xref: /linux/drivers/infiniband/hw/hfi1/sdma.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
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 		if (vl >= ARRAY_SIZE(rht_node->map)) {
966 			ret = -EINVAL;
967 			goto out;
968 		}
969 
970 		rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu,
971 						  sdma_rht_params);
972 		if (!rht_node) {
973 			rht_node = kzalloc(sizeof(*rht_node), GFP_KERNEL);
974 			if (!rht_node) {
975 				ret = -ENOMEM;
976 				goto out;
977 			}
978 
979 			rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
980 			if (!rht_node->map[vl]) {
981 				kfree(rht_node);
982 				ret = -ENOMEM;
983 				goto out;
984 			}
985 			rht_node->cpu_id = cpu;
986 			rht_node->map[vl]->mask = 0;
987 			rht_node->map[vl]->ctr = 1;
988 			rht_node->map[vl]->sde[0] = sde;
989 
990 			ret = rhashtable_insert_fast(dd->sdma_rht,
991 						     &rht_node->node,
992 						     sdma_rht_params);
993 			if (ret) {
994 				kfree(rht_node->map[vl]);
995 				kfree(rht_node);
996 				dd_dev_err(sde->dd, "Failed to set process to sde affinity for cpu %lu\n",
997 					   cpu);
998 				goto out;
999 			}
1000 
1001 		} else {
1002 			int ctr, pow;
1003 
1004 			/* Add new user mappings */
1005 			if (!rht_node->map[vl])
1006 				rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
1007 
1008 			if (!rht_node->map[vl]) {
1009 				ret = -ENOMEM;
1010 				goto out;
1011 			}
1012 
1013 			rht_node->map[vl]->ctr++;
1014 			ctr = rht_node->map[vl]->ctr;
1015 			rht_node->map[vl]->sde[ctr - 1] = sde;
1016 			pow = roundup_pow_of_two(ctr);
1017 			rht_node->map[vl]->mask = pow - 1;
1018 
1019 			/* Populate the sde map table */
1020 			sdma_populate_sde_map(rht_node->map[vl]);
1021 		}
1022 		cpumask_set_cpu(cpu, new_mask);
1023 	}
1024 
1025 	/* Clean up old mappings */
1026 	for_each_cpu(cpu, cpu_online_mask) {
1027 		struct sdma_rht_node *rht_node;
1028 
1029 		/* Don't cleanup sdes that are set in the new mask */
1030 		if (cpumask_test_cpu(cpu, mask))
1031 			continue;
1032 
1033 		rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu,
1034 						  sdma_rht_params);
1035 		if (rht_node) {
1036 			bool empty = true;
1037 			int i;
1038 
1039 			/* Remove mappings for old sde */
1040 			for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1041 				if (rht_node->map[i])
1042 					sdma_cleanup_sde_map(rht_node->map[i],
1043 							     sde);
1044 
1045 			/* Free empty hash table entries */
1046 			for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1047 				if (!rht_node->map[i])
1048 					continue;
1049 
1050 				if (rht_node->map[i]->ctr) {
1051 					empty = false;
1052 					break;
1053 				}
1054 			}
1055 
1056 			if (empty) {
1057 				ret = rhashtable_remove_fast(dd->sdma_rht,
1058 							     &rht_node->node,
1059 							     sdma_rht_params);
1060 				WARN_ON(ret);
1061 
1062 				for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1063 					kfree(rht_node->map[i]);
1064 
1065 				kfree(rht_node);
1066 			}
1067 		}
1068 	}
1069 
1070 	cpumask_copy(&sde->cpu_mask, new_mask);
1071 out:
1072 	mutex_unlock(&process_to_sde_mutex);
1073 out_free:
1074 	free_cpumask_var(mask);
1075 	free_cpumask_var(new_mask);
1076 	return ret ? : strnlen(buf, PAGE_SIZE);
1077 }
1078 
1079 ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf)
1080 {
1081 	mutex_lock(&process_to_sde_mutex);
1082 	if (cpumask_empty(&sde->cpu_mask))
1083 		snprintf(buf, PAGE_SIZE, "%s\n", "empty");
1084 	else
1085 		cpumap_print_to_pagebuf(true, buf, &sde->cpu_mask);
1086 	mutex_unlock(&process_to_sde_mutex);
1087 	return strnlen(buf, PAGE_SIZE);
1088 }
1089 
1090 static void sdma_rht_free(void *ptr, void *arg)
1091 {
1092 	struct sdma_rht_node *rht_node = ptr;
1093 	int i;
1094 
1095 	for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1096 		kfree(rht_node->map[i]);
1097 
1098 	kfree(rht_node);
1099 }
1100 
1101 /**
1102  * sdma_seqfile_dump_cpu_list() - debugfs dump the cpu to sdma mappings
1103  * @s: seq file
1104  * @dd: hfi1_devdata
1105  * @cpuid: cpu id
1106  *
1107  * This routine dumps the process to sde mappings per cpu
1108  */
1109 void sdma_seqfile_dump_cpu_list(struct seq_file *s,
1110 				struct hfi1_devdata *dd,
1111 				unsigned long cpuid)
1112 {
1113 	struct sdma_rht_node *rht_node;
1114 	int i, j;
1115 
1116 	rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpuid,
1117 					  sdma_rht_params);
1118 	if (!rht_node)
1119 		return;
1120 
1121 	seq_printf(s, "cpu%3lu: ", cpuid);
1122 	for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1123 		if (!rht_node->map[i] || !rht_node->map[i]->ctr)
1124 			continue;
1125 
1126 		seq_printf(s, " vl%d: [", i);
1127 
1128 		for (j = 0; j < rht_node->map[i]->ctr; j++) {
1129 			if (!rht_node->map[i]->sde[j])
1130 				continue;
1131 
1132 			if (j > 0)
1133 				seq_puts(s, ",");
1134 
1135 			seq_printf(s, " sdma%2d",
1136 				   rht_node->map[i]->sde[j]->this_idx);
1137 		}
1138 		seq_puts(s, " ]");
1139 	}
1140 
1141 	seq_puts(s, "\n");
1142 }
1143 
1144 /*
1145  * Free the indicated map struct
1146  */
1147 static void sdma_map_free(struct sdma_vl_map *m)
1148 {
1149 	int i;
1150 
1151 	for (i = 0; m && i < m->actual_vls; i++)
1152 		kfree(m->map[i]);
1153 	kfree(m);
1154 }
1155 
1156 /*
1157  * Handle RCU callback
1158  */
1159 static void sdma_map_rcu_callback(struct rcu_head *list)
1160 {
1161 	struct sdma_vl_map *m = container_of(list, struct sdma_vl_map, list);
1162 
1163 	sdma_map_free(m);
1164 }
1165 
1166 /**
1167  * sdma_map_init - called when # vls change
1168  * @dd: hfi1_devdata
1169  * @port: port number
1170  * @num_vls: number of vls
1171  * @vl_engines: per vl engine mapping (optional)
1172  *
1173  * This routine changes the mapping based on the number of vls.
1174  *
1175  * vl_engines is used to specify a non-uniform vl/engine loading. NULL
1176  * implies auto computing the loading and giving each VLs a uniform
1177  * distribution of engines per VL.
1178  *
1179  * The auto algorithm computes the sde_per_vl and the number of extra
1180  * engines.  Any extra engines are added from the last VL on down.
1181  *
1182  * rcu locking is used here to control access to the mapping fields.
1183  *
1184  * If either the num_vls or num_sdma are non-power of 2, the array sizes
1185  * in the struct sdma_vl_map and the struct sdma_map_elem are rounded
1186  * up to the next highest power of 2 and the first entry is reused
1187  * in a round robin fashion.
1188  *
1189  * If an error occurs the map change is not done and the mapping is
1190  * not changed.
1191  *
1192  */
1193 int sdma_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_engines)
1194 {
1195 	int i, j;
1196 	int extra, sde_per_vl;
1197 	int engine = 0;
1198 	u8 lvl_engines[OPA_MAX_VLS];
1199 	struct sdma_vl_map *oldmap, *newmap;
1200 
1201 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
1202 		return 0;
1203 
1204 	if (!vl_engines) {
1205 		/* truncate divide */
1206 		sde_per_vl = dd->num_sdma / num_vls;
1207 		/* extras */
1208 		extra = dd->num_sdma % num_vls;
1209 		vl_engines = lvl_engines;
1210 		/* add extras from last vl down */
1211 		for (i = num_vls - 1; i >= 0; i--, extra--)
1212 			vl_engines[i] = sde_per_vl + (extra > 0 ? 1 : 0);
1213 	}
1214 	/* build new map */
1215 	newmap = kzalloc(
1216 		sizeof(struct sdma_vl_map) +
1217 			roundup_pow_of_two(num_vls) *
1218 			sizeof(struct sdma_map_elem *),
1219 		GFP_KERNEL);
1220 	if (!newmap)
1221 		goto bail;
1222 	newmap->actual_vls = num_vls;
1223 	newmap->vls = roundup_pow_of_two(num_vls);
1224 	newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1225 	/* initialize back-map */
1226 	for (i = 0; i < TXE_NUM_SDMA_ENGINES; i++)
1227 		newmap->engine_to_vl[i] = -1;
1228 	for (i = 0; i < newmap->vls; i++) {
1229 		/* save for wrap around */
1230 		int first_engine = engine;
1231 
1232 		if (i < newmap->actual_vls) {
1233 			int sz = roundup_pow_of_two(vl_engines[i]);
1234 
1235 			/* only allocate once */
1236 			newmap->map[i] = kzalloc(
1237 				sizeof(struct sdma_map_elem) +
1238 					sz * sizeof(struct sdma_engine *),
1239 				GFP_KERNEL);
1240 			if (!newmap->map[i])
1241 				goto bail;
1242 			newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1243 			/* assign engines */
1244 			for (j = 0; j < sz; j++) {
1245 				newmap->map[i]->sde[j] =
1246 					&dd->per_sdma[engine];
1247 				if (++engine >= first_engine + vl_engines[i])
1248 					/* wrap back to first engine */
1249 					engine = first_engine;
1250 			}
1251 			/* assign back-map */
1252 			for (j = 0; j < vl_engines[i]; j++)
1253 				newmap->engine_to_vl[first_engine + j] = i;
1254 		} else {
1255 			/* just re-use entry without allocating */
1256 			newmap->map[i] = newmap->map[i % num_vls];
1257 		}
1258 		engine = first_engine + vl_engines[i];
1259 	}
1260 	/* newmap in hand, save old map */
1261 	spin_lock_irq(&dd->sde_map_lock);
1262 	oldmap = rcu_dereference_protected(dd->sdma_map,
1263 					   lockdep_is_held(&dd->sde_map_lock));
1264 
1265 	/* publish newmap */
1266 	rcu_assign_pointer(dd->sdma_map, newmap);
1267 
1268 	spin_unlock_irq(&dd->sde_map_lock);
1269 	/* success, free any old map after grace period */
1270 	if (oldmap)
1271 		call_rcu(&oldmap->list, sdma_map_rcu_callback);
1272 	return 0;
1273 bail:
1274 	/* free any partial allocation */
1275 	sdma_map_free(newmap);
1276 	return -ENOMEM;
1277 }
1278 
1279 /*
1280  * Clean up allocated memory.
