xref: /openbmc/linux/drivers/infiniband/hw/hfi1/sdma.c (revision 4f6cce39)
1 /*
2  * Copyright(c) 2015, 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/spinlock.h>
49 #include <linux/seqlock.h>
50 #include <linux/netdevice.h>
51 #include <linux/moduleparam.h>
52 #include <linux/bitops.h>
53 #include <linux/timer.h>
54 #include <linux/vmalloc.h>
55 #include <linux/highmem.h>
56 
57 #include "hfi.h"
58 #include "common.h"
59 #include "qp.h"
60 #include "sdma.h"
61 #include "iowait.h"
62 #include "trace.h"
63 
64 /* must be a power of 2 >= 64 <= 32768 */
65 #define SDMA_DESCQ_CNT 2048
66 #define SDMA_DESC_INTR 64
67 #define INVALID_TAIL 0xffff
68 
69 static uint sdma_descq_cnt = SDMA_DESCQ_CNT;
70 module_param(sdma_descq_cnt, uint, S_IRUGO);
71 MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries");
72 
73 static uint sdma_idle_cnt = 250;
74 module_param(sdma_idle_cnt, uint, S_IRUGO);
75 MODULE_PARM_DESC(sdma_idle_cnt, "sdma interrupt idle delay (ns,default 250)");
76 
77 uint mod_num_sdma;
78 module_param_named(num_sdma, mod_num_sdma, uint, S_IRUGO);
79 MODULE_PARM_DESC(num_sdma, "Set max number SDMA engines to use");
80 
81 static uint sdma_desct_intr = SDMA_DESC_INTR;
82 module_param_named(desct_intr, sdma_desct_intr, uint, S_IRUGO | S_IWUSR);
83 MODULE_PARM_DESC(desct_intr, "Number of SDMA descriptor before interrupt");
84 
85 #define SDMA_WAIT_BATCH_SIZE 20
86 /* max wait time for a SDMA engine to indicate it has halted */
87 #define SDMA_ERR_HALT_TIMEOUT 10 /* ms */
88 /* all SDMA engine errors that cause a halt */
89 
90 #define SD(name) SEND_DMA_##name
91 #define ALL_SDMA_ENG_HALT_ERRS \
92 	(SD(ENG_ERR_STATUS_SDMA_WRONG_DW_ERR_SMASK) \
93 	| SD(ENG_ERR_STATUS_SDMA_GEN_MISMATCH_ERR_SMASK) \
94 	| SD(ENG_ERR_STATUS_SDMA_TOO_LONG_ERR_SMASK) \
95 	| SD(ENG_ERR_STATUS_SDMA_TAIL_OUT_OF_BOUNDS_ERR_SMASK) \
96 	| SD(ENG_ERR_STATUS_SDMA_FIRST_DESC_ERR_SMASK) \
97 	| SD(ENG_ERR_STATUS_SDMA_MEM_READ_ERR_SMASK) \
98 	| SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK) \
99 	| SD(ENG_ERR_STATUS_SDMA_LENGTH_MISMATCH_ERR_SMASK) \
100 	| SD(ENG_ERR_STATUS_SDMA_PACKET_DESC_OVERFLOW_ERR_SMASK) \
101 	| SD(ENG_ERR_STATUS_SDMA_HEADER_SELECT_ERR_SMASK) \
102 	| SD(ENG_ERR_STATUS_SDMA_HEADER_ADDRESS_ERR_SMASK) \
103 	| SD(ENG_ERR_STATUS_SDMA_HEADER_LENGTH_ERR_SMASK) \
104 	| SD(ENG_ERR_STATUS_SDMA_TIMEOUT_ERR_SMASK) \
105 	| SD(ENG_ERR_STATUS_SDMA_DESC_TABLE_UNC_ERR_SMASK) \
106 	| SD(ENG_ERR_STATUS_SDMA_ASSEMBLY_UNC_ERR_SMASK) \
107 	| SD(ENG_ERR_STATUS_SDMA_PACKET_TRACKING_UNC_ERR_SMASK) \
108 	| SD(ENG_ERR_STATUS_SDMA_HEADER_STORAGE_UNC_ERR_SMASK) \
109 	| SD(ENG_ERR_STATUS_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SMASK))
110 
111 /* sdma_sendctrl operations */
112 #define SDMA_SENDCTRL_OP_ENABLE    BIT(0)
113 #define SDMA_SENDCTRL_OP_INTENABLE BIT(1)
114 #define SDMA_SENDCTRL_OP_HALT      BIT(2)
115 #define SDMA_SENDCTRL_OP_CLEANUP   BIT(3)
116 
117 /* handle long defines */
118 #define SDMA_EGRESS_PACKET_OCCUPANCY_SMASK \
119 SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SMASK
120 #define SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT \
121 SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT
122 
123 static const char * const sdma_state_names[] = {
124 	[sdma_state_s00_hw_down]                = "s00_HwDown",
125 	[sdma_state_s10_hw_start_up_halt_wait]  = "s10_HwStartUpHaltWait",
126 	[sdma_state_s15_hw_start_up_clean_wait] = "s15_HwStartUpCleanWait",
127 	[sdma_state_s20_idle]                   = "s20_Idle",
128 	[sdma_state_s30_sw_clean_up_wait]       = "s30_SwCleanUpWait",
129 	[sdma_state_s40_hw_clean_up_wait]       = "s40_HwCleanUpWait",
130 	[sdma_state_s50_hw_halt_wait]           = "s50_HwHaltWait",
131 	[sdma_state_s60_idle_halt_wait]         = "s60_IdleHaltWait",
132 	[sdma_state_s80_hw_freeze]		= "s80_HwFreeze",
133 	[sdma_state_s82_freeze_sw_clean]	= "s82_FreezeSwClean",
134 	[sdma_state_s99_running]                = "s99_Running",
135 };
136 
137 #ifdef CONFIG_SDMA_VERBOSITY
138 static const char * const sdma_event_names[] = {
139 	[sdma_event_e00_go_hw_down]   = "e00_GoHwDown",
140 	[sdma_event_e10_go_hw_start]  = "e10_GoHwStart",
141 	[sdma_event_e15_hw_halt_done] = "e15_HwHaltDone",
142 	[sdma_event_e25_hw_clean_up_done] = "e25_HwCleanUpDone",
143 	[sdma_event_e30_go_running]   = "e30_GoRunning",
144 	[sdma_event_e40_sw_cleaned]   = "e40_SwCleaned",
145 	[sdma_event_e50_hw_cleaned]   = "e50_HwCleaned",
146 	[sdma_event_e60_hw_halted]    = "e60_HwHalted",
147 	[sdma_event_e70_go_idle]      = "e70_GoIdle",
148 	[sdma_event_e80_hw_freeze]    = "e80_HwFreeze",
149 	[sdma_event_e81_hw_frozen]    = "e81_HwFrozen",
150 	[sdma_event_e82_hw_unfreeze]  = "e82_HwUnfreeze",
151 	[sdma_event_e85_link_down]    = "e85_LinkDown",
152 	[sdma_event_e90_sw_halted]    = "e90_SwHalted",
153 };
154 #endif
155 
156 static const struct sdma_set_state_action sdma_action_table[] = {
157 	[sdma_state_s00_hw_down] = {
158 		.go_s99_running_tofalse = 1,
159 		.op_enable = 0,
160 		.op_intenable = 0,
161 		.op_halt = 0,
162 		.op_cleanup = 0,
163 	},
164 	[sdma_state_s10_hw_start_up_halt_wait] = {
165 		.op_enable = 0,
166 		.op_intenable = 0,
167 		.op_halt = 1,
168 		.op_cleanup = 0,
169 	},
170 	[sdma_state_s15_hw_start_up_clean_wait] = {
171 		.op_enable = 0,
172 		.op_intenable = 1,
173 		.op_halt = 0,
174 		.op_cleanup = 1,
175 	},
176 	[sdma_state_s20_idle] = {
177 		.op_enable = 0,
178 		.op_intenable = 1,
179 		.op_halt = 0,
180 		.op_cleanup = 0,
181 	},
182 	[sdma_state_s30_sw_clean_up_wait] = {
183 		.op_enable = 0,
184 		.op_intenable = 0,
185 		.op_halt = 0,
186 		.op_cleanup = 0,
187 	},
188 	[sdma_state_s40_hw_clean_up_wait] = {
189 		.op_enable = 0,
190 		.op_intenable = 0,
191 		.op_halt = 0,
192 		.op_cleanup = 1,
193 	},
194 	[sdma_state_s50_hw_halt_wait] = {
195 		.op_enable = 0,
196 		.op_intenable = 0,
197 		.op_halt = 0,
198 		.op_cleanup = 0,
199 	},
200 	[sdma_state_s60_idle_halt_wait] = {
201 		.go_s99_running_tofalse = 1,
202 		.op_enable = 0,
203 		.op_intenable = 0,
204 		.op_halt = 1,
205 		.op_cleanup = 0,
206 	},
207 	[sdma_state_s80_hw_freeze] = {
208 		.op_enable = 0,
209 		.op_intenable = 0,
210 		.op_halt = 0,
211 		.op_cleanup = 0,
212 	},
213 	[sdma_state_s82_freeze_sw_clean] = {
214 		.op_enable = 0,
215 		.op_intenable = 0,
216 		.op_halt = 0,
217 		.op_cleanup = 0,
218 	},
219 	[sdma_state_s99_running] = {
220 		.op_enable = 1,
221 		.op_intenable = 1,
222 		.op_halt = 0,
223 		.op_cleanup = 0,
224 		.go_s99_running_totrue = 1,
225 	},
226 };
227 
228 #define SDMA_TAIL_UPDATE_THRESH 0x1F
229 
230 /* declare all statics here rather than keep sorting */
231 static void sdma_complete(struct kref *);
232 static void sdma_finalput(struct sdma_state *);
233 static void sdma_get(struct sdma_state *);
234 static void sdma_hw_clean_up_task(unsigned long);
235 static void sdma_put(struct sdma_state *);
236 static void sdma_set_state(struct sdma_engine *, enum sdma_states);
237 static void sdma_start_hw_clean_up(struct sdma_engine *);
238 static void sdma_sw_clean_up_task(unsigned long);
239 static void sdma_sendctrl(struct sdma_engine *, unsigned);
240 static void init_sdma_regs(struct sdma_engine *, u32, uint);
241 static void sdma_process_event(
242 	struct sdma_engine *sde,
243 	enum sdma_events event);
244 static void __sdma_process_event(
245 	struct sdma_engine *sde,
246 	enum sdma_events event);
247 static void dump_sdma_state(struct sdma_engine *sde);
248 static void sdma_make_progress(struct sdma_engine *sde, u64 status);
249 static void sdma_desc_avail(struct sdma_engine *sde, unsigned avail);
250 static void sdma_flush_descq(struct sdma_engine *sde);
251 
252 /**
253  * sdma_state_name() - return state string from enum
254  * @state: state
255  */
256 static const char *sdma_state_name(enum sdma_states state)
257 {
258 	return sdma_state_names[state];
259 }
260 
261 static void sdma_get(struct sdma_state *ss)
262 {
263 	kref_get(&ss->kref);
264 }
265 
266 static void sdma_complete(struct kref *kref)
267 {
268 	struct sdma_state *ss =
269 		container_of(kref, struct sdma_state, kref);
270 
271 	complete(&ss->comp);
272 }
273 
274 static void sdma_put(struct sdma_state *ss)
275 {
276 	kref_put(&ss->kref, sdma_complete);
277 }
278 
279 static void sdma_finalput(struct sdma_state *ss)
280 {
281 	sdma_put(ss);
282 	wait_for_completion(&ss->comp);
283 }
284 
285 static inline void write_sde_csr(
286 	struct sdma_engine *sde,
287 	u32 offset0,
288 	u64 value)
289 {
290 	write_kctxt_csr(sde->dd, sde->this_idx, offset0, value);
291 }
292 
293 static inline u64 read_sde_csr(
294 	struct sdma_engine *sde,
295 	u32 offset0)
296 {
297 	return read_kctxt_csr(sde->dd, sde->this_idx, offset0);
298 }
299 
300 /*
301  * sdma_wait_for_packet_egress() - wait for the VL FIFO occupancy for
302  * sdma engine 'sde' to drop to 0.
303  */
304 static void sdma_wait_for_packet_egress(struct sdma_engine *sde,
305 					int pause)
306 {
307 	u64 off = 8 * sde->this_idx;
308 	struct hfi1_devdata *dd = sde->dd;
309 	int lcnt = 0;
310 	u64 reg_prev;
311 	u64 reg = 0;
312 
313 	while (1) {
314 		reg_prev = reg;
315 		reg = read_csr(dd, off + SEND_EGRESS_SEND_DMA_STATUS);
316 
317 		reg &= SDMA_EGRESS_PACKET_OCCUPANCY_SMASK;
318 		reg >>= SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT;
319 		if (reg == 0)
320 			break;
321 		/* counter is reest if accupancy count changes */
322 		if (reg != reg_prev)
323 			lcnt = 0;
324 		if (lcnt++ > 500) {
325 			/* timed out - bounce the link */
326 			dd_dev_err(dd, "%s: engine %u timeout waiting for packets to egress, remaining count %u, bouncing link\n",
327 				   __func__, sde->this_idx, (u32)reg);
328 			queue_work(dd->pport->hfi1_wq,
329 				   &dd->pport->link_bounce_work);
330 			break;
331 		}
332 		udelay(1);
333 	}
334 }
335 
336 /*
337  * sdma_wait() - wait for packet egress to complete for all SDMA engines,
338  * and pause for credit return.
339  */
340 void sdma_wait(struct hfi1_devdata *dd)
341 {
342 	int i;
343 
344 	for (i = 0; i < dd->num_sdma; i++) {
345 		struct sdma_engine *sde = &dd->per_sdma[i];
346 
347 		sdma_wait_for_packet_egress(sde, 0);
348 	}
349 }
350 
351 static inline void sdma_set_desc_cnt(struct sdma_engine *sde, unsigned cnt)
352 {
353 	u64 reg;
354 
355 	if (!(sde->dd->flags & HFI1_HAS_SDMA_TIMEOUT))
356 		return;
357 	reg = cnt;
358 	reg &= SD(DESC_CNT_CNT_MASK);
359 	reg <<= SD(DESC_CNT_CNT_SHIFT);
360 	write_sde_csr(sde, SD(DESC_CNT), reg);
361 }
362 
363 static inline void complete_tx(struct sdma_engine *sde,
364 			       struct sdma_txreq *tx,
365 			       int res)
366 {
367 	/* protect against complete modifying */
368 	struct iowait *wait = tx->wait;
369 	callback_t complete = tx->complete;
370 
371 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
372 	trace_hfi1_sdma_out_sn(sde, tx->sn);
373 	if (WARN_ON_ONCE(sde->head_sn != tx->sn))
374 		dd_dev_err(sde->dd, "expected %llu got %llu\n",
375 			   sde->head_sn, tx->sn);
376 	sde->head_sn++;
377 #endif
378 	__sdma_txclean(sde->dd, tx);
379 	if (complete)
380 		(*complete)(tx, res);
381 	if (wait && iowait_sdma_dec(wait))
382 		iowait_drain_wakeup(wait);
383 }
384 
385 /*
386  * Complete all the sdma requests with a SDMA_TXREQ_S_ABORTED status
387  *
388  * Depending on timing there can be txreqs in two places:
389  * - in the descq ring
390  * - in the flush list
391  *
392  * To avoid ordering issues the descq ring needs to be flushed
393  * first followed by the flush list.
394  *
395  * This routine is called from two places
396  * - From a work queue item
397  * - Directly from the state machine just before setting the
398  *   state to running
399  *
400  * Must be called with head_lock held
401  *
402  */
403 static void sdma_flush(struct sdma_engine *sde)
404 {
405 	struct sdma_txreq *txp, *txp_next;
406 	LIST_HEAD(flushlist);
407 	unsigned long flags;
408 
409 	/* flush from head to tail */
410 	sdma_flush_descq(sde);
411 	spin_lock_irqsave(&sde->flushlist_lock, flags);
412 	/* copy flush list */
413 	list_for_each_entry_safe(txp, txp_next, &sde->flushlist, list) {
414 		list_del_init(&txp->list);
415 		list_add_tail(&txp->list, &flushlist);
416 	}
417 	spin_unlock_irqrestore(&sde->flushlist_lock, flags);
418 	/* flush from flush list */
419 	list_for_each_entry_safe(txp, txp_next, &flushlist, list)
420 		complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED);
421 }
422 
423 /*
424  * Fields a work request for flushing the descq ring
425  * and the flush list
426  *
427  * If the engine has been brought to running during
428  * the scheduling delay, the flush is ignored, assuming
429  * that the process of bringing the engine to running
430  * would have done this flush prior to going to running.
