frees the indicate AHG entry. | |
2861 | */ | |
2862 | void sdma_ahg_free(struct sdma_engine *sde, int ahg_index) | |
2863 | { | |
2864 | if (!sde) | |
2865 | return; | |
2866 | trace_hfi1_ahg_deallocate(sde, ahg_index); | |
2867 | if (ahg_index < 0 || ahg_index > 31) | |
2868 | return; | |
2869 | clear_bit(ahg_index, &sde->ahg_bits); | |
2870 | } | |
2871 | ||
2872 | /* | |
2873 | * SPC freeze handling for SDMA engines. Called when the driver knows | |
2874 | * the SPC is going into a freeze but before the freeze is fully | |
2875 | * settled. Generally an error interrupt. | |
2876 | * | |
2877 | * This event will pull the engine out of running so no more entries can be | |
2878 | * added to the engine's queue. | |
2879 | */ | |
2880 | void sdma_freeze_notify(struct hfi1_devdata *dd, int link_down) | |
2881 | { | |
2882 | int i; | |
2883 | enum sdma_events event = link_down ? sdma_event_e85_link_down : | |
2884 | sdma_event_e80_hw_freeze; | |
2885 | ||
2886 | /* set up the wait but do not wait here */ | |
2887 | atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma); | |
2888 | ||
2889 | /* tell all engines to stop running and wait */ | |
2890 | for (i = 0; i < dd->num_sdma; i++) | |
2891 | sdma_process_event(&dd->per_sdma[i], event); | |
2892 | ||
2893 | /* sdma_freeze() will wait for all engines to have stopped */ | |
2894 | } | |
2895 | ||
2896 | /* | |
2897 | * SPC freeze handling for SDMA engines. Called when the driver knows | |
2898 | * the SPC is fully frozen. | |
2899 | */ | |
2900 | void sdma_freeze(struct hfi1_devdata *dd) | |
2901 | { | |
2902 | int i; | |
2903 | int ret; | |
2904 | ||
2905 | /* | |
2906 | * Make sure all engines have moved out of the running state before | |
2907 | * continuing. | |
2908 | */ | |
2909 | ret = wait_event_interruptible(dd->sdma_unfreeze_wq, | |
2910 | atomic_read(&dd->sdma_unfreeze_count) <= 0); | |
2911 | /* interrupted or count is negative, then unloading - just exit */ | |
2912 | if (ret || atomic_read(&dd->sdma_unfreeze_count) < 0) | |
2913 | return; | |
2914 | ||
2915 | /* set up the count for the next wait */ | |
2916 | atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma); | |
2917 | ||
2918 | /* tell all engines that the SPC is frozen, they can start cleaning */ | |
2919 | for (i = 0; i < dd->num_sdma; i++) | |
2920 | sdma_process_event(&dd->per_sdma[i], sdma_event_e81_hw_frozen); | |
2921 | ||
2922 | /* | |
2923 | * Wait for everyone to finish software clean before exiting. The | |
2924 | * software clean will read engine CSRs, so must be completed before | |
2925 | * the next step, which will clear the engine CSRs. | |
2926 | */ | |
2927 | (void) wait_event_interruptible(dd->sdma_unfreeze_wq, | |
2928 | atomic_read(&dd->sdma_unfreeze_count) <= 0); | |
2929 | /* no need to check results - done no matter what */ | |
2930 | } | |
2931 | ||
2932 | /* | |
2933 | * SPC freeze handling for the SDMA engines. Called after the SPC is unfrozen. | |
2934 | * | |
2935 | * The SPC freeze acts like a SDMA halt and a hardware clean combined. All | |
2936 | * that is left is a software clean. We could do it after the SPC is fully | |
2937 | * frozen, but then we'd have to add another state to wait for the unfreeze. | |
2938 | * Instead, just defer the software clean until the unfreeze step. | |
2939 | */ | |
2940 | void sdma_unfreeze(struct hfi1_devdata *dd) | |
2941 | { | |
2942 | int i; | |
2943 | ||
2944 | /* tell all engines start freeze clean up */ | |
2945 | for (i = 0; i < dd->num_sdma; i++) | |
2946 | sdma_process_event(&dd->per_sdma[i], | |
2947 | sdma_event_e82_hw_unfreeze); | |
2948 | } | |
2949 | ||
2950 | /** | |
2951 | * _sdma_engine_progress_schedule() - schedule progress on engine | |
2952 | * @sde: sdma_engine to schedule progress | |
2953 | * | |
2954 | */ | |
2955 | void _sdma_engine_progress_schedule( | |
2956 | struct sdma_engine *sde) | |
2957 | { | |
2958 | trace_hfi1_sdma_engine_progress(sde, sde->progress_mask); | |
2959 | /* assume we have selected a good cpu */ | |
2960 | write_csr(sde->dd, | |
2961 | CCE_INT_FORCE + (8*(IS_SDMA_START/64)), sde->progress_mask); | |
2962 | } |