1281  *
1282  * This routine is can be called regardless of the success of sdma_init()
1283  *
1284  */
1285 static void sdma_clean(struct hfi1_devdata *dd, size_t num_engines)
1286 {
1287 	size_t i;
1288 	struct sdma_engine *sde;
1289 
1290 	if (dd->sdma_pad_dma) {
1291 		dma_free_coherent(&dd->pcidev->dev, 4,
1292 				  (void *)dd->sdma_pad_dma,
1293 				  dd->sdma_pad_phys);
1294 		dd->sdma_pad_dma = NULL;
1295 		dd->sdma_pad_phys = 0;
1296 	}
1297 	if (dd->sdma_heads_dma) {
1298 		dma_free_coherent(&dd->pcidev->dev, dd->sdma_heads_size,
1299 				  (void *)dd->sdma_heads_dma,
1300 				  dd->sdma_heads_phys);
1301 		dd->sdma_heads_dma = NULL;
1302 		dd->sdma_heads_phys = 0;
1303 	}
1304 	for (i = 0; dd->per_sdma && i < num_engines; ++i) {
1305 		sde = &dd->per_sdma[i];
1306 
1307 		sde->head_dma = NULL;
1308 		sde->head_phys = 0;
1309 
1310 		if (sde->descq) {
1311 			dma_free_coherent(
1312 				&dd->pcidev->dev,
1313 				sde->descq_cnt * sizeof(u64[2]),
1314 				sde->descq,
1315 				sde->descq_phys
1316 			);
1317 			sde->descq = NULL;
1318 			sde->descq_phys = 0;
1319 		}
1320 		kvfree(sde->tx_ring);
1321 		sde->tx_ring = NULL;
1322 	}
1323 	spin_lock_irq(&dd->sde_map_lock);
1324 	sdma_map_free(rcu_access_pointer(dd->sdma_map));
1325 	RCU_INIT_POINTER(dd->sdma_map, NULL);
1326 	spin_unlock_irq(&dd->sde_map_lock);
1327 	synchronize_rcu();
1328 	kfree(dd->per_sdma);
1329 	dd->per_sdma = NULL;
1330 
1331 	if (dd->sdma_rht) {
1332 		rhashtable_free_and_destroy(dd->sdma_rht, sdma_rht_free, NULL);
1333 		kfree(dd->sdma_rht);
1334 		dd->sdma_rht = NULL;
1335 	}
1336 }
1337 
1338 /**
1339  * sdma_init() - called when device probed
1340  * @dd: hfi1_devdata
1341  * @port: port number (currently only zero)
1342  *
1343  * sdma_init initializes the specified number of engines.
1344  *
1345  * The code initializes each sde, its csrs.  Interrupts
1346  * are not required to be enabled.
1347  *
1348  * Returns:
1349  * 0 - success, -errno on failure
1350  */
1351 int sdma_init(struct hfi1_devdata *dd, u8 port)
1352 {
1353 	unsigned this_idx;
1354 	struct sdma_engine *sde;
1355 	struct rhashtable *tmp_sdma_rht;
1356 	u16 descq_cnt;
1357 	void *curr_head;
1358 	struct hfi1_pportdata *ppd = dd->pport + port;
1359 	u32 per_sdma_credits;
1360 	uint idle_cnt = sdma_idle_cnt;
1361 	size_t num_engines = dd->chip_sdma_engines;
1362 	int ret = -ENOMEM;
1363 
1364 	if (!HFI1_CAP_IS_KSET(SDMA)) {
1365 		HFI1_CAP_CLEAR(SDMA_AHG);
1366 		return 0;
1367 	}
1368 	if (mod_num_sdma &&
1369 	    /* can't exceed chip support */
1370 	    mod_num_sdma <= dd->chip_sdma_engines &&
1371 	    /* count must be >= vls */
1372 	    mod_num_sdma >= num_vls)
1373 		num_engines = mod_num_sdma;
1374 
1375 	dd_dev_info(dd, "SDMA mod_num_sdma: %u\n", mod_num_sdma);
1376 	dd_dev_info(dd, "SDMA chip_sdma_engines: %u\n", dd->chip_sdma_engines);
1377 	dd_dev_info(dd, "SDMA chip_sdma_mem_size: %u\n",
1378 		    dd->chip_sdma_mem_size);
1379 
1380 	per_sdma_credits =
1381 		dd->chip_sdma_mem_size / (num_engines * SDMA_BLOCK_SIZE);
1382 
1383 	/* set up freeze waitqueue */
1384 	init_waitqueue_head(&dd->sdma_unfreeze_wq);
1385 	atomic_set(&dd->sdma_unfreeze_count, 0);
1386 
1387 	descq_cnt = sdma_get_descq_cnt();
1388 	dd_dev_info(dd, "SDMA engines %zu descq_cnt %u\n",
1389 		    num_engines, descq_cnt);
1390 
1391 	/* alloc memory for array of send engines */
1392 	dd->per_sdma = kcalloc(num_engines, sizeof(*dd->per_sdma), GFP_KERNEL);
1393 	if (!dd->per_sdma)
1394 		return ret;
1395 
1396 	idle_cnt = ns_to_cclock(dd, idle_cnt);
1397 	if (!sdma_desct_intr)
1398 		sdma_desct_intr = SDMA_DESC_INTR;
1399 
1400 	/* Allocate memory for SendDMA descriptor FIFOs */
1401 	for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1402 		sde = &dd->per_sdma[this_idx];
1403 		sde->dd = dd;
1404 		sde->ppd = ppd;
1405 		sde->this_idx = this_idx;
1406 		sde->descq_cnt = descq_cnt;
1407 		sde->desc_avail = sdma_descq_freecnt(sde);
1408 		sde->sdma_shift = ilog2(descq_cnt);
1409 		sde->sdma_mask = (1 << sde->sdma_shift) - 1;
1410 
1411 		/* Create a mask specifically for each interrupt source */
1412 		sde->int_mask = (u64)1 << (0 * TXE_NUM_SDMA_ENGINES +
1413 					   this_idx);
1414 		sde->progress_mask = (u64)1 << (1 * TXE_NUM_SDMA_ENGINES +
1415 						this_idx);
1416 		sde->idle_mask = (u64)1 << (2 * TXE_NUM_SDMA_ENGINES +
1417 					    this_idx);
1418 		/* Create a combined mask to cover all 3 interrupt sources */
1419 		sde->imask = sde->int_mask | sde->progress_mask |
1420 			     sde->idle_mask;
1421 
1422 		spin_lock_init(&sde->tail_lock);
1423 		seqlock_init(&sde->head_lock);
1424 		spin_lock_init(&sde->senddmactrl_lock);
1425 		spin_lock_init(&sde->flushlist_lock);
1426 		/* insure there is always a zero bit */
1427 		sde->ahg_bits = 0xfffffffe00000000ULL;
1428 
1429 		sdma_set_state(sde, sdma_state_s00_hw_down);
1430 
1431 		/* set up reference counting */
1432 		kref_init(&sde->state.kref);
1433 		init_completion(&sde->state.comp);
1434 
1435 		INIT_LIST_HEAD(&sde->flushlist);
1436 		INIT_LIST_HEAD(&sde->dmawait);
1437 
1438 		sde->tail_csr =
1439 			get_kctxt_csr_addr(dd, this_idx, SD(TAIL));
1440 
1441 		if (idle_cnt)
1442 			dd->default_desc1 =
1443 				SDMA_DESC1_HEAD_TO_HOST_FLAG;
1444 		else
1445 			dd->default_desc1 =
1446 				SDMA_DESC1_INT_REQ_FLAG;
1447 
1448 		tasklet_init(&sde->sdma_hw_clean_up_task, sdma_hw_clean_up_task,
1449 			     (unsigned long)sde);
1450 
1451 		tasklet_init(&sde->sdma_sw_clean_up_task, sdma_sw_clean_up_task,
1452 			     (unsigned long)sde);
1453 		INIT_WORK(&sde->err_halt_worker, sdma_err_halt_wait);
1454 		INIT_WORK(&sde->flush_worker, sdma_field_flush);
1455 
1456 		sde->progress_check_head = 0;
1457 
1458 		setup_timer(&sde->err_progress_check_timer,
1459 			    sdma_err_progress_check, (unsigned long)sde);
1460 
1461 		sde->descq = dma_zalloc_coherent(
1462 			&dd->pcidev->dev,
1463 			descq_cnt * sizeof(u64[2]),
1464 			&sde->descq_phys,
1465 			GFP_KERNEL
1466 		);
1467 		if (!sde->descq)
1468 			goto bail;
1469 		sde->tx_ring =
1470 			kcalloc(descq_cnt, sizeof(struct sdma_txreq *),
1471 				GFP_KERNEL);
1472 		if (!sde->tx_ring)
1473 			sde->tx_ring =
1474 				vzalloc(
1475 					sizeof(struct sdma_txreq *) *
1476 					descq_cnt);
1477 		if (!sde->tx_ring)
1478 			goto bail;
1479 	}
1480 
1481 	dd->sdma_heads_size = L1_CACHE_BYTES * num_engines;
1482 	/* Allocate memory for DMA of head registers to memory */
1483 	dd->sdma_heads_dma = dma_zalloc_coherent(
1484 		&dd->pcidev->dev,
1485 		dd->sdma_heads_size,
1486 		&dd->sdma_heads_phys,
1487 		GFP_KERNEL