431  *
432  */
433 static void sdma_field_flush(struct work_struct *work)
434 {
435 	unsigned long flags;
436 	struct sdma_engine *sde =
437 		container_of(work, struct sdma_engine, flush_worker);
438 
439 	write_seqlock_irqsave(&sde->head_lock, flags);
440 	if (!__sdma_running(sde))
441 		sdma_flush(sde);
442 	write_sequnlock_irqrestore(&sde->head_lock, flags);
443 }
444 
445 static void sdma_err_halt_wait(struct work_struct *work)
446 {
447 	struct sdma_engine *sde = container_of(work, struct sdma_engine,
448 						err_halt_worker);
449 	u64 statuscsr;
450 	unsigned long timeout;
451 
452 	timeout = jiffies + msecs_to_jiffies(SDMA_ERR_HALT_TIMEOUT);
453 	while (1) {
454 		statuscsr = read_sde_csr(sde, SD(STATUS));
455 		statuscsr &= SD(STATUS_ENG_HALTED_SMASK);
456 		if (statuscsr)
457 			break;
458 		if (time_after(jiffies, timeout)) {
459 			dd_dev_err(sde->dd,
460 				   "SDMA engine %d - timeout waiting for engine to halt\n",
461 				   sde->this_idx);
462 			/*
463 			 * Continue anyway.  This could happen if there was
464 			 * an uncorrectable error in the wrong spot.
465 			 */
466 			break;
467 		}
468 		usleep_range(80, 120);
469 	}
470 
471 	sdma_process_event(sde, sdma_event_e15_hw_halt_done);
472 }
473 
474 static void sdma_err_progress_check_schedule(struct sdma_engine *sde)
475 {
476 	if (!is_bx(sde->dd) && HFI1_CAP_IS_KSET(SDMA_AHG)) {
477 		unsigned index;
478 		struct hfi1_devdata *dd = sde->dd;
479 
480 		for (index = 0; index < dd->num_sdma; index++) {
481 			struct sdma_engine *curr_sdma = &dd->per_sdma[index];
482 
483 			if (curr_sdma != sde)
484 				curr_sdma->progress_check_head =
485 							curr_sdma->descq_head;
486 		}
487 		dd_dev_err(sde->dd,
488 			   "SDMA engine %d - check scheduled\n",
489 				sde->this_idx);
490 		mod_timer(&sde->err_progress_check_timer, jiffies + 10);
491 	}
492 }
493 
494 static void sdma_err_progress_check(unsigned long data)
495 {
496 	unsigned index;
497 	struct sdma_engine *sde = (struct sdma_engine *)data;
498 
499 	dd_dev_err(sde->dd, "SDE progress check event\n");
500 	for (index = 0; index < sde->dd->num_sdma; index++) {
501 		struct sdma_engine *curr_sde = &sde->dd->per_sdma[index];
502 		unsigned long flags;
503 
504 		/* check progress on each engine except the current one */
505 		if (curr_sde == sde)
506 			continue;
507 		/*
508 		 * We must lock interrupts when acquiring sde->lock,
509 		 * to avoid a deadlock if interrupt triggers and spins on
510 		 * the same lock on same CPU
511 		 */
512 		spin_lock_irqsave(&curr_sde->tail_lock, flags);
513 		write_seqlock(&curr_sde->head_lock);
514 
515 		/* skip non-running queues */
516 		if (curr_sde->state.current_state != sdma_state_s99_running) {
517 			write_sequnlock(&curr_sde->head_lock);
518 			spin_unlock_irqrestore(&curr_sde->tail_lock, flags);
519 			continue;
520 		}
521 
522 		if ((curr_sde->descq_head != curr_sde->descq_tail) &&
523 		    (curr_sde->descq_head ==
524 				curr_sde->progress_check_head))
525 			__sdma_process_event(curr_sde,
526 					     sdma_event_e90_sw_halted);
527 		write_sequnlock(&curr_sde->head_lock);
528 		spin_unlock_irqrestore(&curr_sde->tail_lock, flags);
529 	}
530 	schedule_work(&sde->err_halt_worker);
531 }
532 
533 static void sdma_hw_clean_up_task(unsigned long opaque)
534 {
535 	struct sdma_engine *sde = (struct sdma_engine *)opaque;
536 	u64 statuscsr;
537 
538 	while (1) {
539 #ifdef CONFIG_SDMA_VERBOSITY
540 		dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
541 			   sde->this_idx, slashstrip(__FILE__), __LINE__,
542 			__func__);
543 #endif
544 		statuscsr = read_sde_csr(sde, SD(STATUS));
545 		statuscsr &= SD(STATUS_ENG_CLEANED_UP_SMASK);
546 		if (statuscsr)
547 			break;
548 		udelay(10);
549 	}
550 
551 	sdma_process_event(sde, sdma_event_e25_hw_clean_up_done);
552 }
553 
554 static inline struct sdma_txreq *get_txhead(struct sdma_engine *sde)
555 {
556 	smp_read_barrier_depends(); /* see sdma_update_tail() */
557 	return sde->tx_ring[sde->tx_head & sde->sdma_mask];
558 }
559 
560 /*
561  * flush ring for recovery
562  */
563 static void sdma_flush_descq(struct sdma_engine *sde)
564 {
565 	u16 head, tail;
566 	int progress = 0;
567 	struct sdma_txreq *txp = get_txhead(sde);
568 
569 	/* The reason for some of the complexity of this code is that
570 	 * not all descriptors have corresponding txps.  So, we have to
571 	 * be able to skip over descs until we wander into the range of
572 	 * the next txp on the list.
573 	 */
574 	head = sde->descq_head & sde->sdma_mask;
575 	tail = sde->descq_tail & sde->sdma_mask;
576 	while (head != tail) {
577 		/* advance head, wrap if needed */
578 		head = ++sde->descq_head & sde->sdma_mask;
579 		/* if now past this txp's descs, do the callback */
580 		if (txp && txp->next_descq_idx == head) {
581 			/* remove from list */
582 			sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
583 			complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED);
584 			trace_hfi1_sdma_progress(sde, head, tail, txp);
585 			txp = get_txhead(sde);
586 		}
587 		progress++;
588 	}
589 	if (progress)
590 		sdma_desc_avail(sde, sdma_descq_freecnt(sde));
591 }
592 
593 static void sdma_sw_clean_up_task(unsigned long opaque)
594 {
595 	struct sdma_engine *sde = (struct sdma_engine *)opaque;
596 	unsigned long flags;
597 
598 	spin_lock_irqsave(&sde->tail_lock, flags);
599 	write_seqlock(&sde->head_lock);
600 
601 	/*
602 	 * At this point, the following should always be true:
603 	 * - We are halted, so no more descriptors are getting retired.
604 	 * - We are not running, so no one is submitting new work.
605 	 * - Only we can send the e40_sw_cleaned, so we can't start
606 	 *   running again until we say so.  So, the active list and
607 	 *   descq are ours to play with.
608 	 */
609 
610 	/*
611 	 * In the error clean up sequence, software clean must be called
612 	 * before the hardware clean so we can use the hardware head in
613 	 * the progress routine.  A hardware clean or SPC unfreeze will
614 	 * reset the hardware head.
615 	 *
616 	 * Process all retired requests. The progress routine will use the
617 	 * latest physical hardware head - we are not running so speed does
618 	 * not matter.
619 	 */
620 	sdma_make_progress(sde, 0);
621 
622 	sdma_flush(sde);
623 
624 	/*
625 	 * Reset our notion of head and tail.
626 	 * Note that the HW registers have been reset via an earlier
627 	 * clean up.
628 	 */
629 	sde->descq_tail = 0;
630 	sde->descq_head = 0;
631 	sde->desc_avail = sdma_descq_freecnt(sde);
632 	*sde->head_dma = 0;
633 
634 	__sdma_process_event(sde, sdma_event_e40_sw_cleaned);
635 
636 	write_sequnlock(&sde->head_lock);
637 	spin_unlock_irqrestore(&sde->tail_lock, flags);
638 }
639 
640 static void sdma_sw_tear_down(struct sdma_engine *sde)
641 {
642 	struct sdma_state *ss = &sde->state;
643 
644 	/* Releasing this reference means the state machine has stopped. */
645 	sdma_put(ss);
646 
647 	/* stop waiting for all unfreeze events to complete */
648 	atomic_set(&sde->dd->sdma_unfreeze_count, -1);
649 	wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
650 }
651 
652 static void sdma_start_hw_clean_up(struct sdma_engine *sde)
653 {
654 	tasklet_hi_schedule(&sde->sdma_hw_clean_up_task);
655 }
656 
657 static void sdma_set_state(struct sdma_engine *sde,
658 			   enum sdma_states next_state)
659 {
660 	struct sdma_state *ss = &sde->state;
661 	const struct sdma_set_state_action *action = sdma_action_table;
662 	unsigned op = 0;
663 
664 	trace_hfi1_sdma_state(
665 		sde,
666 		sdma_state_names[ss->current_state],
667 		sdma_state_names[next_state]);
668 
669 	/* debugging bookkeeping */
670 	ss->previous_state = ss->current_state;
671 	ss->previous_op = ss->current_op;
672 	ss->current_state = next_state;
673 
674 	if (ss->previous_state != sdma_state_s99_running &&
675 	    next_state == sdma_state_s99_running)
676 		sdma_flush(sde);
677 
678 	if (action[next_state].op_enable)
679 		op |= SDMA_SENDCTRL_OP_ENABLE;
680 
681 	if (action[next_state].op_intenable)
682 		op |= SDMA_SENDCTRL_OP_INTENABLE;
683 
684 	if (action[next_state].op_halt)
685 		op |= SDMA_SENDCTRL_OP_HALT;
686 
687 	if (action[next_state].op_cleanup)
688 		op |= SDMA_SENDCTRL_OP_CLEANUP;
689 
690 	if (action[next_state].go_s99_running_tofalse)
691 		ss->go_s99_running = 0;
692 
693 	if (action[next_state].go_s99_running_totrue)
694 		ss->go_s99_running = 1;
695 
696 	ss->current_op = op;
697 	sdma_sendctrl(sde, ss->current_op);
698 }
699 
700 /**
701  * sdma_get_descq_cnt() - called when device probed
702  *
703  * Return a validated descq count.
704  *
705  * This is currently only used in the verbs initialization to build the tx
706  * list.
707  *
708  * This will probably be deleted in favor of a more scalable approach to
709  * alloc tx's.
710  *
711  */
712 u16 sdma_get_descq_cnt(void)
713 {
714 	u16 count = sdma_descq_cnt;
715 
716 	if (!count)
717 		return SDMA_DESCQ_CNT;
718 	/* count must be a power of 2 greater than 64 and less than
719 	 * 32768.   Otherwise return default.
720 	 */
721 	if (!is_power_of_2(count))
722 		return SDMA_DESCQ_CNT;
723 	if (count < 64 || count > 32768)
724 		return SDMA_DESCQ_CNT;
725 	return count;
726 }
727 
728 /**
729  * sdma_engine_get_vl() - return vl for a given sdma engine
730  * @sde: sdma engine
731  *
732  * This function returns the vl mapped to a given engine, or an error if
733  * the mapping can't be found. The mapping fields are protected by RCU.
734  */
735 int sdma_engine_get_vl(struct sdma_engine *sde)
736 {
737 	struct hfi1_devdata *dd = sde->dd;
738 	struct sdma_vl_map *m;
739 	u8 vl;
740 
741 	if (sde->this_idx >= TXE_NUM_SDMA_ENGINES)
742 		return -EINVAL;
743 
744 	rcu_read_lock();
745 	m = rcu_dereference(dd->sdma_map);
746 	if (unlikely(!m)) {
747 		rcu_read_unlock();
748 		return -EINVAL;
749 	}
750 	vl = m->engine_to_vl[sde->this_idx];
751 	rcu_read_unlock();
752 
753 	return vl;
754 }
755 
756 /**
757  * sdma_select_engine_vl() - select sdma engine
758  * @dd: devdata
759  * @selector: a spreading factor
760  * @vl: this vl
761  *
762  *
763  * This function returns an engine based on the selector and a vl.  The
764  * mapping fields are protected by RCU.
765  */
766 struct sdma_engine *sdma_select_engine_vl(
767 	struct hfi1_devdata *dd,
768 	u32 selector,
769 	u8 vl)
770 {
771 	struct sdma_vl_map *m;
772 	struct sdma_map_elem *e;
773 	struct sdma_engine *rval;
774 
775 	/* NOTE This should only happen if SC->VL changed after the initial
776 	 *      checks on the QP/AH
777 	 *      Default will return engine 0 below
778 	 */
779 	if (vl >= num_vls) {
780 		rval = NULL;
781 		goto done;
782 	}
783 
784 	rcu_read_lock();
785 	m = rcu_dereference(dd->sdma_map);
786 	if (unlikely(!m)) {
787 		rcu_read_unlock();
788 		return &dd->per_sdma[0];
789 	}
790 	e = m->map[vl & m->mask];
791 	rval = e->sde[selector & e->mask];
792 	rcu_read_unlock();
793 
794 done:
795 	rval =  !rval ? &dd->per_sdma[0] : rval;
796 	trace_hfi1_sdma_engine_select(dd, selector, vl, rval->this_idx);
797 	return rval;
798 }
799 
800 /**
801  * sdma_select_engine_sc() - select sdma engine
802  * @dd: devdata
803  * @selector: a spreading factor
804  * @sc5: the 5 bit sc
805  *
806  *
807  * This function returns an engine based on the selector and an sc.
808  */
809 struct sdma_engine *sdma_select_engine_sc(
810 	struct hfi1_devdata *dd,
811 	u32 selector,
812 	u8 sc5)
813 {
814 	u8 vl = sc_to_vlt(dd, sc5);
815 
816 	return sdma_select_engine_vl(dd, selector, vl);
817 }
818 
819 struct sdma_rht_map_elem {
820 	u32 mask;
821 	u8 ctr;
822 	struct sdma_engine *sde[0];
823 };
824 
825 struct sdma_rht_node {
826 	unsigned long cpu_id;
827 	struct sdma_rht_map_elem *map[HFI1_MAX_VLS_SUPPORTED];
828 	struct rhash_head node;
829 };
830 
831 #define NR_CPUS_HINT 192
832 
833 static const struct rhashtable_params sdma_rht_params = {
834 	.nelem_hint = NR_CPUS_HINT,
835 	.head_offset = offsetof(struct sdma_rht_node, node),
836 	.key_offset = offsetof(struct sdma_rht_node, cpu_id),
837 	.key_len = FIELD_SIZEOF(struct sdma_rht_node, cpu_id),
838 	.max_size = NR_CPUS,
839 	.min_size = 8,
840 	.automatic_shrinking = true,
841 };
842 
843 /*
844  * sdma_select_user_engine() - select sdma engine based on user setup
845  * @dd: devdata
846  * @selector: a spreading factor
847  * @vl: this vl
848  *
849  * This function returns an sdma engine for a user sdma request.
850  * User defined sdma engine affinity setting is honored when applicable,
851  * otherwise system default sdma engine mapping is used. To ensure correct
852  * ordering, the mapping from <selector, vl> to sde must remain unchanged.
853  */
854 struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd,
855 					    u32 selector, u8 vl)
856 {
857 	struct sdma_rht_node *rht_node;
858 	struct sdma_engine *sde = NULL;
859 	const struct cpumask *current_mask = &current->cpus_allowed;
860 	unsigned long cpu_id;
861 
862 	/*
863 	 * To ensure that always the same sdma engine(s) will be
864 	 * selected make sure the process is pinned to this CPU only.
865 	 */
866 	if (cpumask_weight(current_mask) != 1)
867 		goto out;
868 
869 	cpu_id = smp_processor_id();
870 	rcu_read_lock();
871 	rht_node = rhashtable_lookup_fast(&dd->sdma_rht, &cpu_id,
872 					  sdma_rht_params);
873 
874 	if (rht_node && rht_node->map[vl]) {
875 		struct sdma_rht_map_elem *map = rht_node->map[vl];
876 
877 		sde = map->sde[selector & map->mask];
878 	}
879 	rcu_read_unlock();
880 
881 	if (sde)
882 		return sde;
883 
884 out:
885 	return sdma_select_engine_vl(dd, selector, vl);
886 }
887 
888 static void sdma_populate_sde_map(struct sdma_rht_map_elem *map)
889 {
890 	int i;
891 
892 	for (i = 0; i < roundup_pow_of_two(map->ctr ? : 1) - map->ctr; i++)
893 		map->sde[map->ctr + i] = map->sde[i];
894 }
895 
896 static void sdma_cleanup_sde_map(struct sdma_rht_map_elem *map,
897 				 struct sdma_engine *sde)
898 {
899 	unsigned int i, pow;
900 
901 	/* only need to check the first ctr entries for a match */
902 	for (i = 0; i < map->ctr; i++) {
903 		if (map->sde[i] == sde) {
904 			memmove(&map->sde[i], &map->sde[i + 1],
905 				(map->ctr - i - 1) * sizeof(map->sde[0]));
906 			map->ctr--;
907 			pow = roundup_pow_of_two(map->ctr ? : 1);
908 			map->mask = pow - 1;
909 			sdma_populate_sde_map(map);
910 			break;
911 		}
912 	}
913 }
914 
915 /*
916  * Prevents concurrent reads and writes of the sdma engine cpu_mask
917  */
918 static DEFINE_MUTEX(process_to_sde_mutex);
919 
920 ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf,
921 				size_t count)
922 {
923 	struct hfi1_devdata *dd = sde->dd;
924 	cpumask_var_t mask, new_mask;
925 	unsigned long cpu;
926 	int ret, vl, sz;
927 
928 	vl = sdma_engine_get_vl(sde);
929 	if (unlikely(vl < 0))
930 		return -EINVAL;
931 
932 	ret = zalloc_cpumask_var(&mask, GFP_KERNEL);
933 	if (!ret)
934 		return -ENOMEM;
935 
936 	ret = zalloc_cpumask_var(&new_mask, GFP_KERNEL);
937 	if (!ret) {
938 		free_cpumask_var(mask);
939 		return -ENOMEM;
940 	}
941 	ret = cpulist_parse(buf, mask);
942 	if (ret)
943 		goto out_free;
944 
945 	if (!cpumask_subset(mask, cpu_online_mask)) {
946 		dd_dev_warn(sde->dd, "Invalid CPU mask\n");
947 		ret = -EINVAL;
948 		goto out_free;
949 	}
950 
951 	sz = sizeof(struct sdma_rht_map_elem) +
952 			(TXE_NUM_SDMA_ENGINES * sizeof(struct sdma_engine *));
953 
954 	mutex_lock(&process_to_sde_mutex);
955 
956 	for_each_cpu(cpu, mask) {
957 		struct sdma_rht_node *rht_node;
958 
959 		/* Check if we have this already mapped */
960 		if (cpumask_test_cpu(cpu, &sde->cpu_mask)) {
961 			cpumask_set_cpu(cpu, new_mask);
962 			continue;
963 		}
964 
965 		rht_node = rhashtable_lookup_fast(&dd->sdma_rht, &cpu,
966 						  sdma_rht_params);
967 		if (!rht_node) {
968 			rht_node = kzalloc(sizeof(*rht_node), GFP_KERNEL);
969 			if (!rht_node) {
970 				ret = -ENOMEM;
971 				goto out;
972 			}
973 
974 			rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
975 			if (!rht_node->map[vl]) {
976 				kfree(rht_node);
977 				ret = -ENOMEM;
978 				goto out;
979 			}
980 			rht_node->cpu_id = cpu;
981 			rht_node->map[vl]->mask = 0;
982 			rht_node->map[vl]->ctr = 1;
983 			rht_node->map[vl]->sde[0] = sde;
984 
985 			ret = rhashtable_insert_fast(&dd->sdma_rht,
986 						     &rht_node->node,
987 						     sdma_rht_params);
988 			if (ret) {
989 				kfree(rht_node->map[vl]);
990 				kfree(rht_node);
991 				dd_dev_err(sde->dd, "Failed to set process to sde affinity for cpu %lu\n",
992 					   cpu);
993 				goto out;
994 			}
995 
996 		} else {
997 			int ctr, pow;
998 
999 			/* Add new user mappings */
1000 			if (!rht_node->map[vl])
1001 				rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
1002 
1003 			if (!rht_node->map[vl]) {
1004 				ret = -ENOMEM;
1005 				goto out;
1006 			}
1007 
1008 			rht_node->map[vl]->ctr++;
1009 			ctr = rht_node->map[vl]->ctr;
1010 			rht_node->map[vl]->sde[ctr - 1] = sde;
1011 			pow = roundup_pow_of_two(ctr);
1012 			rht_node->map[vl]->mask = pow - 1;
1013 
1014 			/* Populate the sde map table */
1015 			sdma_populate_sde_map(rht_node->map[vl]);
1016 		}
1017 		cpumask_set_cpu(cpu, new_mask);
1018 	}
1019 
1020 	/* Clean up old mappings */
1021 	for_each_cpu(cpu, cpu_online_mask) {
1022 		struct sdma_rht_node *rht_node;
1023 
1024 		/* Don't cleanup sdes that are set in the new mask */
1025 		if (cpumask_test_cpu(cpu, mask))
1026 			continue;
1027 
1028 		rht_node = rhashtable_lookup_fast(&dd->sdma_rht, &cpu,
1029 						  sdma_rht_params);
1030 		if (rht_node) {
1031 			bool empty = true;
1032 			int i;
1033 
1034 			/* Remove mappings for old sde */
1035 			for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1036 				if (rht_node->map[i])
1037 					sdma_cleanup_sde_map(rht_node->map[i],
1038 							     sde);
1039 
1040 			/* Free empty hash table entries */
1041 			for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1042 				if (!rht_node->map[i])
1043 					continue;
1044 
1045 				if (rht_node->map[i]->ctr) {
1046 					empty = false;
1047 					break;
1048 				}
1049 			}
1050 
1051 			if (empty) {
1052 				ret = rhashtable_remove_fast(&dd->sdma_rht,
1053 							     &rht_node->node,
1054 							     sdma_rht_params);
1055 				WARN_ON(ret);
1056 
1057 				for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1058 					kfree(rht_node->map[i]);
1059 
1060 				kfree(rht_node);
1061 			}
1062 		}
1063 	}
1064 
1065 	cpumask_copy(&sde->cpu_mask, new_mask);
1066 out:
1067 	mutex_unlock(&process_to_sde_mutex);
1068 out_free:
1069 	free_cpumask_var(mask);
1070 	free_cpumask_var(new_mask);
1071 	return ret ? : strnlen(buf, PAGE_SIZE);
1072 }
1073 
1074 ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf)
1075 {
1076 	mutex_lock(&process_to_sde_mutex);
1077 	if (cpumask_empty(&sde->cpu_mask))
1078 		snprintf(buf, PAGE_SIZE, "%s\n", "empty");
1079 	else
1080 		cpumap_print_to_pagebuf(true, buf, &sde->cpu_mask);
1081 	mutex_unlock(&process_to_sde_mutex);
1082 	return strnlen(buf, PAGE_SIZE);
1083 }
1084 
1085 static void sdma_rht_free(void *ptr, void *arg)
1086 {
1087 	struct sdma_rht_node *rht_node = ptr;
1088 	int i;
1089 
1090 	for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1091 		kfree(rht_node->map[i]);
1092 
1093 	kfree(rht_node);
1094 }
1095 
1096 /**
1097  * sdma_seqfile_dump_cpu_list() - debugfs dump the cpu to sdma mappings
1098  * @s: seq file
1099  * @dd: hfi1_devdata
1100  * @cpuid: cpu id
1101  *
1102  * This routine dumps the process to sde mappings per cpu
1103  */
1104 void sdma_seqfile_dump_cpu_list(struct seq_file *s,
1105 				struct hfi1_devdata *dd,
1106 				unsigned long cpuid)
1107 {
1108 	struct sdma_rht_node *rht_node;
1109 	int i, j;
1110 
1111 	rht_node = rhashtable_lookup_fast(&dd->sdma_rht, &cpuid,
1112 					  sdma_rht_params);
1113 	if (!rht_node)
1114 		return;
1115 
1116 	seq_printf(s, "cpu%3lu: ", cpuid);
1117 	for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1118 		if (!rht_node->map[i] || !rht_node->map[i]->ctr)
1119 			continue;
1120 
1121 		seq_printf(s, " vl%d: [", i);
1122 
1123 		for (j = 0; j < rht_node->map[i]->ctr; j++) {
1124 			if (!rht_node->map[i]->sde[j])
1125 				continue;
1126 
1127 			if (j > 0)
1128 				seq_puts(s, ",");
1129 
1130 			seq_printf(s, " sdma%2d",
1131 				   rht_node->map[i]->sde[j]->this_idx);
1132 		}
1133 		seq_puts(s, " ]");
1134 	}
1135 
1136 	seq_puts(s, "\n");
1137 }
1138 
1139 /*
1140  * Free the indicated map struct
1141  */
1142 static void sdma_map_free(struct sdma_vl_map *m)
1143 {
1144 	int i;
1145 
1146 	for (i = 0; m && i < m->actual_vls; i++)
1147 		kfree(m->map[i]);
1148 	kfree(m);
1149 }
1150 
1151 /*
1152  * Handle RCU callback
1153  */
1154 static void sdma_map_rcu_callback(struct rcu_head *list)
1155 {
1156 	struct sdma_vl_map *m = container_of(list, struct sdma_vl_map, list);
1157 
1158 	sdma_map_free(m);
1159 }
1160 
1161 /**
1162  * sdma_map_init - called when # vls change
1163  * @dd: hfi1_devdata
1164  * @port: port number
1165  * @num_vls: number of vls
1166  * @vl_engines: per vl engine mapping (optional)
1167  *
1168  * This routine changes the mapping based on the number of vls.
1169  *
1170  * vl_engines is used to specify a non-uniform vl/engine loading. NULL
1171  * implies auto computing the loading and giving each VLs a uniform
1172  * distribution of engines per VL.
1173  *
1174  * The auto algorithm computes the sde_per_vl and the number of extra
1175  * engines.  Any extra engines are added from the last VL on down.
1176  *
1177  * rcu locking is used here to control access to the mapping fields.
1178  *
1179  * If either the num_vls or num_sdma are non-power of 2, the array sizes
1180  * in the struct sdma_vl_map and the struct sdma_map_elem are rounded
1181  * up to the next highest power of 2 and the first entry is reused
1182  * in a round robin fashion.
1183  *
1184  * If an error occurs the map change is not done and the mapping is
1185  * not changed.
1186  *
1187  */
1188 int sdma_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_engines)
1189 {
1190 	int i, j;
1191 	int extra, sde_per_vl;
1192 	int engine = 0;
1193 	u8 lvl_engines[OPA_MAX_VLS];
1194 	struct sdma_vl_map *oldmap, *newmap;
1195 
1196 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
1197 		return 0;
1198 
1199 	if (!vl_engines) {
1200 		/* truncate divide */
1201 		sde_per_vl = dd->num_sdma / num_vls;
1202 		/* extras */
1203 		extra = dd->num_sdma % num_vls;
1204 		vl_engines = lvl_engines;
1205 		/* add extras from last vl down */
1206 		for (i = num_vls - 1; i >= 0; i--, extra--)
1207 			vl_engines[i] = sde_per_vl + (extra > 0 ? 1 : 0);
1208 	}
1209 	/* build new map */
1210 	newmap = kzalloc(
1211 		sizeof(struct sdma_vl_map) +
1212 			roundup_pow_of_two(num_vls) *
1213 			sizeof(struct sdma_map_elem *),
1214 		GFP_KERNEL);
1215 	if (!newmap)
1216 		goto bail;
1217 	newmap->actual_vls = num_vls;
1218 	newmap->vls = roundup_pow_of_two(num_vls);
1219 	newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1220 	/* initialize back-map */
1221 	for (i = 0; i < TXE_NUM_SDMA_ENGINES; i++)
1222 		newmap->engine_to_vl[i] = -1;
1223 	for (i = 0; i < newmap->vls; i++) {
1224 		/* save for wrap around */
1225 		int first_engine = engine;
1226 
1227 		if (i < newmap->actual_vls) {
1228 			int sz = roundup_pow_of_two(vl_engines[i]);
1229 
1230 			/* only allocate once */
1231 			newmap->map[i] = kzalloc(
1232 				sizeof(struct sdma_map_elem) +
1233 					sz * sizeof(struct sdma_engine *),
1234 				GFP_KERNEL);
1235 			if (!newmap->map[i])
1236 				goto bail;
1237 			newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1238 			/* assign engines */
1239 			for (j = 0; j < sz; j++) {
1240 				newmap->map[i]->sde[j] =
1241 					&dd->per_sdma[engine];
1242 				if (++engine >= first_engine + vl_engines[i])
1243 					/* wrap back to first engine */
1244 					engine = first_engine;
1245 			}
1246 			/* assign back-map */
1247 			for (j = 0; j < vl_engines[i]; j++)
1248 				newmap->engine_to_vl[first_engine + j] = i;
1249 		} else {
1250 			/* just re-use entry without allocating */
1251 			newmap->map[i] = newmap->map[i % num_vls];
1252 		}
1253 		engine = first_engine + vl_engines[i];
1254 	}
1255 	/* newmap in hand, save old map */
1256 	spin_lock_irq(&dd->sde_map_lock);
1257 	oldmap = rcu_dereference_protected(dd->sdma_map,
1258 					   lockdep_is_held(&dd->sde_map_lock));
1259 
1260 	/* publish newmap */
1261 	rcu_assign_pointer(dd->sdma_map, newmap);
1262 
1263 	spin_unlock_irq(&dd->sde_map_lock);
1264 	/* success, free any old map after grace period */
1265 	if (oldmap)
1266 		call_rcu(&oldmap->list, sdma_map_rcu_callback);
1267 	return 0;
1268 bail:
1269 	/* free any partial allocation */
1270 	sdma_map_free(newmap);
1271 	return -ENOMEM;
1272 }
1273 
1274 /*
1275  * Clean up allocated memory.
1276  *
1277  * This routine is can be called regardless of the success of sdma_init()
1278  *
1279  */
1280 static void sdma_clean(struct hfi1_devdata *dd, size_t num_engines)
1281 {
1282 	size_t i;
1283 	struct sdma_engine *sde;
1284 
1285 	if (dd->sdma_pad_dma) {
1286 		dma_free_coherent(&dd->pcidev->dev, 4,
1287 				  (void *)dd->sdma_pad_dma,
1288 				  dd->sdma_pad_phys);
1289 		dd->sdma_pad_dma = NULL;
1290 		dd->sdma_pad_phys = 0;
1291 	}
1292 	if (dd->sdma_heads_dma) {
1293 		dma_free_coherent(&dd->pcidev->dev, dd->sdma_heads_size,
1294 				  (void *)dd->sdma_heads_dma,
1295 				  dd->sdma_heads_phys);
1296 		dd->sdma_heads_dma = NULL;
1297 		dd->sdma_heads_phys = 0;
1298 	}
1299 	for (i = 0; dd->per_sdma && i < num_engines; ++i) {
1300 		sde = &dd->per_sdma[i];
1301 
1302 		sde->head_dma = NULL;
1303 		sde->head_phys = 0;
1304 
1305 		if (sde->descq) {
1306 			dma_free_coherent(
1307 				&dd->pcidev->dev,
1308 				sde->descq_cnt * sizeof(u64[2]),
1309 				sde->descq,
1310 				sde->descq_phys
1311 			);
1312 			sde->descq = NULL;
1313 			sde->descq_phys = 0;
1314 		}
1315 		kvfree(sde->tx_ring);
1316 		sde->tx_ring = NULL;
1317 	}
1318 	spin_lock_irq(&dd->sde_map_lock);
1319 	sdma_map_free(rcu_access_pointer(dd->sdma_map));
1320 	RCU_INIT_POINTER(dd->sdma_map, NULL);
1321 	spin_unlock_irq(&dd->sde_map_lock);
1322 	synchronize_rcu();
1323 	kfree(dd->per_sdma);
1324 	dd->per_sdma = NULL;
1325 }
1326 
1327 /**
1328  * sdma_init() - called when device probed
1329  * @dd: hfi1_devdata
1330  * @port: port number (currently only zero)
1331  *
1332  * sdma_init initializes the specified number of engines.
1333  *
1334  * The code initializes each sde, its csrs.  Interrupts
1335  * are not required to be enabled.