1488 	);
1489 	if (!dd->sdma_heads_dma) {
1490 		dd_dev_err(dd, "failed to allocate SendDMA head memory\n");
1491 		goto bail;
1492 	}
1493 
1494 	/* Allocate memory for pad */
1495 	dd->sdma_pad_dma = dma_zalloc_coherent(
1496 		&dd->pcidev->dev,
1497 		sizeof(u32),
1498 		&dd->sdma_pad_phys,
1499 		GFP_KERNEL
1500 	);
1501 	if (!dd->sdma_pad_dma) {
1502 		dd_dev_err(dd, "failed to allocate SendDMA pad memory\n");
1503 		goto bail;
1504 	}
1505 
1506 	/* assign each engine to different cacheline and init registers */
1507 	curr_head = (void *)dd->sdma_heads_dma;
1508 	for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1509 		unsigned long phys_offset;
1510 
1511 		sde = &dd->per_sdma[this_idx];
1512 
1513 		sde->head_dma = curr_head;
1514 		curr_head += L1_CACHE_BYTES;
1515 		phys_offset = (unsigned long)sde->head_dma -
1516 			      (unsigned long)dd->sdma_heads_dma;
1517 		sde->head_phys = dd->sdma_heads_phys + phys_offset;
1518 		init_sdma_regs(sde, per_sdma_credits, idle_cnt);
1519 	}
1520 	dd->flags |= HFI1_HAS_SEND_DMA;
1521 	dd->flags |= idle_cnt ? HFI1_HAS_SDMA_TIMEOUT : 0;
1522 	dd->num_sdma = num_engines;
1523 	ret = sdma_map_init(dd, port, ppd->vls_operational, NULL);
1524 	if (ret < 0)
1525 		goto bail;
1526 
1527 	tmp_sdma_rht = kzalloc(sizeof(*tmp_sdma_rht), GFP_KERNEL);
1528 	if (!tmp_sdma_rht) {
1529 		ret = -ENOMEM;
1530 		goto bail;
1531 	}
1532 
1533 	ret = rhashtable_init(tmp_sdma_rht, &sdma_rht_params);
1534 	if (ret < 0)
1535 		goto bail;
1536 	dd->sdma_rht = tmp_sdma_rht;
1537 
1538 	dd_dev_info(dd, "SDMA num_sdma: %u\n", dd->num_sdma);
1539 	return 0;
1540 
1541 bail:
1542 	sdma_clean(dd, num_engines);
1543 	return ret;
1544 }
1545 
1546 /**
1547  * sdma_all_running() - called when the link goes up
1548  * @dd: hfi1_devdata
1549  *
1550  * This routine moves all engines to the running state.
1551  */
1552 void sdma_all_running(struct hfi1_devdata *dd)
1553 {
1554 	struct sdma_engine *sde;
1555 	unsigned int i;
1556 
1557 	/* move all engines to running */
1558 	for (i = 0; i < dd->num_sdma; ++i) {
1559 		sde = &dd->per_sdma[i];
1560 		sdma_process_event(sde, sdma_event_e30_go_running);
1561 	}
1562 }
1563 
1564 /**
1565  * sdma_all_idle() - called when the link goes down
1566  * @dd: hfi1_devdata
1567  *
1568  * This routine moves all engines to the idle state.
1569  */
1570 void sdma_all_idle(struct hfi1_devdata *dd)
1571 {
1572 	struct sdma_engine *sde;
1573 	unsigned int i;
1574 
1575 	/* idle all engines */
1576 	for (i = 0; i < dd->num_sdma; ++i) {
1577 		sde = &dd->per_sdma[i];
1578 		sdma_process_event(sde, sdma_event_e70_go_idle);
1579 	}
1580 }
1581 
1582 /**
1583  * sdma_start() - called to kick off state processing for all engines
1584  * @dd: hfi1_devdata
1585  *
1586  * This routine is for kicking off the state processing for all required
1587  * sdma engines.  Interrupts need to be working at this point.
1588  *
1589  */
1590 void sdma_start(struct hfi1_devdata *dd)
1591 {
1592 	unsigned i;
1593 	struct sdma_engine *sde;
1594 
1595 	/* kick off the engines state processing */
1596 	for (i = 0; i < dd->num_sdma; ++i) {
1597 		sde = &dd->per_sdma[i];
1598 		sdma_process_event(sde, sdma_event_e10_go_hw_start);
1599 	}
1600 }
1601 
1602 /**
1603  * sdma_exit() - used when module is removed
1604  * @dd: hfi1_devdata
1605  */
1606 void sdma_exit(struct hfi1_devdata *dd)
1607 {
1608 	unsigned this_idx;
1609 	struct sdma_engine *sde;
1610 
1611 	for (this_idx = 0; dd->per_sdma && this_idx < dd->num_sdma;
1612 			++this_idx) {
1613 		sde = &dd->per_sdma[this_idx];
1614 		if (!list_empty(&sde->dmawait))
1615 			dd_dev_err(dd, "sde %u: dmawait list not empty!\n",
1616 				   sde->this_idx);
1617 		sdma_process_event(sde, sdma_event_e00_go_hw_down);
1618 
1619 		del_timer_sync(&sde->err_progress_check_timer);
1620 
1621 		/*
1622 		 * This waits for the state machine to exit so it is not
1623 		 * necessary to kill the sdma_sw_clean_up_task to make sure
1624 		 * it is not running.
1625 		 */
1626 		sdma_finalput(&sde->state);
1627 	}
1628 	sdma_clean(dd, dd->num_sdma);
1629 }
1630 
1631 /*
1632  * unmap the indicated descriptor
1633  */
1634 static inline void sdma_unmap_desc(
1635 	struct hfi1_devdata *dd,
1636 	struct sdma_desc *descp)
1637 {
1638 	switch (sdma_mapping_type(descp)) {
1639 	case SDMA_MAP_SINGLE:
1640 		dma_unmap_single(
1641 			&dd->pcidev->dev,
1642 			sdma_mapping_addr(descp),
1643 			sdma_mapping_len(descp),
1644 			DMA_TO_DEVICE);
1645 		break;
1646 	case SDMA_MAP_PAGE:
1647 		dma_unmap_page(
1648 			&dd->pcidev->dev,
1649 			sdma_mapping_addr(descp),
1650 			sdma_mapping_len(descp),
1651 			DMA_TO_DEVICE);
1652 		break;
1653 	}
1654 }
1655 
1656 /*
1657  * return the mode as indicated by the first
1658  * descriptor in the tx.
1659  */
1660 static inline u8 ahg_mode(struct sdma_txreq *tx)
1661 {
1662 	return (tx->descp[0].qw[1] & SDMA_DESC1_HEADER_MODE_SMASK)
1663 		>> SDMA_DESC1_HEADER_MODE_SHIFT;
1664 }
1665 
1666 /**
1667  * __sdma_txclean() - clean tx of mappings, descp *kmalloc's
1668  * @dd: hfi1_devdata for unmapping
1669  * @tx: tx request to clean
1670  *
1671  * This is used in the progress routine to clean the tx or
1672  * by the ULP to toss an in-process tx build.
1673  *
1674  * The code can be called multiple times without issue.
1675  *
1676  */
1677 void __sdma_txclean(
1678 	struct hfi1_devdata *dd,
1679 	struct sdma_txreq *tx)
1680 {
1681 	u16 i;
1682 
1683 	if (tx->num_desc) {
1684 		u8 skip = 0, mode = ahg_mode(tx);
1685 
1686 		/* unmap first */
1687 		sdma_unmap_desc(dd, &tx->descp[0]);
1688 		/* determine number of AHG descriptors to skip */
1689 		if (mode > SDMA_AHG_APPLY_UPDATE1)
1690 			skip = mode >> 1;
1691 		for (i = 1 + skip; i < tx->num_desc; i++)
1692 			sdma_unmap_desc(dd, &tx->descp[i]);
1693 		tx->num_desc = 0;
1694 	}
1695 	kfree(tx->coalesce_buf);
1696 	tx->coalesce_buf = NULL;
1697 	/* kmalloc'ed descp */
1698 	if (unlikely(tx->desc_limit > ARRAY_SIZE(tx->descs))) {
1699 		tx->desc_limit = ARRAY_SIZE(tx->descs);
1700 		kfree(tx->descp);
1701 	}
1702 }
1703 
1704 static inline u16 sdma_gethead(struct sdma_engine *sde)
1705 {
1706 	struct hfi1_devdata *dd = sde->dd;
1707 	int use_dmahead;
1708 	u16 hwhead;
1709 
1710 #ifdef CONFIG_SDMA_VERBOSITY
1711 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1712 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1713 #endif
1714 
1715 retry:
1716 	use_dmahead = HFI1_CAP_IS_KSET(USE_SDMA_HEAD) && __sdma_running(sde) &&
1717 					(dd->flags & HFI1_HAS_SDMA_TIMEOUT);
1718 	hwhead = use_dmahead ?