1336  *
1337  * Returns:
1338  * 0 - success, -errno on failure
1339  */
1340 int sdma_init(struct hfi1_devdata *dd, u8 port)
1341 {
1342 	unsigned this_idx;
1343 	struct sdma_engine *sde;
1344 	u16 descq_cnt;
1345 	void *curr_head;
1346 	struct hfi1_pportdata *ppd = dd->pport + port;
1347 	u32 per_sdma_credits;
1348 	uint idle_cnt = sdma_idle_cnt;
1349 	size_t num_engines = dd->chip_sdma_engines;
1350 
1351 	if (!HFI1_CAP_IS_KSET(SDMA)) {
1352 		HFI1_CAP_CLEAR(SDMA_AHG);
1353 		return 0;
1354 	}
1355 	if (mod_num_sdma &&
1356 	    /* can't exceed chip support */
1357 	    mod_num_sdma <= dd->chip_sdma_engines &&
1358 	    /* count must be >= vls */
1359 	    mod_num_sdma >= num_vls)
1360 		num_engines = mod_num_sdma;
1361 
1362 	dd_dev_info(dd, "SDMA mod_num_sdma: %u\n", mod_num_sdma);
1363 	dd_dev_info(dd, "SDMA chip_sdma_engines: %u\n", dd->chip_sdma_engines);
1364 	dd_dev_info(dd, "SDMA chip_sdma_mem_size: %u\n",
1365 		    dd->chip_sdma_mem_size);
1366 
1367 	per_sdma_credits =
1368 		dd->chip_sdma_mem_size / (num_engines * SDMA_BLOCK_SIZE);
1369 
1370 	/* set up freeze waitqueue */
1371 	init_waitqueue_head(&dd->sdma_unfreeze_wq);
1372 	atomic_set(&dd->sdma_unfreeze_count, 0);
1373 
1374 	descq_cnt = sdma_get_descq_cnt();
1375 	dd_dev_info(dd, "SDMA engines %zu descq_cnt %u\n",
1376 		    num_engines, descq_cnt);
1377 
1378 	/* alloc memory for array of send engines */
1379 	dd->per_sdma = kcalloc(num_engines, sizeof(*dd->per_sdma), GFP_KERNEL);
1380 	if (!dd->per_sdma)
1381 		return -ENOMEM;
1382 
1383 	idle_cnt = ns_to_cclock(dd, idle_cnt);
1384 	if (!sdma_desct_intr)
1385 		sdma_desct_intr = SDMA_DESC_INTR;
1386 
1387 	/* Allocate memory for SendDMA descriptor FIFOs */
1388 	for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1389 		sde = &dd->per_sdma[this_idx];
1390 		sde->dd = dd;
1391 		sde->ppd = ppd;
1392 		sde->this_idx = this_idx;
1393 		sde->descq_cnt = descq_cnt;
1394 		sde->desc_avail = sdma_descq_freecnt(sde);
1395 		sde->sdma_shift = ilog2(descq_cnt);
1396 		sde->sdma_mask = (1 << sde->sdma_shift) - 1;
1397 
1398 		/* Create a mask specifically for each interrupt source */
1399 		sde->int_mask = (u64)1 << (0 * TXE_NUM_SDMA_ENGINES +
1400 					   this_idx);
1401 		sde->progress_mask = (u64)1 << (1 * TXE_NUM_SDMA_ENGINES +
1402 						this_idx);
1403 		sde->idle_mask = (u64)1 << (2 * TXE_NUM_SDMA_ENGINES +
1404 					    this_idx);
1405 		/* Create a combined mask to cover all 3 interrupt sources */
1406 		sde->imask = sde->int_mask | sde->progress_mask |
1407 			     sde->idle_mask;
1408 
1409 		spin_lock_init(&sde->tail_lock);
1410 		seqlock_init(&sde->head_lock);
1411 		spin_lock_init(&sde->senddmactrl_lock);
1412 		spin_lock_init(&sde->flushlist_lock);
1413 		/* insure there is always a zero bit */
1414 		sde->ahg_bits = 0xfffffffe00000000ULL;
1415 
1416 		sdma_set_state(sde, sdma_state_s00_hw_down);
1417 
1418 		/* set up reference counting */
1419 		kref_init(&sde->state.kref);
1420 		init_completion(&sde->state.comp);
1421 
1422 		INIT_LIST_HEAD(&sde->flushlist);
1423 		INIT_LIST_HEAD(&sde->dmawait);
1424 
1425 		sde->tail_csr =
1426 			get_kctxt_csr_addr(dd, this_idx, SD(TAIL));
1427 
1428 		if (idle_cnt)
1429 			dd->default_desc1 =
1430 				SDMA_DESC1_HEAD_TO_HOST_FLAG;
1431 		else
1432 			dd->default_desc1 =
1433 				SDMA_DESC1_INT_REQ_FLAG;
1434 
1435 		tasklet_init(&sde->sdma_hw_clean_up_task, sdma_hw_clean_up_task,
1436 			     (unsigned long)sde);
1437 
1438 		tasklet_init(&sde->sdma_sw_clean_up_task, sdma_sw_clean_up_task,
1439 			     (unsigned long)sde);
1440 		INIT_WORK(&sde->err_halt_worker, sdma_err_halt_wait);
1441 		INIT_WORK(&sde->flush_worker, sdma_field_flush);
1442 
1443 		sde->progress_check_head = 0;
1444 
1445 		setup_timer(&sde->err_progress_check_timer,
1446 			    sdma_err_progress_check, (unsigned long)sde);
1447 
1448 		sde->descq = dma_zalloc_coherent(
1449 			&dd->pcidev->dev,
1450 			descq_cnt * sizeof(u64[2]),
1451 			&sde->descq_phys,
1452 			GFP_KERNEL
1453 		);
1454 		if (!sde->descq)
1455 			goto bail;
1456 		sde->tx_ring =
1457 			kcalloc(descq_cnt, sizeof(struct sdma_txreq *),
1458 				GFP_KERNEL);
1459 		if (!sde->tx_ring)
1460 			sde->tx_ring =
1461 				vzalloc(
1462 					sizeof(struct sdma_txreq *) *
1463 					descq_cnt);
1464 		if (!sde->tx_ring)
1465 			goto bail;
1466 	}
1467 
1468 	dd->sdma_heads_size = L1_CACHE_BYTES * num_engines;
1469 	/* Allocate memory for DMA of head registers to memory */
1470 	dd->sdma_heads_dma = dma_zalloc_coherent(
1471 		&dd->pcidev->dev,
1472 		dd->sdma_heads_size,
1473 		&dd->sdma_heads_phys,
1474 		GFP_KERNEL
1475 	);
1476 	if (!dd->sdma_heads_dma) {
1477 		dd_dev_err(dd, "failed to allocate SendDMA head memory\n");
1478 		goto bail;
1479 	}
1480 
1481 	/* Allocate memory for pad */
1482 	dd->sdma_pad_dma = dma_zalloc_coherent(
1483 		&dd->pcidev->dev,
1484 		sizeof(u32),
1485 		&dd->sdma_pad_phys,
1486 		GFP_KERNEL
1487 	);
1488 	if (!dd->sdma_pad_dma) {
1489 		dd_dev_err(dd, "failed to allocate SendDMA pad memory\n");
1490 		goto bail;
1491 	}
1492 
1493 	/* assign each engine to different cacheline and init registers */
1494 	curr_head = (void *)dd->sdma_heads_dma;
1495 	for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1496 		unsigned long phys_offset;
1497 
1498 		sde = &dd->per_sdma[this_idx];
1499 
1500 		sde->head_dma = curr_head;
1501 		curr_head += L1_CACHE_BYTES;
1502 		phys_offset = (unsigned long)sde->head_dma -
1503 			      (unsigned long)dd->sdma_heads_dma;
1504 		sde->head_phys = dd->sdma_heads_phys + phys_offset;
1505 		init_sdma_regs(sde, per_sdma_credits, idle_cnt);
1506 	}
1507 	dd->flags |= HFI1_HAS_SEND_DMA;
1508 	dd->flags |= idle_cnt ? HFI1_HAS_SDMA_TIMEOUT : 0;
1509 	dd->num_sdma = num_engines;
1510 	if (sdma_map_init(dd, port, ppd->vls_operational, NULL))
1511 		goto bail;
1512 
1513 	if (rhashtable_init(&dd->sdma_rht, &sdma_rht_params))
1514 		goto bail;
1515 
1516 	dd_dev_info(dd, "SDMA num_sdma: %u\n", dd->num_sdma);
1517 	return 0;
1518 
1519 bail:
1520 	sdma_clean(dd, num_engines);
1521 	return -ENOMEM;
1522 }
1523 
1524 /**
1525  * sdma_all_running() - called when the link goes up
1526  * @dd: hfi1_devdata
1527  *
1528  * This routine moves all engines to the running state.
1529  */
1530 void sdma_all_running(struct hfi1_devdata *dd)
1531 {
1532 	struct sdma_engine *sde;
1533 	unsigned int i;
1534 
1535 	/* move all engines to running */
1536 	for (i = 0; i < dd->num_sdma; ++i) {
1537 		sde = &dd->per_sdma[i];
1538 		sdma_process_event(sde, sdma_event_e30_go_running);
1539 	}
1540 }
1541 
1542 /**
1543  * sdma_all_idle() - called when the link goes down
1544  * @dd: hfi1_devdata
1545  *
1546  * This routine moves all engines to the idle state.
1547  */
1548 void sdma_all_idle(struct hfi1_devdata *dd)
1549 {
1550 	struct sdma_engine *sde;
1551 	unsigned int i;
1552 
1553 	/* idle all engines */
1554 	for (i = 0; i < dd->num_sdma; ++i) {
1555 		sde = &dd->per_sdma[i];
1556 		sdma_process_event(sde, sdma_event_e70_go_idle);
1557 	}
1558 }
1559 
1560 /**
1561  * sdma_start() - called to kick off state processing for all engines
1562  * @dd: hfi1_devdata
1563  *
1564  * This routine is for kicking off the state processing for all required
1565  * sdma engines.  Interrupts need to be working at this point.
1566  *
1567  */
1568 void sdma_start(struct hfi1_devdata *dd)
1569 {
1570 	unsigned i;
1571 	struct sdma_engine *sde;
1572 
1573 	/* kick off the engines state processing */
1574 	for (i = 0; i < dd->num_sdma; ++i) {
1575 		sde = &dd->per_sdma[i];
1576 		sdma_process_event(sde, sdma_event_e10_go_hw_start);
1577 	}
1578 }
1579 
1580 /**
1581  * sdma_exit() - used when module is removed
1582  * @dd: hfi1_devdata
1583  */
1584 void sdma_exit(struct hfi1_devdata *dd)
1585 {
1586 	unsigned this_idx;
1587 	struct sdma_engine *sde;
1588 
1589 	for (this_idx = 0; dd->per_sdma && this_idx < dd->num_sdma;
1590 			++this_idx) {
1591 		sde = &dd->per_sdma[this_idx];
1592 		if (!list_empty(&sde->dmawait))
1593 			dd_dev_err(dd, "sde %u: dmawait list not empty!\n",
1594 				   sde->this_idx);
1595 		sdma_process_event(sde, sdma_event_e00_go_hw_down);
1596 
1597 		del_timer_sync(&sde->err_progress_check_timer);
1598 
1599 		/*
1600 		 * This waits for the state machine to exit so it is not
1601 		 * necessary to kill the sdma_sw_clean_up_task to make sure
1602 		 * it is not running.
1603 		 */
1604 		sdma_finalput(&sde->state);
1605 	}
1606 	sdma_clean(dd, dd->num_sdma);
1607 	rhashtable_free_and_destroy(&dd->sdma_rht, sdma_rht_free, NULL);
1608 }
1609 
1610 /*
1611  * unmap the indicated descriptor
1612  */
1613 static inline void sdma_unmap_desc(
1614 	struct hfi1_devdata *dd,
1615 	struct sdma_desc *descp)
1616 {
1617 	switch (sdma_mapping_type(descp)) {
1618 	case SDMA_MAP_SINGLE:
1619 		dma_unmap_single(
1620 			&dd->pcidev->dev,
1621 			sdma_mapping_addr(descp),
1622 			sdma_mapping_len(descp),
1623 			DMA_TO_DEVICE);
1624 		break;
1625 	case SDMA_MAP_PAGE:
1626 		dma_unmap_page(
1627 			&dd->pcidev->dev,
1628 			sdma_mapping_addr(descp),
1629 			sdma_mapping_len(descp),
1630 			DMA_TO_DEVICE);
1631 		break;
1632 	}
1633 }
1634 
1635 /*
1636  * return the mode as indicated by the first
1637  * descriptor in the tx.
1638  */
1639 static inline u8 ahg_mode(struct sdma_txreq *tx)
1640 {
1641 	return (tx->descp[0].qw[1] & SDMA_DESC1_HEADER_MODE_SMASK)
1642 		>> SDMA_DESC1_HEADER_MODE_SHIFT;
1643 }
1644 
1645 /**
1646  * __sdma_txclean() - clean tx of mappings, descp *kmalloc's
1647  * @dd: hfi1_devdata for unmapping
1648  * @tx: tx request to clean
1649  *
1650  * This is used in the progress routine to clean the tx or
1651  * by the ULP to toss an in-process tx build.
1652  *
1653  * The code can be called multiple times without issue.
1654  *
1655  */
1656 void __sdma_txclean(
1657 	struct hfi1_devdata *dd,
1658 	struct sdma_txreq *tx)
1659 {
1660 	u16 i;
1661 
1662 	if (tx->num_desc) {
1663 		u8 skip = 0, mode = ahg_mode(tx);
1664 
1665 		/* unmap first */
1666 		sdma_unmap_desc(dd, &tx->descp[0]);
1667 		/* determine number of AHG descriptors to skip */
1668 		if (mode > SDMA_AHG_APPLY_UPDATE1)
1669 			skip = mode >> 1;
1670 		for (i = 1 + skip; i < tx->num_desc; i++)
1671 			sdma_unmap_desc(dd, &tx->descp[i]);
1672 		tx->num_desc = 0;
1673 	}
1674 	kfree(tx->coalesce_buf);
1675 	tx->coalesce_buf = NULL;
1676 	/* kmalloc'ed descp */
1677 	if (unlikely(tx->desc_limit > ARRAY_SIZE(tx->descs))) {
1678 		tx->desc_limit = ARRAY_SIZE(tx->descs);
1679 		kfree(tx->descp);
1680 	}
1681 }
1682 
1683 static inline u16 sdma_gethead(struct sdma_engine *sde)
1684 {
1685 	struct hfi1_devdata *dd = sde->dd;
1686 	int use_dmahead;
1687 	u16 hwhead;
1688 
1689 #ifdef CONFIG_SDMA_VERBOSITY
1690 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1691 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1692 #endif
1693 
1694 retry:
1695 	use_dmahead = HFI1_CAP_IS_KSET(USE_SDMA_HEAD) && __sdma_running(sde) &&
1696 					(dd->flags & HFI1_HAS_SDMA_TIMEOUT);
1697 	hwhead = use_dmahead ?
1698 		(u16)le64_to_cpu(*sde->head_dma) :
1699 		(u16)read_sde_csr(sde, SD(HEAD));
1700 
1701 	if (unlikely(HFI1_CAP_IS_KSET(SDMA_HEAD_CHECK))) {
1702 		u16 cnt;
1703 		u16 swtail;
1704 		u16 swhead;
1705 		int sane;
1706 
1707 		swhead = sde->descq_head & sde->sdma_mask;
1708 		/* this code is really bad for cache line trading */
1709 		swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
1710 		cnt = sde->descq_cnt;
1711 
1712 		if (swhead < swtail)
1713 			/* not wrapped */
1714 			sane = (hwhead >= swhead) & (hwhead <= swtail);
1715 		else if (swhead > swtail)
1716 			/* wrapped around */
1717 			sane = ((hwhead >= swhead) && (hwhead < cnt)) ||
1718 				(hwhead <= swtail);
1719 		else
1720 			/* empty */
1721 			sane = (hwhead == swhead);
1722 
1723 		if (unlikely(!sane)) {
1724 			dd_dev_err(dd, "SDMA(%u) bad head (%s) hwhd=%hu swhd=%hu swtl=%hu cnt=%hu\n",
1725 				   sde->this_idx,
1726 				   use_dmahead ? "dma" : "kreg",
1727 				   hwhead, swhead, swtail, cnt);
1728 			if (use_dmahead) {
1729 				/* try one more time, using csr */
1730 				use_dmahead = 0;
1731 				goto retry;
1732 			}
1733 			/* proceed as if no progress */
1734 			hwhead = swhead;
1735 		}
1736 	}
1737 	return hwhead;
1738 }
1739 
1740 /*
1741  * This is called when there are send DMA descriptors that might be
1742  * available.
1743  *
1744  * This is called with head_lock held.
1745  */
1746 static void sdma_desc_avail(struct sdma_engine *sde, unsigned avail)
1747 {
1748 	struct iowait *wait, *nw;
1749 	struct iowait *waits[SDMA_WAIT_BATCH_SIZE];
1750 	unsigned i, n = 0, seq;
1751 	struct sdma_txreq *stx;
1752 	struct hfi1_ibdev *dev = &sde->dd->verbs_dev;
1753 
1754 #ifdef CONFIG_SDMA_VERBOSITY
1755 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", sde->this_idx,
1756 		   slashstrip(__FILE__), __LINE__, __func__);
1757 	dd_dev_err(sde->dd, "avail: %u\n", avail);
1758 #endif
1759 
1760 	do {
1761 		seq = read_seqbegin(&dev->iowait_lock);
1762 		if (!list_empty(&sde->dmawait)) {
1763 			/* at least one item */
1764 			write_seqlock(&dev->iowait_lock);
1765 			/* Harvest waiters wanting DMA descriptors */
1766 			list_for_each_entry_safe(
1767 					wait,
1768 					nw,
1769 					&sde->dmawait,
1770 					list) {
1771 				u16 num_desc = 0;
1772 
1773 				if (!wait->wakeup)
1774 					continue;
1775 				if (n == ARRAY_SIZE(waits))
1776 					break;
1777 				if (!list_empty(&wait->tx_head)) {
1778 					stx = list_first_entry(
1779 						&wait->tx_head,
1780 						struct sdma_txreq,
1781 						list);
1782 					num_desc = stx->num_desc;
1783 				}
1784 				if (num_desc > avail)
1785 					break;
1786 				avail -= num_desc;
1787 				list_del_init(&wait->list);
1788 				waits[n++] = wait;
1789 			}
1790 			write_sequnlock(&dev->iowait_lock);
1791 			break;
1792 		}
1793 	} while (read_seqretry(&dev->iowait_lock, seq));
1794 
1795 	for (i = 0; i < n; i++)
1796 		waits[i]->wakeup(waits[i], SDMA_AVAIL_REASON);
1797 }
1798 
1799 /* head_lock must be held */
1800 static void sdma_make_progress(struct sdma_engine *sde, u64 status)
1801 {
1802 	struct sdma_txreq *txp = NULL;
1803 	int progress = 0;
1804 	u16 hwhead, swhead;
1805 	int idle_check_done = 0;
1806 
1807 	hwhead = sdma_gethead(sde);
1808 
1809 	/* The reason for some of the complexity of this code is that
1810 	 * not all descriptors have corresponding txps.  So, we have to
1811 	 * be able to skip over descs until we wander into the range of
1812 	 * the next txp on the list.
1813 	 */
1814 
1815 retry:
1816 	txp = get_txhead(sde);
1817 	swhead = sde->descq_head & sde->sdma_mask;
1818 	trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1819 	while (swhead != hwhead) {
1820 		/* advance head, wrap if needed */
1821 		swhead = ++sde->descq_head & sde->sdma_mask;
1822 
1823 		/* if now past this txp's descs, do the callback */
1824 		if (txp && txp->next_descq_idx == swhead) {
1825 			/* remove from list */
1826 			sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
1827 			complete_tx(sde, txp, SDMA_TXREQ_S_OK);
1828 			/* see if there is another txp */
1829 			txp = get_txhead(sde);
1830 		}
1831 		trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1832 		progress++;
1833 	}
1834 
1835 	/*
1836 	 * The SDMA idle interrupt is not guaranteed to be ordered with respect
1837 	 * to updates to the the dma_head location in host memory. The head
1838 	 * value read might not be fully up to date. If there are pending
1839 	 * descriptors and the SDMA idle interrupt fired then read from the
1840 	 * CSR SDMA head instead to get the latest value from the hardware.
1841 	 * The hardware SDMA head should be read at most once in this invocation
1842 	 * of sdma_make_progress(..) which is ensured by idle_check_done flag
1843 	 */
1844 	if ((status & sde->idle_mask) && !idle_check_done) {
1845 		u16 swtail;
1846 
1847 		swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
1848 		if (swtail != hwhead) {
1849 			hwhead = (u16)read_sde_csr(sde, SD(HEAD));
1850 			idle_check_done = 1;
1851 			goto retry;
1852 		}
1853 	}
1854 
1855 	sde->last_status = status;
1856 	if (progress)
1857 		sdma_desc_avail(sde, sdma_descq_freecnt(sde));
1858 }
1859 
1860 /*
1861  * sdma_engine_interrupt() - interrupt handler for engine
1862  * @sde: sdma engine
1863  * @status: sdma interrupt reason
1864  *
1865  * Status is a mask of the 3 possible interrupts for this engine.  It will
1866  * contain bits _only_ for this SDMA engine.  It will contain at least one
1867  * bit, it may contain more.
1868  */
1869 void sdma_engine_interrupt(struct sdma_engine *sde, u64 status)
1870 {
1871 	trace_hfi1_sdma_engine_interrupt(sde, status);
1872 	write_seqlock(&sde->head_lock);
1873 	sdma_set_desc_cnt(sde, sdma_desct_intr);
1874 	if (status & sde->idle_mask)
1875 		sde->idle_int_cnt++;
1876 	else if (status & sde->progress_mask)
1877 		sde->progress_int_cnt++;
1878 	else if (status & sde->int_mask)
1879 		sde->sdma_int_cnt++;
1880 	sdma_make_progress(sde, status);
1881 	write_sequnlock(&sde->head_lock);
1882 }
1883 
1884 /**
1885  * sdma_engine_error() - error handler for engine
1886  * @sde: sdma engine
1887  * @status: sdma interrupt reason
1888  */
1889 void sdma_engine_error(struct sdma_engine *sde, u64 status)
1890 {
1891 	unsigned long flags;
1892 
1893 #ifdef CONFIG_SDMA_VERBOSITY
1894 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) error status 0x%llx state %s\n",
1895 		   sde->this_idx,
1896 		   (unsigned long long)status,
1897 		   sdma_state_names[sde->state.current_state]);
1898 #endif
1899 	spin_lock_irqsave(&sde->tail_lock, flags);
1900 	write_seqlock(&sde->head_lock);
1901 	if (status & ALL_SDMA_ENG_HALT_ERRS)
1902 		__sdma_process_event(sde, sdma_event_e60_hw_halted);
1903 	if (status & ~SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK)) {
1904 		dd_dev_err(sde->dd,
1905 			   "SDMA (%u) engine error: 0x%llx state %s\n",
1906 			   sde->this_idx,
1907 			   (unsigned long long)status,
1908 			   sdma_state_names[sde->state.current_state]);
1909 		dump_sdma_state(sde);
1910 	}
1911 	write_sequnlock(&sde->head_lock);
1912 	spin_unlock_irqrestore(&sde->tail_lock, flags);
1913 }
1914 
1915 static void sdma_sendctrl(struct sdma_engine *sde, unsigned op)
1916 {
1917 	u64 set_senddmactrl = 0;
1918 	u64 clr_senddmactrl = 0;
1919 	unsigned long flags;
1920 
1921 #ifdef CONFIG_SDMA_VERBOSITY
1922 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) senddmactrl E=%d I=%d H=%d C=%d\n",
1923 		   sde->this_idx,
1924 		   (op & SDMA_SENDCTRL_OP_ENABLE) ? 1 : 0,
1925 		   (op & SDMA_SENDCTRL_OP_INTENABLE) ? 1 : 0,
1926 		   (op & SDMA_SENDCTRL_OP_HALT) ? 1 : 0,
1927 		   (op & SDMA_SENDCTRL_OP_CLEANUP) ? 1 : 0);
1928 #endif
1929 
1930 	if (op & SDMA_SENDCTRL_OP_ENABLE)
1931 		set_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1932 	else
1933 		clr_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1934 
1935 	if (op & SDMA_SENDCTRL_OP_INTENABLE)
1936 		set_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1937 	else
1938 		clr_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1939 
1940 	if (op & SDMA_SENDCTRL_OP_HALT)
1941 		set_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1942 	else
1943 		clr_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1944 
1945 	spin_lock_irqsave(&sde->senddmactrl_lock, flags);
1946 
1947 	sde->p_senddmactrl |= set_senddmactrl;
1948 	sde->p_senddmactrl &= ~clr_senddmactrl;
1949 
1950 	if (op & SDMA_SENDCTRL_OP_CLEANUP)
1951 		write_sde_csr(sde, SD(CTRL),
1952 			      sde->p_senddmactrl |
1953 			      SD(CTRL_SDMA_CLEANUP_SMASK));
1954 	else
1955 		write_sde_csr(sde, SD(CTRL), sde->p_senddmactrl);
1956 
1957 	spin_unlock_irqrestore(&sde->senddmactrl_lock, flags);
1958 
1959 #ifdef CONFIG_SDMA_VERBOSITY
1960 	sdma_dumpstate(sde);
1961 #endif
1962 }
1963 
1964 static void sdma_setlengen(struct sdma_engine *sde)
1965 {
1966 #ifdef CONFIG_SDMA_VERBOSITY
1967 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1968 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1969 #endif
1970 
1971 	/*
1972 	 * Set SendDmaLenGen and clear-then-set the MSB of the generation
1973 	 * count to enable generation checking and load the internal
1974 	 * generation counter.
1975 	 */
1976 	write_sde_csr(sde, SD(LEN_GEN),
1977 		      (sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT));
1978 	write_sde_csr(sde, SD(LEN_GEN),
1979 		      ((sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT)) |
1980 		      (4ULL << SD(LEN_GEN_GENERATION_SHIFT)));
1981 }
1982 
1983 static inline void sdma_update_tail(struct sdma_engine *sde, u16 tail)
1984 {
1985 	/* Commit writes to memory and advance the tail on the chip */
1986 	smp_wmb(); /* see get_txhead() */
1987 	writeq(tail, sde->tail_csr);
1988 }
1989 
1990 /*
1991  * This is called when changing to state s10_hw_start_up_halt_wait as
1992  * a result of send buffer errors or send DMA descriptor errors.
1993  */
1994 static void sdma_hw_start_up(struct sdma_engine *sde)
1995 {
1996 	u64 reg;
1997 
1998 #ifdef CONFIG_SDMA_VERBOSITY
1999 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2000 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2001 #endif
2002 
2003 	sdma_setlengen(sde);
2004 	sdma_update_tail(sde, 0); /* Set SendDmaTail */
2005 	*sde->head_dma = 0;
2006 
2007 	reg = SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_MASK) <<
2008 	      SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SHIFT);
2009 	write_sde_csr(sde, SD(ENG_ERR_CLEAR), reg);
2010 }
2011 
2012 /*
2013  * set_sdma_integrity
2014  *
2015  * Set the SEND_DMA_CHECK_ENABLE register for send DMA engine 'sde'.
2016  */
2017 static void set_sdma_integrity(struct sdma_engine *sde)
2018 {
2019 	struct hfi1_devdata *dd = sde->dd;
2020 
2021 	write_sde_csr(sde, SD(CHECK_ENABLE),
2022 		      hfi1_pkt_base_sdma_integrity(dd));
2023 }
2024 
2025 static void init_sdma_regs(
2026 	struct sdma_engine *sde,
2027 	u32 credits,
2028 	uint idle_cnt)
2029 {
2030 	u8 opval, opmask;
2031 #ifdef CONFIG_SDMA_VERBOSITY
2032 	struct hfi1_devdata *dd = sde->dd;
2033 
2034 	dd_dev_err(dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2035 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2036 #endif
2037 
2038 	write_sde_csr(sde, SD(BASE_ADDR), sde->descq_phys);
2039 	sdma_setlengen(sde);
2040 	sdma_update_tail(sde, 0); /* Set SendDmaTail */
2041 	write_sde_csr(sde, SD(RELOAD_CNT), idle_cnt);
2042 	write_sde_csr(sde, SD(DESC_CNT), 0);
2043 	write_sde_csr(sde, SD(HEAD_ADDR), sde->head_phys);
2044 	write_sde_csr(sde, SD(MEMORY),
2045 		      ((u64)credits << SD(MEMORY_SDMA_MEMORY_CNT_SHIFT)) |
2046 		      ((u64)(credits * sde->this_idx) <<
2047 		       SD(MEMORY_SDMA_MEMORY_INDEX_SHIFT)));
2048 	write_sde_csr(sde, SD(ENG_ERR_MASK), ~0ull);
2049 	set_sdma_integrity(sde);
2050 	opmask = OPCODE_CHECK_MASK_DISABLED;
2051 	opval = OPCODE_CHECK_VAL_DISABLED;
2052 	write_sde_csr(sde, SD(CHECK_OPCODE),
2053 		      (opmask << SEND_CTXT_CHECK_OPCODE_MASK_SHIFT) |
2054 		      (opval << SEND_CTXT_CHECK_OPCODE_VALUE_SHIFT));
2055 }
2056 
2057 #ifdef CONFIG_SDMA_VERBOSITY
2058 
2059 #define sdma_dumpstate_helper0(reg) do { \
2060 		csr = read_csr(sde->dd, reg); \
2061 		dd_dev_err(sde->dd, "%36s     0x%016llx\n", #reg, csr); \
2062 	} while (0)
2063 
2064 #define sdma_dumpstate_helper(reg) do { \
2065 		csr = read_sde_csr(sde, reg); \
2066 		dd_dev_err(sde->dd, "%36s[%02u] 0x%016llx\n", \
2067 			#reg, sde->this_idx, csr); \
2068 	} while (0)
2069 
2070 #define sdma_dumpstate_helper2(reg) do { \
2071 		csr = read_csr(sde->dd, reg + (8 * i)); \
2072 		dd_dev_err(sde->dd, "%33s_%02u     0x%016llx\n", \
2073 				#reg, i, csr); \
2074 	} while (0)
2075 
2076 void sdma_dumpstate(struct sdma_engine *sde)
2077 {
2078 	u64 csr;
2079 	unsigned i;
2080 
2081 	sdma_dumpstate_helper(SD(CTRL));
2082 	sdma_dumpstate_helper(SD(STATUS));
2083 	sdma_dumpstate_helper0(SD(ERR_STATUS));
2084 	sdma_dumpstate_helper0(SD(ERR_MASK));
2085 	sdma_dumpstate_helper(SD(ENG_ERR_STATUS));
2086 	sdma_dumpstate_helper(SD(ENG_ERR_MASK));
2087 
2088 	for (i = 0; i < CCE_NUM_INT_CSRS; ++i) {
2089 		sdma_dumpstate_helper2(CCE_INT_STATUS);
2090 		sdma_dumpstate_helper2(CCE_INT_MASK);
2091 		sdma_dumpstate_helper2(CCE_INT_BLOCKED);
2092 	}
2093 
2094 	sdma_dumpstate_helper(SD(TAIL));
2095 	sdma_dumpstate_helper(SD(HEAD));
2096 	sdma_dumpstate_helper(SD(PRIORITY_THLD));
2097 	sdma_dumpstate_helper(SD(IDLE_CNT));
2098 	sdma_dumpstate_helper(SD(RELOAD_CNT));
2099 	sdma_dumpstate_helper(SD(DESC_CNT));
2100 	sdma_dumpstate_helper(SD(DESC_FETCHED_CNT));
2101 	sdma_dumpstate_helper(SD(MEMORY));
2102 	sdma_dumpstate_helper0(SD(ENGINES));
2103 	sdma_dumpstate_helper0(SD(MEM_SIZE));
2104 	/* sdma_dumpstate_helper(SEND_EGRESS_SEND_DMA_STATUS);  */
2105 	sdma_dumpstate_helper(SD(BASE_ADDR));
2106 	sdma_dumpstate_helper(SD(LEN_GEN));
2107 	sdma_dumpstate_helper(SD(HEAD_ADDR));
2108 	sdma_dumpstate_helper(SD(CHECK_ENABLE));
2109 	sdma_dumpstate_helper(SD(CHECK_VL));
2110 	sdma_dumpstate_helper(SD(CHECK_JOB_KEY));
2111 	sdma_dumpstate_helper(SD(CHECK_PARTITION_KEY));
2112 	sdma_dumpstate_helper(SD(CHECK_SLID));
2113 	sdma_dumpstate_helper(SD(CHECK_OPCODE));
2114 }
2115 #endif
2116 
2117 static void dump_sdma_state(struct sdma_engine *sde)
2118 {
2119 	struct hw_sdma_desc *descq;
2120 	struct hw_sdma_desc *descqp;
2121 	u64 desc[2];
2122 	u64 addr;
2123 	u8 gen;
2124 	u16 len;
2125 	u16 head, tail, cnt;
2126 
2127 	head = sde->descq_head & sde->sdma_mask;
2128 	tail = sde->descq_tail & sde->sdma_mask;
2129 	cnt = sdma_descq_freecnt(sde);
2130 	descq = sde->descq;
2131 
2132 	dd_dev_err(sde->dd,
2133 		   "SDMA (%u) descq_head: %u descq_tail: %u freecnt: %u FLE %d\n",
2134 		   sde->this_idx, head, tail, cnt,
2135 		   !list_empty(&sde->flushlist));
2136 
2137 	/* print info for each entry in the descriptor queue */
2138 	while (head != tail) {
2139 		char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2140 
2141 		descqp = &sde->descq[head];
2142 		desc[0] = le64_to_cpu(descqp->qw[0]);
2143 		desc[1] = le64_to_cpu(descqp->qw[1]);
2144 		flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2145 		flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2146 				'H' : '-';
2147 		flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2148 		flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2149 		addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2150 			& SDMA_DESC0_PHY_ADDR_MASK;
2151 		gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2152 			& SDMA_DESC1_GENERATION_MASK;
2153 		len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2154 			& SDMA_DESC0_BYTE_COUNT_MASK;
2155 		dd_dev_err(sde->dd,
2156 			   "SDMA sdmadesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2157 			   head, flags, addr, gen, len);
2158 		dd_dev_err(sde->dd,
2159 			   "\tdesc0:0x%016llx desc1 0x%016llx\n",
2160 			   desc[0], desc[1]);
2161 		if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2162 			dd_dev_err(sde->dd,
2163 				   "\taidx: %u amode: %u alen: %u\n",
2164 				   (u8)((desc[1] &
2165 					 SDMA_DESC1_HEADER_INDEX_SMASK) >>
2166 					SDMA_DESC1_HEADER_INDEX_SHIFT),
2167 				   (u8)((desc[1] &
2168 					 SDMA_DESC1_HEADER_MODE_SMASK) >>
2169 					SDMA_DESC1_HEADER_MODE_SHIFT),
2170 				   (u8)((desc[1] &
2171 					 SDMA_DESC1_HEADER_DWS_SMASK) >>
2172 					SDMA_DESC1_HEADER_DWS_SHIFT));
2173 		head++;
2174 		head &= sde->sdma_mask;
2175 	}
2176 }
2177 
2178 #define SDE_FMT \
2179 	"SDE %u CPU %d 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"
2180 /**
2181  * sdma_seqfile_dump_sde() - debugfs dump of sde
2182  * @s: seq file
2183  * @sde: send dma engine to dump
2184  *
2185  * This routine dumps the sde to the indicated seq file.