1719 		(u16)le64_to_cpu(*sde->head_dma) :
1720 		(u16)read_sde_csr(sde, SD(HEAD));
1721 
1722 	if (unlikely(HFI1_CAP_IS_KSET(SDMA_HEAD_CHECK))) {
1723 		u16 cnt;
1724 		u16 swtail;
1725 		u16 swhead;
1726 		int sane;
1727 
1728 		swhead = sde->descq_head & sde->sdma_mask;
1729 		/* this code is really bad for cache line trading */
1730 		swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
1731 		cnt = sde->descq_cnt;
1732 
1733 		if (swhead < swtail)
1734 			/* not wrapped */
1735 			sane = (hwhead >= swhead) & (hwhead <= swtail);
1736 		else if (swhead > swtail)
1737 			/* wrapped around */
1738 			sane = ((hwhead >= swhead) && (hwhead < cnt)) ||
1739 				(hwhead <= swtail);
1740 		else
1741 			/* empty */
1742 			sane = (hwhead == swhead);
1743 
1744 		if (unlikely(!sane)) {
1745 			dd_dev_err(dd, "SDMA(%u) bad head (%s) hwhd=%hu swhd=%hu swtl=%hu cnt=%hu\n",
1746 				   sde->this_idx,
1747 				   use_dmahead ? "dma" : "kreg",
1748 				   hwhead, swhead, swtail, cnt);
1749 			if (use_dmahead) {
1750 				/* try one more time, using csr */
1751 				use_dmahead = 0;
1752 				goto retry;
1753 			}
1754 			/* proceed as if no progress */
1755 			hwhead = swhead;
1756 		}
1757 	}
1758 	return hwhead;
1759 }
1760 
1761 /*
1762  * This is called when there are send DMA descriptors that might be
1763  * available.
1764  *
1765  * This is called with head_lock held.
1766  */
1767 static void sdma_desc_avail(struct sdma_engine *sde, unsigned avail)
1768 {
1769 	struct iowait *wait, *nw;
1770 	struct iowait *waits[SDMA_WAIT_BATCH_SIZE];
1771 	unsigned i, n = 0, seq;
1772 	struct sdma_txreq *stx;
1773 	struct hfi1_ibdev *dev = &sde->dd->verbs_dev;
1774 
1775 #ifdef CONFIG_SDMA_VERBOSITY
1776 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", sde->this_idx,
1777 		   slashstrip(__FILE__), __LINE__, __func__);
1778 	dd_dev_err(sde->dd, "avail: %u\n", avail);
1779 #endif
1780 
1781 	do {
1782 		seq = read_seqbegin(&dev->iowait_lock);
1783 		if (!list_empty(&sde->dmawait)) {
1784 			/* at least one item */
1785 			write_seqlock(&dev->iowait_lock);
1786 			/* Harvest waiters wanting DMA descriptors */
1787 			list_for_each_entry_safe(
1788 					wait,
1789 					nw,
1790 					&sde->dmawait,
1791 					list) {
1792 				u16 num_desc = 0;
1793 
1794 				if (!wait->wakeup)
1795 					continue;
1796 				if (n == ARRAY_SIZE(waits))
1797 					break;
1798 				if (!list_empty(&wait->tx_head)) {
1799 					stx = list_first_entry(
1800 						&wait->tx_head,
1801 						struct sdma_txreq,
1802 						list);
1803 					num_desc = stx->num_desc;
1804 				}
1805 				if (num_desc > avail)
1806 					break;
1807 				avail -= num_desc;
1808 				list_del_init(&wait->list);
1809 				waits[n++] = wait;
1810 			}
1811 			write_sequnlock(&dev->iowait_lock);
1812 			break;
1813 		}
1814 	} while (read_seqretry(&dev->iowait_lock, seq));
1815 
1816 	for (i = 0; i < n; i++)
1817 		waits[i]->wakeup(waits[i], SDMA_AVAIL_REASON);
1818 }
1819 
1820 /* head_lock must be held */
1821 static void sdma_make_progress(struct sdma_engine *sde, u64 status)
1822 {
1823 	struct sdma_txreq *txp = NULL;
1824 	int progress = 0;
1825 	u16 hwhead, swhead;
1826 	int idle_check_done = 0;
1827 
1828 	hwhead = sdma_gethead(sde);
1829 
1830 	/* The reason for some of the complexity of this code is that
1831 	 * not all descriptors have corresponding txps.  So, we have to
1832 	 * be able to skip over descs until we wander into the range of
1833 	 * the next txp on the list.
1834 	 */
1835 
1836 retry:
1837 	txp = get_txhead(sde);
1838 	swhead = sde->descq_head & sde->sdma_mask;
1839 	trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1840 	while (swhead != hwhead) {
1841 		/* advance head, wrap if needed */
1842 		swhead = ++sde->descq_head & sde->sdma_mask;
1843 
1844 		/* if now past this txp's descs, do the callback */
1845 		if (txp && txp->next_descq_idx == swhead) {
1846 			/* remove from list */
1847 			sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
1848 			complete_tx(sde, txp, SDMA_TXREQ_S_OK);
1849 			/* see if there is another txp */
1850 			txp = get_txhead(sde);
1851 		}
1852 		trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1853 		progress++;
1854 	}
1855 
1856 	/*
1857 	 * The SDMA idle interrupt is not guaranteed to be ordered with respect
1858 	 * to updates to the the dma_head location in host memory. The head
1859 	 * value read might not be fully up to date. If there are pending
1860 	 * descriptors and the SDMA idle interrupt fired then read from the
1861 	 * CSR SDMA head instead to get the latest value from the hardware.
1862 	 * The hardware SDMA head should be read at most once in this invocation
1863 	 * of sdma_make_progress(..) which is ensured by idle_check_done flag
1864 	 */
1865 	if ((status & sde->idle_mask) && !idle_check_done) {
1866 		u16 swtail;
1867 
1868 		swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
1869 		if (swtail != hwhead) {
1870 			hwhead = (u16)read_sde_csr(sde, SD(HEAD));
1871 			idle_check_done = 1;
1872 			goto retry;
1873 		}
1874 	}
1875 
1876 	sde->last_status = status;
1877 	if (progress)
1878 		sdma_desc_avail(sde, sdma_descq_freecnt(sde));
1879 }
1880 
1881 /*
1882  * sdma_engine_interrupt() - interrupt handler for engine
1883  * @sde: sdma engine
1884  * @status: sdma interrupt reason
1885  *
1886  * Status is a mask of the 3 possible interrupts for this engine.  It will
1887  * contain bits _only_ for this SDMA engine.  It will contain at least one
1888  * bit, it may contain more.
1889  */
1890 void sdma_engine_interrupt(struct sdma_engine *sde, u64 status)
1891 {
1892 	trace_hfi1_sdma_engine_interrupt(sde, status);
1893 	write_seqlock(&sde->head_lock);
1894 	sdma_set_desc_cnt(sde, sdma_desct_intr);
1895 	if (status & sde->idle_mask)
1896 		sde->idle_int_cnt++;
1897 	else if (status & sde->progress_mask)
1898 		sde->progress_int_cnt++;
1899 	else if (status & sde->int_mask)
1900 		sde->sdma_int_cnt++;
1901 	sdma_make_progress(sde, status);
1902 	write_sequnlock(&sde->head_lock);
1903 }
1904 
1905 /**
1906  * sdma_engine_error() - error handler for engine
1907  * @sde: sdma engine
1908  * @status: sdma interrupt reason
1909  */
1910 void sdma_engine_error(struct sdma_engine *sde, u64 status)
1911 {
1912 	unsigned long flags;
1913 
1914 #ifdef CONFIG_SDMA_VERBOSITY
1915 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) error status 0x%llx state %s\n",
1916 		   sde->this_idx,
1917 		   (unsigned long long)status,
1918 		   sdma_state_names[sde->state.current_state]);
1919 #endif
1920 	spin_lock_irqsave(&sde->tail_lock, flags);
1921 	write_seqlock(&sde->head_lock);
1922 	if (status & ALL_SDMA_ENG_HALT_ERRS)
1923 		__sdma_process_event(sde, sdma_event_e60_hw_halted);
1924 	if (status & ~SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK)) {
1925 		dd_dev_err(sde->dd,
1926 			   "SDMA (%u) engine error: 0x%llx state %s\n",
1927 			   sde->this_idx,
1928 			   (unsigned long long)status,
1929 			   sdma_state_names[sde->state.current_state]);
1930 		dump_sdma_state(sde);
1931 	}
1932 	write_sequnlock(&sde->head_lock);
1933 	spin_unlock_irqrestore(&sde->tail_lock, flags);
1934 }
1935 
1936 static void sdma_sendctrl(struct sdma_engine *sde, unsigned op)
1937 {
1938 	u64 set_senddmactrl = 0;
1939 	u64 clr_senddmactrl = 0;
1940 	unsigned long flags;
1941 
1942 #ifdef CONFIG_SDMA_VERBOSITY
1943 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) senddmactrl E=%d I=%d H=%d C=%d\n",
1944 		   sde->this_idx,
1945 		   (op & SDMA_SENDCTRL_OP_ENABLE) ? 1 : 0,
1946 		   (op & SDMA_SENDCTRL_OP_INTENABLE) ? 1 : 0,
1947 		   (op & SDMA_SENDCTRL_OP_HALT) ? 1 : 0,
1948 		   (op & SDMA_SENDCTRL_OP_CLEANUP) ? 1 : 0);
1949 #endif
1950 
1951 	if (op & SDMA_SENDCTRL_OP_ENABLE)
1952 		set_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1953 	else
1954 		clr_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1955 
1956 	if (op & SDMA_SENDCTRL_OP_INTENABLE)
1957 		set_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1958 	else
1959 		clr_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1960 
1961 	if (op & SDMA_SENDCTRL_OP_HALT)
1962 		set_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1963 	else
1964 		clr_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1965 
1966 	spin_lock_irqsave(&sde->senddmactrl_lock, flags);
1967 
1968 	sde->p_senddmactrl |= set_senddmactrl;
1969 	sde->p_senddmactrl &= ~clr_senddmactrl;
1970 
1971 	if (op & SDMA_SENDCTRL_OP_CLEANUP)
1972 		write_sde_csr(sde, SD(CTRL),
1973 			      sde->p_senddmactrl |
1974 			      SD(CTRL_SDMA_CLEANUP_SMASK));
1975 	else
1976 		write_sde_csr(sde, SD(CTRL), sde->p_senddmactrl);
1977 
1978 	spin_unlock_irqrestore(&sde->senddmactrl_lock, flags);
1979 
1980 #ifdef CONFIG_SDMA_VERBOSITY
1981 	sdma_dumpstate(sde);
1982 #endif
1983 }
1984 
1985 static void sdma_setlengen(struct sdma_engine *sde)
1986 {
1987 #ifdef CONFIG_SDMA_VERBOSITY
1988 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1989 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1990 #endif
1991 
1992 	/*
1993 	 * Set SendDmaLenGen and clear-then-set the MSB of the generation
1994 	 * count to enable generation checking and load the internal
1995 	 * generation counter.
1996 	 */
1997 	write_sde_csr(sde, SD(LEN_GEN),
1998 		      (sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT));
1999 	write_sde_csr(sde, SD(LEN_GEN),
2000 		      ((sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT)) |
2001 		      (4ULL << SD(LEN_GEN_GENERATION_SHIFT)));
2002 }
2003 
2004 static inline void sdma_update_tail(struct sdma_engine *sde, u16 tail)
2005 {
2006 	/* Commit writes to memory and advance the tail on the chip */
2007 	smp_wmb(); /* see get_txhead() */
2008 	writeq(tail, sde->tail_csr);
2009 }
2010 
2011 /*
2012  * This is called when changing to state s10_hw_start_up_halt_wait as
2013  * a result of send buffer errors or send DMA descriptor errors.
2014  */
2015 static void sdma_hw_start_up(struct sdma_engine *sde)
2016 {
2017 	u64 reg;
2018 
2019 #ifdef CONFIG_SDMA_VERBOSITY
2020 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2021 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2022 #endif
2023 
2024 	sdma_setlengen(sde);
2025 	sdma_update_tail(sde, 0); /* Set SendDmaTail */
2026 	*sde->head_dma = 0;
2027 
2028 	reg = SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_MASK) <<
2029 	      SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SHIFT);
2030 	write_sde_csr(sde, SD(ENG_ERR_CLEAR), reg);
2031 }
2032 
2033 /*
2034  * set_sdma_integrity
2035  *
2036  * Set the SEND_DMA_CHECK_ENABLE register for send DMA engine 'sde'.
2037  */
2038 static void set_sdma_integrity(struct sdma_engine *sde)
2039 {
2040 	struct hfi1_devdata *dd = sde->dd;
2041 
2042 	write_sde_csr(sde, SD(CHECK_ENABLE),
2043 		      hfi1_pkt_base_sdma_integrity(dd));
2044 }
2045 
2046 static void init_sdma_regs(
2047 	struct sdma_engine *sde,
2048 	u32 credits,
2049 	uint idle_cnt)
2050 {
2051 	u8 opval, opmask;
2052 #ifdef CONFIG_SDMA_VERBOSITY
2053 	struct hfi1_devdata *dd = sde->dd;
2054 
2055 	dd_dev_err(dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2056 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2057 #endif
2058 
2059 	write_sde_csr(sde, SD(BASE_ADDR), sde->descq_phys);
2060 	sdma_setlengen(sde);
2061 	sdma_update_tail(sde, 0); /* Set SendDmaTail */
2062 	write_sde_csr(sde, SD(RELOAD_CNT), idle_cnt);
2063 	write_sde_csr(sde, SD(DESC_CNT), 0);
2064 	write_sde_csr(sde, SD(HEAD_ADDR), sde->head_phys);
2065 	write_sde_csr(sde, SD(MEMORY),
2066 		      ((u64)credits << SD(MEMORY_SDMA_MEMORY_CNT_SHIFT)) |
2067 		      ((u64)(credits * sde->this_idx) <<
2068 		       SD(MEMORY_SDMA_MEMORY_INDEX_SHIFT)));
2069 	write_sde_csr(sde, SD(ENG_ERR_MASK), ~0ull);
2070 	set_sdma_integrity(sde);
2071 	opmask = OPCODE_CHECK_MASK_DISABLED;
2072 	opval = OPCODE_CHECK_VAL_DISABLED;
2073 	write_sde_csr(sde, SD(CHECK_OPCODE),
2074 		      (opmask << SEND_CTXT_CHECK_OPCODE_MASK_SHIFT) |
2075 		      (opval << SEND_CTXT_CHECK_OPCODE_VALUE_SHIFT));
2076 }
2077 
2078 #ifdef CONFIG_SDMA_VERBOSITY
2079 
2080 #define sdma_dumpstate_helper0(reg) do { \
2081 		csr = read_csr(sde->dd, reg); \
2082 		dd_dev_err(sde->dd, "%36s     0x%016llx\n", #reg, csr); \
2083 	} while (0)
2084 
2085 #define sdma_dumpstate_helper(reg) do { \
2086 		csr = read_sde_csr(sde, reg); \
2087 		dd_dev_err(sde->dd, "%36s[%02u] 0x%016llx\n", \
2088 			#reg, sde->this_idx, csr); \
2089 	} while (0)
2090 
2091 #define sdma_dumpstate_helper2(reg) do { \
2092 		csr = read_csr(sde->dd, reg + (8 * i)); \
2093 		dd_dev_err(sde->dd, "%33s_%02u     0x%016llx\n", \
2094 				#reg, i, csr); \
2095 	} while (0)
2096 
2097 void sdma_dumpstate(struct sdma_engine *sde)
2098 {
2099 	u64 csr;
2100 	unsigned i;
2101 
2102 	sdma_dumpstate_helper(SD(CTRL));
2103 	sdma_dumpstate_helper(SD(STATUS));
2104 	sdma_dumpstate_helper0(SD(ERR_STATUS));
2105 	sdma_dumpstate_helper0(SD(ERR_MASK));
2106 	sdma_dumpstate_helper(SD(ENG_ERR_STATUS));
2107 	sdma_dumpstate_helper(SD(ENG_ERR_MASK));
2108 
2109 	for (i = 0; i < CCE_NUM_INT_CSRS; ++i) {
2110 		sdma_dumpstate_helper2(CCE_INT_STATUS);
2111 		sdma_dumpstate_helper2(CCE_INT_MASK);
2112 		sdma_dumpstate_helper2(CCE_INT_BLOCKED);
2113 	}
2114 
2115 	sdma_dumpstate_helper(SD(TAIL));
2116 	sdma_dumpstate_helper(SD(HEAD));
2117 	sdma_dumpstate_helper(SD(PRIORITY_THLD));
2118 	sdma_dumpstate_helper(SD(IDLE_CNT));
2119 	sdma_dumpstate_helper(SD(RELOAD_CNT));
2120 	sdma_dumpstate_helper(SD(DESC_CNT));
2121 	sdma_dumpstate_helper(SD(DESC_FETCHED_CNT));
2122 	sdma_dumpstate_helper(SD(MEMORY));
2123 	sdma_dumpstate_helper0(SD(ENGINES));
2124 	sdma_dumpstate_helper0(SD(MEM_SIZE));
2125 	/* sdma_dumpstate_helper(SEND_EGRESS_SEND_DMA_STATUS);  */
2126 	sdma_dumpstate_helper(SD(BASE_ADDR));
2127 	sdma_dumpstate_helper(SD(LEN_GEN));
2128 	sdma_dumpstate_helper(SD(HEAD_ADDR));
2129 	sdma_dumpstate_helper(SD(CHECK_ENABLE));
2130 	sdma_dumpstate_helper(SD(CHECK_VL));
2131 	sdma_dumpstate_helper(SD(CHECK_JOB_KEY));
2132 	sdma_dumpstate_helper(SD(CHECK_PARTITION_KEY));
2133 	sdma_dumpstate_helper(SD(CHECK_SLID));
2134 	sdma_dumpstate_helper(SD(CHECK_OPCODE));
2135 }
2136 #endif
2137 
2138 static void dump_sdma_state(struct sdma_engine *sde)
2139 {
2140 	struct hw_sdma_desc *descq;
2141 	struct hw_sdma_desc *descqp;
2142 	u64 desc[2];
2143 	u64 addr;
2144 	u8 gen;
2145 	u16 len;
2146 	u16 head, tail, cnt;
2147 
2148 	head = sde->descq_head & sde->sdma_mask;
2149 	tail = sde->descq_tail & sde->sdma_mask;
2150 	cnt = sdma_descq_freecnt(sde);
2151 	descq = sde->descq;
2152 
2153 	dd_dev_err(sde->dd,
2154 		   "SDMA (%u) descq_head: %u descq_tail: %u freecnt: %u FLE %d\n",
2155 		   sde->this_idx, head, tail, cnt,
2156 		   !list_empty(&sde->flushlist));
2157 
2158 	/* print info for each entry in the descriptor queue */
2159 	while (head != tail) {
2160 		char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2161 
2162 		descqp = &sde->descq[head];
2163 		desc[0] = le64_to_cpu(descqp->qw[0]);
2164 		desc[1] = le64_to_cpu(descqp->qw[1]);
2165 		flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2166 		flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2167 				'H' : '-';
2168 		flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2169 		flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2170 		addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2171 			& SDMA_DESC0_PHY_ADDR_MASK;
2172 		gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2173 			& SDMA_DESC1_GENERATION_MASK;
2174 		len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2175 			& SDMA_DESC0_BYTE_COUNT_MASK;
2176 		dd_dev_err(sde->dd,
2177 			   "SDMA sdmadesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2178 			   head, flags, addr, gen, len);
2179 		dd_dev_err(sde->dd,
2180 			   "\tdesc0:0x%016llx desc1 0x%016llx\n",
2181 			   desc[0], desc[1]);
2182 		if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2183 			dd_dev_err(sde->dd,
2184 				   "\taidx: %u amode: %u alen: %u\n",
2185 				   (u8)((desc[1] &
2186 					 SDMA_DESC1_HEADER_INDEX_SMASK) >>
2187 					SDMA_DESC1_HEADER_INDEX_SHIFT),
2188 				   (u8)((desc[1] &
2189 					 SDMA_DESC1_HEADER_MODE_SMASK) >>
2190 					SDMA_DESC1_HEADER_MODE_SHIFT),
2191 				   (u8)((desc[1] &
2192 					 SDMA_DESC1_HEADER_DWS_SMASK) >>
2193 					SDMA_DESC1_HEADER_DWS_SHIFT));
2194 		head++;
2195 		head &= sde->sdma_mask;
2196 	}
2197 }
2198 
2199 #define SDE_FMT \
2200 	"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"
2201 /**
2202  * sdma_seqfile_dump_sde() - debugfs dump of sde
2203  * @s: seq file
2204  * @sde: send dma engine to dump
2205  *
2206  * This routine dumps the sde to the indicated seq file.