2186  */
2187 void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *sde)
2188 {
2189 	u16 head, tail;
2190 	struct hw_sdma_desc *descqp;
2191 	u64 desc[2];
2192 	u64 addr;
2193 	u8 gen;
2194 	u16 len;
2195 
2196 	head = sde->descq_head & sde->sdma_mask;
2197 	tail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
2198 	seq_printf(s, SDE_FMT, sde->this_idx,
2199 		   sde->cpu,
2200 		   sdma_state_name(sde->state.current_state),
2201 		   (unsigned long long)read_sde_csr(sde, SD(CTRL)),
2202 		   (unsigned long long)read_sde_csr(sde, SD(STATUS)),
2203 		   (unsigned long long)read_sde_csr(sde, SD(ENG_ERR_STATUS)),
2204 		   (unsigned long long)read_sde_csr(sde, SD(TAIL)), tail,
2205 		   (unsigned long long)read_sde_csr(sde, SD(HEAD)), head,
2206 		   (unsigned long long)le64_to_cpu(*sde->head_dma),
2207 		   (unsigned long long)read_sde_csr(sde, SD(MEMORY)),
2208 		   (unsigned long long)read_sde_csr(sde, SD(LEN_GEN)),
2209 		   (unsigned long long)read_sde_csr(sde, SD(RELOAD_CNT)),
2210 		   (unsigned long long)sde->last_status,
2211 		   (unsigned long long)sde->ahg_bits,
2212 		   sde->tx_tail,
2213 		   sde->tx_head,
2214 		   sde->descq_tail,
2215 		   sde->descq_head,
2216 		   !list_empty(&sde->flushlist),
2217 		   sde->descq_full_count,
2218 		   (unsigned long long)read_sde_csr(sde, SEND_DMA_CHECK_SLID));
2219 
2220 	/* print info for each entry in the descriptor queue */
2221 	while (head != tail) {
2222 		char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2223 
2224 		descqp = &sde->descq[head];
2225 		desc[0] = le64_to_cpu(descqp->qw[0]);
2226 		desc[1] = le64_to_cpu(descqp->qw[1]);
2227 		flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2228 		flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2229 				'H' : '-';
2230 		flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2231 		flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2232 		addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2233 			& SDMA_DESC0_PHY_ADDR_MASK;
2234 		gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2235 			& SDMA_DESC1_GENERATION_MASK;
2236 		len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2237 			& SDMA_DESC0_BYTE_COUNT_MASK;
2238 		seq_printf(s,
2239 			   "\tdesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2240 			   head, flags, addr, gen, len);
2241 		if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2242 			seq_printf(s, "\t\tahgidx: %u ahgmode: %u\n",
2243 				   (u8)((desc[1] &
2244 					 SDMA_DESC1_HEADER_INDEX_SMASK) >>
2245 					SDMA_DESC1_HEADER_INDEX_SHIFT),
2246 				   (u8)((desc[1] &
2247 					 SDMA_DESC1_HEADER_MODE_SMASK) >>
2248 					SDMA_DESC1_HEADER_MODE_SHIFT));
2249 		head = (head + 1) & sde->sdma_mask;
2250 	}
2251 }
2252 
2253 /*
2254  * add the generation number into
2255  * the qw1 and return
2256  */
2257 static inline u64 add_gen(struct sdma_engine *sde, u64 qw1)
2258 {
2259 	u8 generation = (sde->descq_tail >> sde->sdma_shift) & 3;
2260 
2261 	qw1 &= ~SDMA_DESC1_GENERATION_SMASK;
2262 	qw1 |= ((u64)generation & SDMA_DESC1_GENERATION_MASK)
2263 			<< SDMA_DESC1_GENERATION_SHIFT;
2264 	return qw1;
2265 }
2266 
2267 /*
2268  * This routine submits the indicated tx
2269  *
2270  * Space has already been guaranteed and
2271  * tail side of ring is locked.
2272  *
2273  * The hardware tail update is done
2274  * in the caller and that is facilitated
2275  * by returning the new tail.
2276  *
2277  * There is special case logic for ahg
2278  * to not add the generation number for
2279  * up to 2 descriptors that follow the
2280  * first descriptor.
2281  *
2282  */
2283 static inline u16 submit_tx(struct sdma_engine *sde, struct sdma_txreq *tx)
2284 {
2285 	int i;
2286 	u16 tail;
2287 	struct sdma_desc *descp = tx->descp;
2288 	u8 skip = 0, mode = ahg_mode(tx);
2289 
2290 	tail = sde->descq_tail & sde->sdma_mask;
2291 	sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2292 	sde->descq[tail].qw[1] = cpu_to_le64(add_gen(sde, descp->qw[1]));
2293 	trace_hfi1_sdma_descriptor(sde, descp->qw[0], descp->qw[1],
2294 				   tail, &sde->descq[tail]);
2295 	tail = ++sde->descq_tail & sde->sdma_mask;
2296 	descp++;
2297 	if (mode > SDMA_AHG_APPLY_UPDATE1)
2298 		skip = mode >> 1;
2299 	for (i = 1; i < tx->num_desc; i++, descp++) {
2300 		u64 qw1;
2301 
2302 		sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2303 		if (skip) {
2304 			/* edits don't have generation */
2305 			qw1 = descp->qw[1];
2306 			skip--;
2307 		} else {
2308 			/* replace generation with real one for non-edits */
2309 			qw1 = add_gen(sde, descp->qw[1]);
2310 		}
2311 		sde->descq[tail].qw[1] = cpu_to_le64(qw1);
2312 		trace_hfi1_sdma_descriptor(sde, descp->qw[0], qw1,
2313 					   tail, &sde->descq[tail]);
2314 		tail = ++sde->descq_tail & sde->sdma_mask;
2315 	}
2316 	tx->next_descq_idx = tail;
2317 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2318 	tx->sn = sde->tail_sn++;
2319 	trace_hfi1_sdma_in_sn(sde, tx->sn);
2320 	WARN_ON_ONCE(sde->tx_ring[sde->tx_tail & sde->sdma_mask]);
2321 #endif
2322 	sde->tx_ring[sde->tx_tail++ & sde->sdma_mask] = tx;
2323 	sde->desc_avail -= tx->num_desc;
2324 	return tail;
2325 }
2326 
2327 /*
2328  * Check for progress
2329  */
2330 static int sdma_check_progress(
2331 	struct sdma_engine *sde,
2332 	struct iowait *wait,
2333 	struct sdma_txreq *tx)
2334 {
2335 	int ret;
2336 
2337 	sde->desc_avail = sdma_descq_freecnt(sde);
2338 	if (tx->num_desc <= sde->desc_avail)
2339 		return -EAGAIN;
2340 	/* pulse the head_lock */
2341 	if (wait && wait->sleep) {
2342 		unsigned seq;
2343 
2344 		seq = raw_seqcount_begin(
2345 			(const seqcount_t *)&sde->head_lock.seqcount);
2346 		ret = wait->sleep(sde, wait, tx, seq);
2347 		if (ret == -EAGAIN)
2348 			sde->desc_avail = sdma_descq_freecnt(sde);
2349 	} else {
2350 		ret = -EBUSY;
2351 	}
2352 	return ret;
2353 }
2354 
2355 /**
2356  * sdma_send_txreq() - submit a tx req to ring
2357  * @sde: sdma engine to use
2358  * @wait: wait structure to use when full (may be NULL)
2359  * @tx: sdma_txreq to submit
2360  *
2361  * The call submits the tx into the ring.  If a iowait structure is non-NULL
2362  * the packet will be queued to the list in wait.
2363  *
2364  * Return:
2365  * 0 - Success, -EINVAL - sdma_txreq incomplete, -EBUSY - no space in
2366  * ring (wait == NULL)
2367  * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2368  */
2369 int sdma_send_txreq(struct sdma_engine *sde,
2370 		    struct iowait *wait,
2371 		    struct sdma_txreq *tx)
2372 {
2373 	int ret = 0;
2374 	u16 tail;
2375 	unsigned long flags;
2376 
2377 	/* user should have supplied entire packet */
2378 	if (unlikely(tx->tlen))
2379 		return -EINVAL;
2380 	tx->wait = wait;
2381 	spin_lock_irqsave(&sde->tail_lock, flags);
2382 retry:
2383 	if (unlikely(!__sdma_running(sde)))
2384 		goto unlock_noconn;
2385 	if (unlikely(tx->num_desc > sde->desc_avail))
2386 		goto nodesc;
2387 	tail = submit_tx(sde, tx);
2388 	if (wait)
2389 		iowait_sdma_inc(wait);
2390 	sdma_update_tail(sde, tail);
2391 unlock:
2392 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2393 	return ret;
2394 unlock_noconn:
2395 	if (wait)
2396 		iowait_sdma_inc(wait);
2397 	tx->next_descq_idx = 0;
2398 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2399 	tx->sn = sde->tail_sn++;
2400 	trace_hfi1_sdma_in_sn(sde, tx->sn);
2401 #endif
2402 	spin_lock(&sde->flushlist_lock);
2403 	list_add_tail(&tx->list, &sde->flushlist);
2404 	spin_unlock(&sde->flushlist_lock);
2405 	if (wait) {
2406 		wait->tx_count++;
2407 		wait->count += tx->num_desc;
2408 	}
2409 	schedule_work(&sde->flush_worker);
2410 	ret = -ECOMM;
2411 	goto unlock;
2412 nodesc:
2413 	ret = sdma_check_progress(sde, wait, tx);
2414 	if (ret == -EAGAIN) {
2415 		ret = 0;
2416 		goto retry;
2417 	}
2418 	sde->descq_full_count++;
2419 	goto unlock;
2420 }
2421 
2422 /**
2423  * sdma_send_txlist() - submit a list of tx req to ring
2424  * @sde: sdma engine to use
2425  * @wait: wait structure to use when full (may be NULL)
2426  * @tx_list: list of sdma_txreqs to submit
2427  * @count: pointer to a u32 which, after return will contain the total number of
2428  *         sdma_txreqs removed from the tx_list. This will include sdma_txreqs
2429  *         whose SDMA descriptors are submitted to the ring and the sdma_txreqs
2430  *         which are added to SDMA engine flush list if the SDMA engine state is
2431  *         not running.
2432  *
2433  * The call submits the list into the ring.
2434  *
2435  * If the iowait structure is non-NULL and not equal to the iowait list
2436  * the unprocessed part of the list  will be appended to the list in wait.
2437  *
2438  * In all cases, the tx_list will be updated so the head of the tx_list is
2439  * the list of descriptors that have yet to be transmitted.
2440  *
2441  * The intent of this call is to provide a more efficient
2442  * way of submitting multiple packets to SDMA while holding the tail
2443  * side locking.
2444  *
2445  * Return:
2446  * 0 - Success,
2447  * -EINVAL - sdma_txreq incomplete, -EBUSY - no space in ring (wait == NULL)
2448  * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2449  */
2450 int sdma_send_txlist(struct sdma_engine *sde, struct iowait *wait,
2451 		     struct list_head *tx_list, u32 *count_out)
2452 {
2453 	struct sdma_txreq *tx, *tx_next;
2454 	int ret = 0;
2455 	unsigned long flags;
2456 	u16 tail = INVALID_TAIL;
2457 	u32 submit_count = 0, flush_count = 0, total_count;
2458 
2459 	spin_lock_irqsave(&sde->tail_lock, flags);
2460 retry:
2461 	list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2462 		tx->wait = wait;
2463 		if (unlikely(!__sdma_running(sde)))
2464 			goto unlock_noconn;
2465 		if (unlikely(tx->num_desc > sde->desc_avail))
2466 			goto nodesc;
2467 		if (unlikely(tx->tlen)) {
2468 			ret = -EINVAL;
2469 			goto update_tail;
2470 		}
2471 		list_del_init(&tx->list);
2472 		tail = submit_tx(sde, tx);
2473 		submit_count++;
2474 		if (tail != INVALID_TAIL &&
2475 		    (submit_count & SDMA_TAIL_UPDATE_THRESH) == 0) {
2476 			sdma_update_tail(sde, tail);
2477 			tail = INVALID_TAIL;
2478 		}
2479 	}
2480 update_tail:
2481 	total_count = submit_count + flush_count;
2482 	if (wait)
2483 		iowait_sdma_add(wait, total_count);
2484 	if (tail != INVALID_TAIL)
2485 		sdma_update_tail(sde, tail);
2486 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2487 	*count_out = total_count;
2488 	return ret;
2489 unlock_noconn:
2490 	spin_lock(&sde->flushlist_lock);
2491 	list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2492 		tx->wait = wait;
2493 		list_del_init(&tx->list);
2494 		tx->next_descq_idx = 0;
2495 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2496 		tx->sn = sde->tail_sn++;
2497 		trace_hfi1_sdma_in_sn(sde, tx->sn);
2498 #endif
2499 		list_add_tail(&tx->list, &sde->flushlist);
2500 		flush_count++;
2501 		if (wait) {
2502 			wait->tx_count++;
2503 			wait->count += tx->num_desc;
2504 		}
2505 	}
2506 	spin_unlock(&sde->flushlist_lock);
2507 	schedule_work(&sde->flush_worker);
2508 	ret = -ECOMM;
2509 	goto update_tail;
2510 nodesc:
2511 	ret = sdma_check_progress(sde, wait, tx);
2512 	if (ret == -EAGAIN) {
2513 		ret = 0;
2514 		goto retry;
2515 	}
2516 	sde->descq_full_count++;
2517 	goto update_tail;
2518 }
2519 
2520 static void sdma_process_event(struct sdma_engine *sde, enum sdma_events event)
2521 {
2522 	unsigned long flags;
2523 
2524 	spin_lock_irqsave(&sde->tail_lock, flags);
2525 	write_seqlock(&sde->head_lock);
2526 
2527 	__sdma_process_event(sde, event);
2528 
2529 	if (sde->state.current_state == sdma_state_s99_running)
2530 		sdma_desc_avail(sde, sdma_descq_freecnt(sde));
2531 
2532 	write_sequnlock(&sde->head_lock);
2533 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2534 }
2535 
2536 static void __sdma_process_event(struct sdma_engine *sde,
2537 				 enum sdma_events event)
2538 {
2539 	struct sdma_state *ss = &sde->state;
2540 	int need_progress = 0;
2541 
2542 	/* CONFIG SDMA temporary */
2543 #ifdef CONFIG_SDMA_VERBOSITY
2544 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) [%s] %s\n", sde->this_idx,
2545 		   sdma_state_names[ss->current_state],
2546 		   sdma_event_names[event]);
2547 #endif
2548 
2549 	switch (ss->current_state) {
2550 	case sdma_state_s00_hw_down:
2551 		switch (event) {
2552 		case sdma_event_e00_go_hw_down:
2553 			break;
2554 		case sdma_event_e30_go_running:
2555 			/*
2556 			 * If down, but running requested (usually result
2557 			 * of link up, then we need to start up.