2207  */
2208 void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *sde)
2209 {
2210 	u16 head, tail;
2211 	struct hw_sdma_desc *descqp;
2212 	u64 desc[2];
2213 	u64 addr;
2214 	u8 gen;
2215 	u16 len;
2216 
2217 	head = sde->descq_head & sde->sdma_mask;
2218 	tail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
2219 	seq_printf(s, SDE_FMT, sde->this_idx,
2220 		   sde->cpu,
2221 		   sdma_state_name(sde->state.current_state),
2222 		   (unsigned long long)read_sde_csr(sde, SD(CTRL)),
2223 		   (unsigned long long)read_sde_csr(sde, SD(STATUS)),
2224 		   (unsigned long long)read_sde_csr(sde, SD(ENG_ERR_STATUS)),
2225 		   (unsigned long long)read_sde_csr(sde, SD(TAIL)), tail,
2226 		   (unsigned long long)read_sde_csr(sde, SD(HEAD)), head,
2227 		   (unsigned long long)le64_to_cpu(*sde->head_dma),
2228 		   (unsigned long long)read_sde_csr(sde, SD(MEMORY)),
2229 		   (unsigned long long)read_sde_csr(sde, SD(LEN_GEN)),
2230 		   (unsigned long long)read_sde_csr(sde, SD(RELOAD_CNT)),
2231 		   (unsigned long long)sde->last_status,
2232 		   (unsigned long long)sde->ahg_bits,
2233 		   sde->tx_tail,
2234 		   sde->tx_head,
2235 		   sde->descq_tail,
2236 		   sde->descq_head,
2237 		   !list_empty(&sde->flushlist),
2238 		   sde->descq_full_count,
2239 		   (unsigned long long)read_sde_csr(sde, SEND_DMA_CHECK_SLID));
2240 
2241 	/* print info for each entry in the descriptor queue */
2242 	while (head != tail) {
2243 		char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2244 
2245 		descqp = &sde->descq[head];
2246 		desc[0] = le64_to_cpu(descqp->qw[0]);
2247 		desc[1] = le64_to_cpu(descqp->qw[1]);
2248 		flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2249 		flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2250 				'H' : '-';
2251 		flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2252 		flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2253 		addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2254 			& SDMA_DESC0_PHY_ADDR_MASK;
2255 		gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2256 			& SDMA_DESC1_GENERATION_MASK;
2257 		len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2258 			& SDMA_DESC0_BYTE_COUNT_MASK;
2259 		seq_printf(s,
2260 			   "\tdesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2261 			   head, flags, addr, gen, len);
2262 		if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2263 			seq_printf(s, "\t\tahgidx: %u ahgmode: %u\n",
2264 				   (u8)((desc[1] &
2265 					 SDMA_DESC1_HEADER_INDEX_SMASK) >>
2266 					SDMA_DESC1_HEADER_INDEX_SHIFT),
2267 				   (u8)((desc[1] &
2268 					 SDMA_DESC1_HEADER_MODE_SMASK) >>
2269 					SDMA_DESC1_HEADER_MODE_SHIFT));
2270 		head = (head + 1) & sde->sdma_mask;
2271 	}
2272 }
2273 
2274 /*
2275  * add the generation number into
2276  * the qw1 and return
2277  */
2278 static inline u64 add_gen(struct sdma_engine *sde, u64 qw1)
2279 {
2280 	u8 generation = (sde->descq_tail >> sde->sdma_shift) & 3;
2281 
2282 	qw1 &= ~SDMA_DESC1_GENERATION_SMASK;
2283 	qw1 |= ((u64)generation & SDMA_DESC1_GENERATION_MASK)
2284 			<< SDMA_DESC1_GENERATION_SHIFT;
2285 	return qw1;
2286 }
2287 
2288 /*
2289  * This routine submits the indicated tx
2290  *
2291  * Space has already been guaranteed and
2292  * tail side of ring is locked.
2293  *
2294  * The hardware tail update is done
2295  * in the caller and that is facilitated
2296  * by returning the new tail.
2297  *
2298  * There is special case logic for ahg
2299  * to not add the generation number for
2300  * up to 2 descriptors that follow the
2301  * first descriptor.
2302  *
2303  */
2304 static inline u16 submit_tx(struct sdma_engine *sde, struct sdma_txreq *tx)
2305 {
2306 	int i;
2307 	u16 tail;
2308 	struct sdma_desc *descp = tx->descp;
2309 	u8 skip = 0, mode = ahg_mode(tx);
2310 
2311 	tail = sde->descq_tail & sde->sdma_mask;
2312 	sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2313 	sde->descq[tail].qw[1] = cpu_to_le64(add_gen(sde, descp->qw[1]));
2314 	trace_hfi1_sdma_descriptor(sde, descp->qw[0], descp->qw[1],
2315 				   tail, &sde->descq[tail]);
2316 	tail = ++sde->descq_tail & sde->sdma_mask;
2317 	descp++;
2318 	if (mode > SDMA_AHG_APPLY_UPDATE1)
2319 		skip = mode >> 1;
2320 	for (i = 1; i < tx->num_desc; i++, descp++) {
2321 		u64 qw1;
2322 
2323 		sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2324 		if (skip) {
2325 			/* edits don't have generation */
2326 			qw1 = descp->qw[1];
2327 			skip--;
2328 		} else {
2329 			/* replace generation with real one for non-edits */
2330 			qw1 = add_gen(sde, descp->qw[1]);
2331 		}
2332 		sde->descq[tail].qw[1] = cpu_to_le64(qw1);
2333 		trace_hfi1_sdma_descriptor(sde, descp->qw[0], qw1,
2334 					   tail, &sde->descq[tail]);
2335 		tail = ++sde->descq_tail & sde->sdma_mask;
2336 	}
2337 	tx->next_descq_idx = tail;
2338 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2339 	tx->sn = sde->tail_sn++;
2340 	trace_hfi1_sdma_in_sn(sde, tx->sn);
2341 	WARN_ON_ONCE(sde->tx_ring[sde->tx_tail & sde->sdma_mask]);
2342 #endif
2343 	sde->tx_ring[sde->tx_tail++ & sde->sdma_mask] = tx;
2344 	sde->desc_avail -= tx->num_desc;
2345 	return tail;
2346 }
2347 
2348 /*
2349  * Check for progress
2350  */
2351 static int sdma_check_progress(
2352 	struct sdma_engine *sde,
2353 	struct iowait *wait,
2354 	struct sdma_txreq *tx)
2355 {
2356 	int ret;
2357 
2358 	sde->desc_avail = sdma_descq_freecnt(sde);
2359 	if (tx->num_desc <= sde->desc_avail)
2360 		return -EAGAIN;
2361 	/* pulse the head_lock */
2362 	if (wait && wait->sleep) {
2363 		unsigned seq;
2364 
2365 		seq = raw_seqcount_begin(
2366 			(const seqcount_t *)&sde->head_lock.seqcount);
2367 		ret = wait->sleep(sde, wait, tx, seq);
2368 		if (ret == -EAGAIN)
2369 			sde->desc_avail = sdma_descq_freecnt(sde);
2370 	} else {
2371 		ret = -EBUSY;
2372 	}
2373 	return ret;
2374 }
2375 
2376 /**
2377  * sdma_send_txreq() - submit a tx req to ring
2378  * @sde: sdma engine to use
2379  * @wait: wait structure to use when full (may be NULL)
2380  * @tx: sdma_txreq to submit
2381  *
2382  * The call submits the tx into the ring.  If a iowait structure is non-NULL
2383  * the packet will be queued to the list in wait.
2384  *
2385  * Return:
2386  * 0 - Success, -EINVAL - sdma_txreq incomplete, -EBUSY - no space in
2387  * ring (wait == NULL)
2388  * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2389  */
2390 int sdma_send_txreq(struct sdma_engine *sde,
2391 		    struct iowait *wait,
2392 		    struct sdma_txreq *tx)
2393 {
2394 	int ret = 0;
2395 	u16 tail;
2396 	unsigned long flags;
2397 
2398 	/* user should have supplied entire packet */
2399 	if (unlikely(tx->tlen))
2400 		return -EINVAL;
2401 	tx->wait = wait;
2402 	spin_lock_irqsave(&sde->tail_lock, flags);
2403 retry:
2404 	if (unlikely(!__sdma_running(sde)))
2405 		goto unlock_noconn;
2406 	if (unlikely(tx->num_desc > sde->desc_avail))
2407 		goto nodesc;
2408 	tail = submit_tx(sde, tx);
2409 	if (wait)
2410 		iowait_sdma_inc(wait);
2411 	sdma_update_tail(sde, tail);
2412 unlock:
2413 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2414 	return ret;
2415 unlock_noconn:
2416 	if (wait)
2417 		iowait_sdma_inc(wait);
2418 	tx->next_descq_idx = 0;
2419 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2420 	tx->sn = sde->tail_sn++;
2421 	trace_hfi1_sdma_in_sn(sde, tx->sn);
2422 #endif
2423 	spin_lock(&sde->flushlist_lock);
2424 	list_add_tail(&tx->list, &sde->flushlist);
2425 	spin_unlock(&sde->flushlist_lock);
2426 	if (wait) {
2427 		wait->tx_count++;
2428 		wait->count += tx->num_desc;
2429 	}
2430 	schedule_work(&sde->flush_worker);
2431 	ret = -ECOMM;
2432 	goto unlock;
2433 nodesc:
2434 	ret = sdma_check_progress(sde, wait, tx);
2435 	if (ret == -EAGAIN) {
2436 		ret = 0;
2437 		goto retry;
2438 	}
2439 	sde->descq_full_count++;
2440 	goto unlock;
2441 }
2442 
2443 /**
2444  * sdma_send_txlist() - submit a list of tx req to ring
2445  * @sde: sdma engine to use
2446  * @wait: wait structure to use when full (may be NULL)
2447  * @tx_list: list of sdma_txreqs to submit
2448  * @count: pointer to a u32 which, after return will contain the total number of
2449  *         sdma_txreqs removed from the tx_list. This will include sdma_txreqs
2450  *         whose SDMA descriptors are submitted to the ring and the sdma_txreqs
2451  *         which are added to SDMA engine flush list if the SDMA engine state is
2452  *         not running.
2453  *
2454  * The call submits the list into the ring.
2455  *
2456  * If the iowait structure is non-NULL and not equal to the iowait list
2457  * the unprocessed part of the list  will be appended to the list in wait.
2458  *
2459  * In all cases, the tx_list will be updated so the head of the tx_list is
2460  * the list of descriptors that have yet to be transmitted.
2461  *
2462  * The intent of this call is to provide a more efficient
2463  * way of submitting multiple packets to SDMA while holding the tail
2464  * side locking.
2465  *
2466  * Return:
2467  * 0 - Success,
2468  * -EINVAL - sdma_txreq incomplete, -EBUSY - no space in ring (wait == NULL)
2469  * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2470  */
2471 int sdma_send_txlist(struct sdma_engine *sde, struct iowait *wait,
2472 		     struct list_head *tx_list, u32 *count_out)
2473 {
2474 	struct sdma_txreq *tx, *tx_next;
2475 	int ret = 0;
2476 	unsigned long flags;
2477 	u16 tail = INVALID_TAIL;
2478 	u32 submit_count = 0, flush_count = 0, total_count;
2479 
2480 	spin_lock_irqsave(&sde->tail_lock, flags);
2481 retry:
2482 	list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2483 		tx->wait = wait;
2484 		if (unlikely(!__sdma_running(sde)))
2485 			goto unlock_noconn;
2486 		if (unlikely(tx->num_desc > sde->desc_avail))
2487 			goto nodesc;
2488 		if (unlikely(tx->tlen)) {
2489 			ret = -EINVAL;
2490 			goto update_tail;
2491 		}
2492 		list_del_init(&tx->list);
2493 		tail = submit_tx(sde, tx);
2494 		submit_count++;
2495 		if (tail != INVALID_TAIL &&
2496 		    (submit_count & SDMA_TAIL_UPDATE_THRESH) == 0) {
2497 			sdma_update_tail(sde, tail);
2498 			tail = INVALID_TAIL;
2499 		}
2500 	}
2501 update_tail:
2502 	total_count = submit_count + flush_count;
2503 	if (wait)
2504 		iowait_sdma_add(wait, total_count);
2505 	if (tail != INVALID_TAIL)
2506 		sdma_update_tail(sde, tail);
2507 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2508 	*count_out = total_count;
2509 	return ret;
2510 unlock_noconn:
2511 	spin_lock(&sde->flushlist_lock);
2512 	list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2513 		tx->wait = wait;
2514 		list_del_init(&tx->list);
2515 		tx->next_descq_idx = 0;
2516 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2517 		tx->sn = sde->tail_sn++;
2518 		trace_hfi1_sdma_in_sn(sde, tx->sn);
2519 #endif
2520 		list_add_tail(&tx->list, &sde->flushlist);
2521 		flush_count++;
2522 		if (wait) {
2523 			wait->tx_count++;
2524 			wait->count += tx->num_desc;
2525 		}
2526 	}
2527 	spin_unlock(&sde->flushlist_lock);
2528 	schedule_work(&sde->flush_worker);
2529 	ret = -ECOMM;
2530 	goto update_tail;
2531 nodesc:
2532 	ret = sdma_check_progress(sde, wait, tx);
2533 	if (ret == -EAGAIN) {
2534 		ret = 0;
2535 		goto retry;
2536 	}
2537 	sde->descq_full_count++;
2538 	goto update_tail;
2539 }
2540 
2541 static void sdma_process_event(struct sdma_engine *sde, enum sdma_events event)
2542 {
2543 	unsigned long flags;
2544 
2545 	spin_lock_irqsave(&sde->tail_lock, flags);
2546 	write_seqlock(&sde->head_lock);
2547 
2548 	__sdma_process_event(sde, event);
2549 
2550 	if (sde->state.current_state == sdma_state_s99_running)
2551 		sdma_desc_avail(sde, sdma_descq_freecnt(sde));
2552 
2553 	write_sequnlock(&sde->head_lock);
2554 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2555 }
2556 
2557 static void __sdma_process_event(struct sdma_engine *sde,
2558 				 enum sdma_events event)
2559 {
2560 	struct sdma_state *ss = &sde->state;
2561 	int need_progress = 0;
2562 
2563 	/* CONFIG SDMA temporary */
2564 #ifdef CONFIG_SDMA_VERBOSITY
2565 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) [%s] %s\n", sde->this_idx,
2566 		   sdma_state_names[ss->current_state],
2567 		   sdma_event_names[event]);
2568 #endif
2569 
2570 	switch (ss->current_state) {
2571 	case sdma_state_s00_hw_down:
2572 		switch (event) {
2573 		case sdma_event_e00_go_hw_down:
2574 			break;
2575 		case sdma_event_e30_go_running:
2576 			/*
2577 			 * If down, but running requested (usually result
2578 			 * of link up, then we need to start up.
2579 			 * This can happen when hw down is requested while
2580 			 * bringing the link up with traffic active on
2581 			 * 7220, e.g.
2582 			 */
2583 			ss->go_s99_running = 1;
2584 			/* fall through and start dma engine */
2585 		case sdma_event_e10_go_hw_start:
2586 			/* This reference means the state machine is started */
2587 			sdma_get(&sde->state);
2588 			sdma_set_state(sde,
2589 				       sdma_state_s10_hw_start_up_halt_wait);
2590 			break;
2591 		case sdma_event_e15_hw_halt_done:
2592 			break;
2593 		case sdma_event_e25_hw_clean_up_done:
2594 			break;
2595 		case sdma_event_e40_sw_cleaned:
2596 			sdma_sw_tear_down(sde);
2597 			break;
2598 		case sdma_event_e50_hw_cleaned:
2599 			break;
2600 		case sdma_event_e60_hw_halted:
2601 			break;
2602 		case sdma_event_e70_go_idle:
2603 			break;
2604 		case sdma_event_e80_hw_freeze:
2605 			break;
2606 		case sdma_event_e81_hw_frozen:
2607 			break;
2608 		case sdma_event_e82_hw_unfreeze:
2609 			break;
2610 		case sdma_event_e85_link_down:
2611 			break;
2612 		case sdma_event_e90_sw_halted:
2613 			break;
2614 		}
2615 		break;
2616 
2617 	case sdma_state_s10_hw_start_up_halt_wait:
2618 		switch (event) {
2619 		case sdma_event_e00_go_hw_down:
2620 			sdma_set_state(sde, sdma_state_s00_hw_down);
2621 			sdma_sw_tear_down(sde);
2622 			break;
2623 		case sdma_event_e10_go_hw_start:
2624 			break;
2625 		case sdma_event_e15_hw_halt_done:
2626 			sdma_set_state(sde,
2627 				       sdma_state_s15_hw_start_up_clean_wait);
2628 			sdma_start_hw_clean_up(sde);
2629 			break;
2630 		case sdma_event_e25_hw_clean_up_done:
2631 			break;
2632 		case sdma_event_e30_go_running:
2633 			ss->go_s99_running = 1;
2634 			break;
2635 		case sdma_event_e40_sw_cleaned:
2636 			break;
2637 		case sdma_event_e50_hw_cleaned:
2638 			break;
2639 		case sdma_event_e60_hw_halted:
2640 			schedule_work(&sde->err_halt_worker);
2641 			break;
2642 		case sdma_event_e70_go_idle:
2643 			ss->go_s99_running = 0;
2644 			break;
2645 		case sdma_event_e80_hw_freeze:
2646 			break;
2647 		case sdma_event_e81_hw_frozen:
2648 			break;
2649 		case sdma_event_e82_hw_unfreeze:
2650 			break;
2651 		case sdma_event_e85_link_down:
2652 			break;
2653 		case sdma_event_e90_sw_halted:
2654 			break;
2655 		}
2656 		break;
2657 
2658 	case sdma_state_s15_hw_start_up_clean_wait:
2659 		switch (event) {
2660 		case sdma_event_e00_go_hw_down:
2661 			sdma_set_state(sde, sdma_state_s00_hw_down);
2662 			sdma_sw_tear_down(sde);
2663 			break;
2664 		case sdma_event_e10_go_hw_start:
2665 			break;
2666 		case sdma_event_e15_hw_halt_done:
2667 			break;
2668 		case sdma_event_e25_hw_clean_up_done:
2669 			sdma_hw_start_up(sde);
2670 			sdma_set_state(sde, ss->go_s99_running ?