2558 			 * This can happen when hw down is requested while
2559 			 * bringing the link up with traffic active on
2560 			 * 7220, e.g.
2561 			 */
2562 			ss->go_s99_running = 1;
2563 			/* fall through and start dma engine */
2564 		case sdma_event_e10_go_hw_start:
2565 			/* This reference means the state machine is started */
2566 			sdma_get(&sde->state);
2567 			sdma_set_state(sde,
2568 				       sdma_state_s10_hw_start_up_halt_wait);
2569 			break;
2570 		case sdma_event_e15_hw_halt_done:
2571 			break;
2572 		case sdma_event_e25_hw_clean_up_done:
2573 			break;
2574 		case sdma_event_e40_sw_cleaned:
2575 			sdma_sw_tear_down(sde);
2576 			break;
2577 		case sdma_event_e50_hw_cleaned:
2578 			break;
2579 		case sdma_event_e60_hw_halted:
2580 			break;
2581 		case sdma_event_e70_go_idle:
2582 			break;
2583 		case sdma_event_e80_hw_freeze:
2584 			break;
2585 		case sdma_event_e81_hw_frozen:
2586 			break;
2587 		case sdma_event_e82_hw_unfreeze:
2588 			break;
2589 		case sdma_event_e85_link_down:
2590 			break;
2591 		case sdma_event_e90_sw_halted:
2592 			break;
2593 		}
2594 		break;
2595 
2596 	case sdma_state_s10_hw_start_up_halt_wait:
2597 		switch (event) {
2598 		case sdma_event_e00_go_hw_down:
2599 			sdma_set_state(sde, sdma_state_s00_hw_down);
2600 			sdma_sw_tear_down(sde);
2601 			break;
2602 		case sdma_event_e10_go_hw_start:
2603 			break;
2604 		case sdma_event_e15_hw_halt_done:
2605 			sdma_set_state(sde,
2606 				       sdma_state_s15_hw_start_up_clean_wait);
2607 			sdma_start_hw_clean_up(sde);
2608 			break;
2609 		case sdma_event_e25_hw_clean_up_done:
2610 			break;
2611 		case sdma_event_e30_go_running:
2612 			ss->go_s99_running = 1;
2613 			break;
2614 		case sdma_event_e40_sw_cleaned:
2615 			break;
2616 		case sdma_event_e50_hw_cleaned:
2617 			break;
2618 		case sdma_event_e60_hw_halted:
2619 			schedule_work(&sde->err_halt_worker);
2620 			break;
2621 		case sdma_event_e70_go_idle:
2622 			ss->go_s99_running = 0;
2623 			break;
2624 		case sdma_event_e80_hw_freeze:
2625 			break;
2626 		case sdma_event_e81_hw_frozen:
2627 			break;
2628 		case sdma_event_e82_hw_unfreeze:
2629 			break;
2630 		case sdma_event_e85_link_down:
2631 			break;
2632 		case sdma_event_e90_sw_halted:
2633 			break;
2634 		}
2635 		break;
2636 
2637 	case sdma_state_s15_hw_start_up_clean_wait:
2638 		switch (event) {
2639 		case sdma_event_e00_go_hw_down:
2640 			sdma_set_state(sde, sdma_state_s00_hw_down);
2641 			sdma_sw_tear_down(sde);
2642 			break;
2643 		case sdma_event_e10_go_hw_start:
2644 			break;
2645 		case sdma_event_e15_hw_halt_done:
2646 			break;
2647 		case sdma_event_e25_hw_clean_up_done:
2648 			sdma_hw_start_up(sde);
2649 			sdma_set_state(sde, ss->go_s99_running ?
2650 				       sdma_state_s99_running :
2651 				       sdma_state_s20_idle);
2652 			break;
2653 		case sdma_event_e30_go_running:
2654 			ss->go_s99_running = 1;
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 			break;
2662 		case sdma_event_e70_go_idle:
2663 			ss->go_s99_running = 0;
2664 			break;
2665 		case sdma_event_e80_hw_freeze:
2666 			break;
2667 		case sdma_event_e81_hw_frozen:
2668 			break;
2669 		case sdma_event_e82_hw_unfreeze:
2670 			break;
2671 		case sdma_event_e85_link_down:
2672 			break;
2673 		case sdma_event_e90_sw_halted:
2674 			break;
2675 		}
2676 		break;
2677 
2678 	case sdma_state_s20_idle:
2679 		switch (event) {
2680 		case sdma_event_e00_go_hw_down:
2681 			sdma_set_state(sde, sdma_state_s00_hw_down);
2682 			sdma_sw_tear_down(sde);
2683 			break;
2684 		case sdma_event_e10_go_hw_start:
2685 			break;
2686 		case sdma_event_e15_hw_halt_done:
2687 			break;
2688 		case sdma_event_e25_hw_clean_up_done:
2689 			break;
2690 		case sdma_event_e30_go_running:
2691 			sdma_set_state(sde, sdma_state_s99_running);
2692 			ss->go_s99_running = 1;
2693 			break;
2694 		case sdma_event_e40_sw_cleaned:
2695 			break;
2696 		case sdma_event_e50_hw_cleaned:
2697 			break;
2698 		case sdma_event_e60_hw_halted:
2699 			sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
2700 			schedule_work(&sde->err_halt_worker);
2701 			break;
2702 		case sdma_event_e70_go_idle:
2703 			break;
2704 		case sdma_event_e85_link_down:
2705 			/* fall through */
2706 		case sdma_event_e80_hw_freeze:
2707 			sdma_set_state(sde, sdma_state_s80_hw_freeze);
2708 			atomic_dec(&sde->dd->sdma_unfreeze_count);
2709 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2710 			break;
2711 		case sdma_event_e81_hw_frozen:
2712 			break;
2713 		case sdma_event_e82_hw_unfreeze:
2714 			break;
2715 		case sdma_event_e90_sw_halted:
2716 			break;
2717 		}
2718 		break;
2719 
2720 	case sdma_state_s30_sw_clean_up_wait:
2721 		switch (event) {
2722 		case sdma_event_e00_go_hw_down:
2723 			sdma_set_state(sde, sdma_state_s00_hw_down);
2724 			break;
2725 		case sdma_event_e10_go_hw_start:
2726 			break;
2727 		case sdma_event_e15_hw_halt_done:
2728 			break;
2729 		case sdma_event_e25_hw_clean_up_done:
2730 			break;
2731 		case sdma_event_e30_go_running:
2732 			ss->go_s99_running = 1;
2733 			break;
2734 		case sdma_event_e40_sw_cleaned:
2735 			sdma_set_state(sde, sdma_state_s40_hw_clean_up_wait);
2736 			sdma_start_hw_clean_up(sde);
2737 			break;
2738 		case sdma_event_e50_hw_cleaned:
2739 			break;
2740 		case sdma_event_e60_hw_halted:
2741 			break;
2742 		case sdma_event_e70_go_idle:
2743 			ss->go_s99_running = 0;
2744 			break;
2745 		case sdma_event_e80_hw_freeze:
2746 			break;
2747 		case sdma_event_e81_hw_frozen:
2748 			break;
2749 		case sdma_event_e82_hw_unfreeze:
2750 			break;
2751 		case sdma_event_e85_link_down:
2752 			ss->go_s99_running = 0;
2753 			break;
2754 		case sdma_event_e90_sw_halted:
2755 			break;
2756 		}
2757 		break;
2758 
2759 	case sdma_state_s40_hw_clean_up_wait:
2760 		switch (event) {
2761 		case sdma_event_e00_go_hw_down:
2762 			sdma_set_state(sde, sdma_state_s00_hw_down);
2763 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2764 			break;
2765 		case sdma_event_e10_go_hw_start:
2766 			break;
2767 		case sdma_event_e15_hw_halt_done:
2768 			break;
2769 		case sdma_event_e25_hw_clean_up_done:
2770 			sdma_hw_start_up(sde);
2771 			sdma_set_state(sde, ss->go_s99_running ?
2772 				       sdma_state_s99_running :
2773 				       sdma_state_s20_idle);
2774 			break;
2775 		case sdma_event_e30_go_running:
2776 			ss->go_s99_running = 1;
2777 			break;
2778 		case sdma_event_e40_sw_cleaned:
2779 			break;
2780 		case sdma_event_e50_hw_cleaned:
2781 			break;
2782 		case sdma_event_e60_hw_halted:
2783 			break;
2784 		case sdma_event_e70_go_idle:
2785 			ss->go_s99_running = 0;
2786 			break;
2787 		case sdma_event_e80_hw_freeze:
2788 			break;
2789 		case sdma_event_e81_hw_frozen:
2790 			break;
2791 		case sdma_event_e82_hw_unfreeze:
2792 			break;
2793 		case sdma_event_e85_link_down:
2794 			ss->go_s99_running = 0;
2795 			break;
2796 		case sdma_event_e90_sw_halted:
2797 			break;
2798 		}
2799 		break;
2800 
2801 	case sdma_state_s50_hw_halt_wait:
2802 		switch (event) {
2803 		case sdma_event_e00_go_hw_down:
2804 			sdma_set_state(sde, sdma_state_s00_hw_down);
2805 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2806 			break;
2807 		case sdma_event_e10_go_hw_start:
2808 			break;
2809 		case sdma_event_e15_hw_halt_done:
2810 			sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2811 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2812 			break;
2813 		case sdma_event_e25_hw_clean_up_done:
2814 			break;
2815 		case sdma_event_e30_go_running:
2816 			ss->go_s99_running = 1;
2817 			break;
2818 		case sdma_event_e40_sw_cleaned:
2819 			break;
2820 		case sdma_event_e50_hw_cleaned:
2821 			break;
2822 		case sdma_event_e60_hw_halted:
2823 			schedule_work(&sde->err_halt_worker);
2824 			break;
2825 		case sdma_event_e70_go_idle:
2826 			ss->go_s99_running = 0;
2827 			break;
2828 		case sdma_event_e80_hw_freeze:
2829 			break;
2830 		case sdma_event_e81_hw_frozen:
2831 			break;
2832 		case sdma_event_e82_hw_unfreeze:
2833 			break;
2834 		case sdma_event_e85_link_down:
2835 			ss->go_s99_running = 0;
2836 			break;
2837 		case sdma_event_e90_sw_halted:
2838 			break;
2839 		}
2840 		break;
2841 
2842 	case sdma_state_s60_idle_halt_wait:
2843 		switch (event) {
2844 		case sdma_event_e00_go_hw_down:
2845 			sdma_set_state(sde, sdma_state_s00_hw_down);
2846 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2847 			break;
2848 		case sdma_event_e10_go_hw_start:
2849 			break;
2850 		case sdma_event_e15_hw_halt_done:
2851 			sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2852 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2853 			break;
2854 		case sdma_event_e25_hw_clean_up_done:
2855 			break;
2856 		case sdma_event_e30_go_running:
2857 			ss->go_s99_running = 1;
2858 			break;
2859 		case sdma_event_e40_sw_cleaned:
2860 			break;
2861 		case sdma_event_e50_hw_cleaned:
2862 			break;
2863 		case sdma_event_e60_hw_halted:
2864 			schedule_work(&sde->err_halt_worker);
2865 			break;
2866 		case sdma_event_e70_go_idle:
2867 			ss->go_s99_running = 0;
2868 			break;
2869 		case sdma_event_e80_hw_freeze:
2870 			break;
2871 		case sdma_event_e81_hw_frozen:
2872 			break;
2873 		case sdma_event_e82_hw_unfreeze:
2874 			break;
2875 		case sdma_event_e85_link_down:
2876 			break;
2877 		case sdma_event_e90_sw_halted:
2878 			break;
2879 		}
2880 		break;
2881 
2882 	case sdma_state_s80_hw_freeze:
2883 		switch (event) {
2884 		case sdma_event_e00_go_hw_down:
2885 			sdma_set_state(sde, sdma_state_s00_hw_down);
2886 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2887 			break;
2888 		case sdma_event_e10_go_hw_start:
2889 			break;
2890 		case sdma_event_e15_hw_halt_done:
2891 			break;
2892 		case sdma_event_e25_hw_clean_up_done:
2893 			break;
2894 		case sdma_event_e30_go_running:
2895 			ss->go_s99_running = 1;
2896 			break;
2897 		case sdma_event_e40_sw_cleaned:
2898 			break;
2899 		case sdma_event_e50_hw_cleaned:
2900 			break;
2901 		case sdma_event_e60_hw_halted:
2902 			break;
2903 		case sdma_event_e70_go_idle:
2904 			ss->go_s99_running = 0;
2905 			break;
2906 		case sdma_event_e80_hw_freeze:
2907 			break;
2908 		case sdma_event_e81_hw_frozen:
2909 			sdma_set_state(sde, sdma_state_s82_freeze_sw_clean);
2910 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2911 			break;
2912 		case sdma_event_e82_hw_unfreeze:
2913 			break;
2914 		case sdma_event_e85_link_down:
2915 			break;
2916 		case sdma_event_e90_sw_halted:
2917 			break;
2918 		}
2919 		break;
2920 
2921 	case sdma_state_s82_freeze_sw_clean:
2922 		switch (event) {
2923 		case sdma_event_e00_go_hw_down:
2924 			sdma_set_state(sde, sdma_state_s00_hw_down);
2925 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2926 			break;
2927 		case sdma_event_e10_go_hw_start:
2928 			break;
2929 		case sdma_event_e15_hw_halt_done:
2930 			break;
2931 		case sdma_event_e25_hw_clean_up_done:
2932 			break;
2933 		case sdma_event_e30_go_running:
2934 			ss->go_s99_running = 1;
2935 			break;
2936 		case sdma_event_e40_sw_cleaned:
2937 			/* notify caller this engine is done cleaning */
2938 			atomic_dec(&sde->dd->sdma_unfreeze_count);
2939 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2940 			break;
2941 		case sdma_event_e50_hw_cleaned:
2942 			break;
2943 		case sdma_event_e60_hw_halted:
2944 			break;
2945 		case sdma_event_e70_go_idle:
2946 			ss->go_s99_running = 0;
2947 			break;
2948 		case sdma_event_e80_hw_freeze:
2949 			break;
2950 		case sdma_event_e81_hw_frozen:
2951 			break;
2952 		case sdma_event_e82_hw_unfreeze:
2953 			sdma_hw_start_up(sde);
2954 			sdma_set_state(sde, ss->go_s99_running ?
2955 				       sdma_state_s99_running :
2956 				       sdma_state_s20_idle);
2957 			break;
2958 		case sdma_event_e85_link_down:
2959 			break;
2960 		case sdma_event_e90_sw_halted:
2961 			break;
2962 		}
2963 		break;
2964 
2965 	case sdma_state_s99_running:
2966 		switch (event) {
2967 		case sdma_event_e00_go_hw_down:
2968 			sdma_set_state(sde, sdma_state_s00_hw_down);
2969 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2970 			break;
2971 		case sdma_event_e10_go_hw_start:
2972 			break;
2973 		case sdma_event_e15_hw_halt_done:
2974 			break;
2975 		case sdma_event_e25_hw_clean_up_done:
2976 			break;
2977 		case sdma_event_e30_go_running:
2978 			break;
2979 		case sdma_event_e40_sw_cleaned:
2980 			break;
2981 		case sdma_event_e50_hw_cleaned:
2982 			break;
2983 		case sdma_event_e60_hw_halted:
2984 			need_progress = 1;
2985 			sdma_err_progress_check_schedule(sde);
2986 		case sdma_event_e90_sw_halted:
2987 			/*
2988 			* SW initiated halt does not perform engines
2989 			* progress check
2990 			*/
2991 			sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
2992 			schedule_work(&sde->err_halt_worker);
2993 			break;
2994 		case sdma_event_e70_go_idle:
2995 			sdma_set_state(sde, sdma_state_s60_idle_halt_wait);
2996 			break;
2997 		case sdma_event_e85_link_down:
2998 			ss->go_s99_running = 0;
2999 			/* fall through */
3000 		case sdma_event_e80_hw_freeze:
3001 			sdma_set_state(sde, sdma_state_s80_hw_freeze);
3002 			atomic_dec(&sde->dd->sdma_unfreeze_count);
3003 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
3004 			break;
3005 		case sdma_event_e81_hw_frozen:
3006 			break;
3007 		case sdma_event_e82_hw_unfreeze:
3008 			break;
3009 		}
3010 		break;
3011 	}
3012 
3013 	ss->last_event = event;
3014 	if (need_progress)
3015 		sdma_make_progress(sde, 0);
3016 }
3017 
3018 /*
3019  * _extend_sdma_tx_descs() - helper to extend txreq
3020  *
3021  * This is called once the initial nominal allocation
3022  * of descriptors in the sdma_txreq is exhausted.
3023  *
3024  * The code will bump the allocation up to the max
3025  * of MAX_DESC (64) descriptors. There doesn't seem
3026  * much point in an interim step. The last descriptor
3027  * is reserved for coalesce buffer in order to support
3028  * cases where input packet has >MAX_DESC iovecs.
3029  *
3030  */
3031 static int _extend_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3032 {
3033 	int i;
3034 
3035 	/* Handle last descriptor */
3036 	if (unlikely((tx->num_desc == (MAX_DESC - 1)))) {
3037 		/* if tlen is 0, it is for padding, release last descriptor */
3038 		if (!tx->tlen) {
3039 			tx->desc_limit = MAX_DESC;
3040 		} else if (!tx->coalesce_buf) {
3041 			/* allocate coalesce buffer with space for padding */
3042 			tx->coalesce_buf = kmalloc(tx->tlen + sizeof(u32),
3043 						   GFP_ATOMIC);
3044 			if (!tx->coalesce_buf)
3045 				goto enomem;
3046 			tx->coalesce_idx = 0;
3047 		}
3048 		return 0;
3049 	}
3050 
3051 	if (unlikely(tx->num_desc == MAX_DESC))
3052 		goto enomem;
3053 
3054 	tx->descp = kmalloc_array(
3055 			MAX_DESC,
3056 			sizeof(struct sdma_desc),
3057 			GFP_ATOMIC);
3058 	if (!tx->descp)
3059 		goto enomem;
3060 
3061 	/* reserve last descriptor for coalescing */
3062 	tx->desc_limit = MAX_DESC - 1;
3063 	/* copy ones already built */
3064 	for (i = 0; i < tx->num_desc; i++)
3065 		tx->descp[i] = tx->descs[i];
3066 	return 0;
3067 enomem:
3068 	__sdma_txclean(dd, tx);
3069 	return -ENOMEM;
3070 }
3071 
3072 /*
3073  * ext_coal_sdma_tx_descs() - extend or coalesce sdma tx descriptors
3074  *
3075  * This is called once the initial nominal allocation of descriptors
3076  * in the sdma_txreq is exhausted.
3077  *
3078  * This function calls _extend_sdma_tx_descs to extend or allocate
3079  * coalesce buffer. If there is a allocated coalesce buffer, it will
3080  * copy the input packet data into the coalesce buffer. It also adds
3081  * coalesce buffer descriptor once when whole packet is received.
3082  *
3083  * Return:
3084  * <0 - error
3085  * 0 - coalescing, don't populate descriptor
3086  * 1 - continue with populating descriptor
3087  */
3088 int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx,
3089 			   int type, void *kvaddr, struct page *page,
3090 			   unsigned long offset, u16 len)
3091 {
3092 	int pad_len, rval;
3093 	dma_addr_t addr;
3094 
3095 	rval = _extend_sdma_tx_descs(dd, tx);
3096 	if (rval) {
3097 		__sdma_txclean(dd, tx);
3098 		return rval;
3099 	}
3100 
3101 	/* If coalesce buffer is allocated, copy data into it */
3102 	if (tx->coalesce_buf) {
3103 		if (type == SDMA_MAP_NONE) {
3104 			__sdma_txclean(dd, tx);
3105 			return -EINVAL;
3106 		}
3107 
3108 		if (type == SDMA_MAP_PAGE) {
3109 			kvaddr = kmap(page);
3110 			kvaddr += offset;
3111 		} else if (WARN_ON(!kvaddr)) {
3112 			__sdma_txclean(dd, tx);
3113 			return -EINVAL;
3114 		}
3115 
3116 		memcpy(tx->coalesce_buf + tx->coalesce_idx, kvaddr, len);
3117 		tx->coalesce_idx += len;
3118 		if (type == SDMA_MAP_PAGE)
3119 			kunmap(page);
3120 
3121 		/* If there is more data, return */
3122 		if (tx->tlen - tx->coalesce_idx)
3123 			return 0;
3124 
3125 		/* Whole packet is received; add any padding */
3126 		pad_len = tx->packet_len & (sizeof(u32) - 1);
3127 		if (pad_len) {
3128 			pad_len = sizeof(u32) - pad_len;
3129 			memset(tx->coalesce_buf + tx->coalesce_idx, 0, pad_len);
3130 			/* padding is taken care of for coalescing case */
3131 			tx->packet_len += pad_len;
3132 			tx->tlen += pad_len;
3133 		}
3134 
3135 		/* dma map the coalesce buffer */
3136 		addr = dma_map_single(&dd->pcidev->dev,
3137 				      tx->coalesce_buf,
3138 				      tx->tlen,
3139 				      DMA_TO_DEVICE);
3140 
3141 		if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) {
3142 			__sdma_txclean(dd, tx);
3143 			return -ENOSPC;
3144 		}
3145 
3146 		/* Add descriptor for coalesce buffer */
3147 		tx->desc_limit = MAX_DESC;
3148 		return _sdma_txadd_daddr(dd, SDMA_MAP_SINGLE, tx,
3149 					 addr, tx->tlen);
3150 	}
3151 
3152 	return 1;
3153 }
3154 
3155 /* Update sdes when the lmc changes */
3156 void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid)
3157 {
3158 	struct sdma_engine *sde;
3159 	int i;
3160 	u64 sreg;
3161 
3162 	sreg = ((mask & SD(CHECK_SLID_MASK_MASK)) <<
3163 		SD(CHECK_SLID_MASK_SHIFT)) |
3164 		(((lid & mask) & SD(CHECK_SLID_VALUE_MASK)) <<
3165 		SD(CHECK_SLID_VALUE_SHIFT));
3166 
3167 	for (i = 0; i < dd->num_sdma; i++) {
3168 		hfi1_cdbg(LINKVERB, "SendDmaEngine[%d].SLID_CHECK = 0x%x",
3169 			  i, (u32)sreg);
3170 		sde = &dd->per_sdma[i];
3171 		write_sde_csr(sde, SD(CHECK_SLID), sreg);
3172 	}
3173 }
3174 
3175 /* tx not dword sized - pad */
3176 int _pad_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3177 {
3178 	int rval = 0;
3179 
3180 	tx->num_desc++;
3181 	if ((unlikely(tx->num_desc == tx->desc_limit))) {
3182 		rval = _extend_sdma_tx_descs(dd, tx);
3183 		if (rval) {
3184 			__sdma_txclean(dd, tx);
3185 			return rval;
3186 		}
3187 	}
3188 	/* finish the one just added */
3189 	make_tx_sdma_desc(
3190 		tx,
3191 		SDMA_MAP_NONE,
3192 		dd->sdma_pad_phys,
3193 		sizeof(u32) - (tx->packet_len & (sizeof(u32) - 1)));
3194 	_sdma_close_tx(dd, tx);
3195 	return rval;
3196 }
3197 
3198 /*
3199  * Add ahg to the sdma_txreq
3200  *
3201  * The logic will consume up to 3
3202  * descriptors at the beginning of
3203  * sdma_txreq.
3204  */
3205 void _sdma_txreq_ahgadd(
3206 	struct sdma_txreq *tx,
3207 	u8 num_ahg,
3208 	u8 ahg_entry,
3209 	u32 *ahg,
3210 	u8 ahg_hlen)
3211 {
3212 	u32 i, shift = 0, desc = 0;
3213 	u8 mode;
3214 
3215 	WARN_ON_ONCE(num_ahg > 9 || (ahg_hlen & 3) || ahg_hlen == 4);
3216 	/* compute mode */
3217 	if (num_ahg == 1)
3218 		mode = SDMA_AHG_APPLY_UPDATE1;
3219 	else if (num_ahg <= 5)
3220 		mode = SDMA_AHG_APPLY_UPDATE2;
3221 	else
3222 		mode = SDMA_AHG_APPLY_UPDATE3;
3223 	tx->num_desc++;
3224 	/* initialize to consumed descriptors to zero */
3225 	switch (mode) {
3226 	case SDMA_AHG_APPLY_UPDATE3:
3227 		tx->num_desc++;
3228 		tx->descs[2].qw[0] = 0;
3229 		tx->descs[2].qw[1] = 0;
3230 		/* FALLTHROUGH */
3231 	case SDMA_AHG_APPLY_UPDATE2:
3232 		tx->num_desc++;
3233 		tx->descs[1].qw[0] = 0;
3234 		tx->descs[1].qw[1] = 0;
3235 		break;
3236 	}
3237 	ahg_hlen >>= 2;
3238 	tx->descs[0].qw[1] |=
3239 		(((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK)
3240 			<< SDMA_DESC1_HEADER_INDEX_SHIFT) |
3241 		(((u64)ahg_hlen & SDMA_DESC1_HEADER_DWS_MASK)
3242 			<< SDMA_DESC1_HEADER_DWS_SHIFT) |
3243 		(((u64)mode & SDMA_DESC1_HEADER_MODE_MASK)
3244 			<< SDMA_DESC1_HEADER_MODE_SHIFT) |
3245 		(((u64)ahg[0] & SDMA_DESC1_HEADER_UPDATE1_MASK)
3246 			<< SDMA_DESC1_HEADER_UPDATE1_SHIFT);
3247 	for (i = 0; i < (num_ahg - 1); i++) {
3248 		if (!shift && !(i & 2))
3249 			desc++;
3250 		tx->descs[desc].qw[!!(i & 2)] |=
3251 			(((u64)ahg[i + 1])
3252 				<< shift);
3253 		shift = (shift + 32) & 63;
3254 	}
3255 }
3256 
3257 /**
3258  * sdma_ahg_alloc - allocate an AHG entry
3259  * @sde: engine to allocate from
3260  *
3261  * Return:
3262  * 0-31 when successful, -EOPNOTSUPP if AHG is not enabled,
3263  * -ENOSPC if an entry is not available
3264  */
3265 int sdma_ahg_alloc(struct sdma_engine *sde)
3266 {
3267 	int nr;
3268 	int oldbit;
3269 
3270 	if (!sde) {
3271 		trace_hfi1_ahg_allocate(sde, -EINVAL);
3272 		return -EINVAL;
3273 	}
3274 	while (1) {
3275 		nr = ffz(ACCESS_ONCE(sde->ahg_bits));
3276 		if (nr > 31) {
3277 			trace_hfi1_ahg_allocate(sde, -ENOSPC);
3278 			return -ENOSPC;
3279 		}
3280 		oldbit = test_and_set_bit(nr, &sde->ahg_bits);
3281 		if (!oldbit)
3282 			break;
3283 		cpu_relax();
3284 	}
3285 	trace_hfi1_ahg_allocate(sde, nr);
3286 	return nr;
3287 }
3288 
3289 /**
3290  * sdma_ahg_free - free an AHG entry
3291  * @sde: engine to return AHG entry
3292  * @ahg_index: index to free
3293  *
3294  * This routine frees the indicate AHG entry.
3295  */
3296 void sdma_ahg_free(struct sdma_engine *sde, int ahg_index)
3297 {
3298 	if (!sde)
3299 		return;
3300 	trace_hfi1_ahg_deallocate(sde, ahg_index);
3301 	if (ahg_index < 0 || ahg_index > 31)
3302 		return;
3303 	clear_bit(ahg_index, &sde->ahg_bits);
3304 }
3305 
3306 /*
3307  * SPC freeze handling for SDMA engines.  Called when the driver knows
3308  * the SPC is going into a freeze but before the freeze is fully
3309  * settled.  Generally an error interrupt.
3310  *
3311  * This event will pull the engine out of running so no more entries can be
3312  * added to the engine's queue.
3313  */
3314 void sdma_freeze_notify(struct hfi1_devdata *dd, int link_down)
3315 {
3316 	int i;
3317 	enum sdma_events event = link_down ? sdma_event_e85_link_down :
3318 					     sdma_event_e80_hw_freeze;
3319 
3320 	/* set up the wait but do not wait here */
3321 	atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3322 
3323 	/* tell all engines to stop running and wait */
3324 	for (i = 0; i < dd->num_sdma; i++)
3325 		sdma_process_event(&dd->per_sdma[i], event);
3326 
3327 	/* sdma_freeze() will wait for all engines to have stopped */
3328 }
3329 
3330 /*
3331  * SPC freeze handling for SDMA engines.  Called when the driver knows
3332  * the SPC is fully frozen.
3333  */
3334 void sdma_freeze(struct hfi1_devdata *dd)
3335 {
3336 	int i;
3337 	int ret;
3338 
3339 	/*
3340 	 * Make sure all engines have moved out of the running state before
3341 	 * continuing.
3342 	 */
3343 	ret = wait_event_interruptible(dd->sdma_unfreeze_wq,
3344 				       atomic_read(&dd->sdma_unfreeze_count) <=
3345 				       0);
3346 	/* interrupted or count is negative, then unloading - just exit */
3347 	if (ret || atomic_read(&dd->sdma_unfreeze_count) < 0)
3348 		return;
3349 
3350 	/* set up the count for the next wait */
3351 	atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3352 
3353 	/* tell all engines that the SPC is frozen, they can start cleaning */
3354 	for (i = 0; i < dd->num_sdma; i++)
3355 		sdma_process_event(&dd->per_sdma[i], sdma_event_e81_hw_frozen);
3356 
3357 	/*
3358 	 * Wait for everyone to finish software clean before exiting.  The
3359 	 * software clean will read engine CSRs, so must be completed before
3360 	 * the next step, which will clear the engine CSRs.
3361 	 */
3362 	(void)wait_event_interruptible(dd->sdma_unfreeze_wq,
3363 				atomic_read(&dd->sdma_unfreeze_count) <= 0);
3364 	/* no need to check results - done no matter what */
3365 }
3366 
3367 /*
3368  * SPC freeze handling for the SDMA engines.  Called after the SPC is unfrozen.
3369  *
3370  * The SPC freeze acts like a SDMA halt and a hardware clean combined.  All
3371  * that is left is a software clean.  We could do it after the SPC is fully
3372  * frozen, but then we'd have to add another state to wait for the unfreeze.
3373  * Instead, just defer the software clean until the unfreeze step.
3374  */
3375 void sdma_unfreeze(struct hfi1_devdata *dd)
3376 {
3377 	int i;
3378 
3379 	/* tell all engines start freeze clean up */
3380 	for (i = 0; i < dd->num_sdma; i++)
3381 		sdma_process_event(&dd->per_sdma[i],
3382 				   sdma_event_e82_hw_unfreeze);
3383 }
3384 
3385 /**
3386  * _sdma_engine_progress_schedule() - schedule progress on engine
3387  * @sde: sdma_engine to schedule progress
3388  *
3389  */
3390 void _sdma_engine_progress_schedule(
3391 	struct sdma_engine *sde)
3392 {
3393 	trace_hfi1_sdma_engine_progress(sde, sde->progress_mask);
3394 	/* assume we have selected a good cpu */
3395 	write_csr(sde->dd,
3396 		  CCE_INT_FORCE + (8 * (IS_SDMA_START / 64)),
3397 		  sde->progress_mask);
3398 }
3399