2671 				       sdma_state_s99_running :
2672 				       sdma_state_s20_idle);
2673 			break;
2674 		case sdma_event_e30_go_running:
2675 			ss->go_s99_running = 1;
2676 			break;
2677 		case sdma_event_e40_sw_cleaned:
2678 			break;
2679 		case sdma_event_e50_hw_cleaned:
2680 			break;
2681 		case sdma_event_e60_hw_halted:
2682 			break;
2683 		case sdma_event_e70_go_idle:
2684 			ss->go_s99_running = 0;
2685 			break;
2686 		case sdma_event_e80_hw_freeze:
2687 			break;
2688 		case sdma_event_e81_hw_frozen:
2689 			break;
2690 		case sdma_event_e82_hw_unfreeze:
2691 			break;
2692 		case sdma_event_e85_link_down:
2693 			break;
2694 		case sdma_event_e90_sw_halted:
2695 			break;
2696 		}
2697 		break;
2698 
2699 	case sdma_state_s20_idle:
2700 		switch (event) {
2701 		case sdma_event_e00_go_hw_down:
2702 			sdma_set_state(sde, sdma_state_s00_hw_down);
2703 			sdma_sw_tear_down(sde);
2704 			break;
2705 		case sdma_event_e10_go_hw_start:
2706 			break;
2707 		case sdma_event_e15_hw_halt_done:
2708 			break;
2709 		case sdma_event_e25_hw_clean_up_done:
2710 			break;
2711 		case sdma_event_e30_go_running:
2712 			sdma_set_state(sde, sdma_state_s99_running);
2713 			ss->go_s99_running = 1;
2714 			break;
2715 		case sdma_event_e40_sw_cleaned:
2716 			break;
2717 		case sdma_event_e50_hw_cleaned:
2718 			break;
2719 		case sdma_event_e60_hw_halted:
2720 			sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
2721 			schedule_work(&sde->err_halt_worker);
2722 			break;
2723 		case sdma_event_e70_go_idle:
2724 			break;
2725 		case sdma_event_e85_link_down:
2726 			/* fall through */
2727 		case sdma_event_e80_hw_freeze:
2728 			sdma_set_state(sde, sdma_state_s80_hw_freeze);
2729 			atomic_dec(&sde->dd->sdma_unfreeze_count);
2730 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2731 			break;
2732 		case sdma_event_e81_hw_frozen:
2733 			break;
2734 		case sdma_event_e82_hw_unfreeze:
2735 			break;
2736 		case sdma_event_e90_sw_halted:
2737 			break;
2738 		}
2739 		break;
2740 
2741 	case sdma_state_s30_sw_clean_up_wait:
2742 		switch (event) {
2743 		case sdma_event_e00_go_hw_down:
2744 			sdma_set_state(sde, sdma_state_s00_hw_down);
2745 			break;
2746 		case sdma_event_e10_go_hw_start:
2747 			break;
2748 		case sdma_event_e15_hw_halt_done:
2749 			break;
2750 		case sdma_event_e25_hw_clean_up_done:
2751 			break;
2752 		case sdma_event_e30_go_running:
2753 			ss->go_s99_running = 1;
2754 			break;
2755 		case sdma_event_e40_sw_cleaned:
2756 			sdma_set_state(sde, sdma_state_s40_hw_clean_up_wait);
2757 			sdma_start_hw_clean_up(sde);
2758 			break;
2759 		case sdma_event_e50_hw_cleaned:
2760 			break;
2761 		case sdma_event_e60_hw_halted:
2762 			break;
2763 		case sdma_event_e70_go_idle:
2764 			ss->go_s99_running = 0;
2765 			break;
2766 		case sdma_event_e80_hw_freeze:
2767 			break;
2768 		case sdma_event_e81_hw_frozen:
2769 			break;
2770 		case sdma_event_e82_hw_unfreeze:
2771 			break;
2772 		case sdma_event_e85_link_down:
2773 			ss->go_s99_running = 0;
2774 			break;
2775 		case sdma_event_e90_sw_halted:
2776 			break;
2777 		}
2778 		break;
2779 
2780 	case sdma_state_s40_hw_clean_up_wait:
2781 		switch (event) {
2782 		case sdma_event_e00_go_hw_down:
2783 			sdma_set_state(sde, sdma_state_s00_hw_down);
2784 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2785 			break;
2786 		case sdma_event_e10_go_hw_start:
2787 			break;
2788 		case sdma_event_e15_hw_halt_done:
2789 			break;
2790 		case sdma_event_e25_hw_clean_up_done:
2791 			sdma_hw_start_up(sde);
2792 			sdma_set_state(sde, ss->go_s99_running ?
2793 				       sdma_state_s99_running :
2794 				       sdma_state_s20_idle);
2795 			break;
2796 		case sdma_event_e30_go_running:
2797 			ss->go_s99_running = 1;
2798 			break;
2799 		case sdma_event_e40_sw_cleaned:
2800 			break;
2801 		case sdma_event_e50_hw_cleaned:
2802 			break;
2803 		case sdma_event_e60_hw_halted:
2804 			break;
2805 		case sdma_event_e70_go_idle:
2806 			ss->go_s99_running = 0;
2807 			break;
2808 		case sdma_event_e80_hw_freeze:
2809 			break;
2810 		case sdma_event_e81_hw_frozen:
2811 			break;
2812 		case sdma_event_e82_hw_unfreeze:
2813 			break;
2814 		case sdma_event_e85_link_down:
2815 			ss->go_s99_running = 0;
2816 			break;
2817 		case sdma_event_e90_sw_halted:
2818 			break;
2819 		}
2820 		break;
2821 
2822 	case sdma_state_s50_hw_halt_wait:
2823 		switch (event) {
2824 		case sdma_event_e00_go_hw_down:
2825 			sdma_set_state(sde, sdma_state_s00_hw_down);
2826 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2827 			break;
2828 		case sdma_event_e10_go_hw_start:
2829 			break;
2830 		case sdma_event_e15_hw_halt_done:
2831 			sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2832 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2833 			break;
2834 		case sdma_event_e25_hw_clean_up_done:
2835 			break;
2836 		case sdma_event_e30_go_running:
2837 			ss->go_s99_running = 1;
2838 			break;
2839 		case sdma_event_e40_sw_cleaned:
2840 			break;
2841 		case sdma_event_e50_hw_cleaned:
2842 			break;
2843 		case sdma_event_e60_hw_halted:
2844 			schedule_work(&sde->err_halt_worker);
2845 			break;
2846 		case sdma_event_e70_go_idle:
2847 			ss->go_s99_running = 0;
2848 			break;
2849 		case sdma_event_e80_hw_freeze:
2850 			break;
2851 		case sdma_event_e81_hw_frozen:
2852 			break;
2853 		case sdma_event_e82_hw_unfreeze:
2854 			break;
2855 		case sdma_event_e85_link_down:
2856 			ss->go_s99_running = 0;
2857 			break;
2858 		case sdma_event_e90_sw_halted:
2859 			break;
2860 		}
2861 		break;
2862 
2863 	case sdma_state_s60_idle_halt_wait:
2864 		switch (event) {
2865 		case sdma_event_e00_go_hw_down:
2866 			sdma_set_state(sde, sdma_state_s00_hw_down);
2867 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2868 			break;
2869 		case sdma_event_e10_go_hw_start:
2870 			break;
2871 		case sdma_event_e15_hw_halt_done:
2872 			sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2873 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2874 			break;
2875 		case sdma_event_e25_hw_clean_up_done:
2876 			break;
2877 		case sdma_event_e30_go_running:
2878 			ss->go_s99_running = 1;
2879 			break;
2880 		case sdma_event_e40_sw_cleaned:
2881 			break;
2882 		case sdma_event_e50_hw_cleaned:
2883 			break;
2884 		case sdma_event_e60_hw_halted:
2885 			schedule_work(&sde->err_halt_worker);
2886 			break;
2887 		case sdma_event_e70_go_idle:
2888 			ss->go_s99_running = 0;
2889 			break;
2890 		case sdma_event_e80_hw_freeze:
2891 			break;
2892 		case sdma_event_e81_hw_frozen:
2893 			break;
2894 		case sdma_event_e82_hw_unfreeze:
2895 			break;
2896 		case sdma_event_e85_link_down:
2897 			break;
2898 		case sdma_event_e90_sw_halted:
2899 			break;
2900 		}
2901 		break;
2902 
2903 	case sdma_state_s80_hw_freeze:
2904 		switch (event) {
2905 		case sdma_event_e00_go_hw_down:
2906 			sdma_set_state(sde, sdma_state_s00_hw_down);
2907 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2908 			break;
2909 		case sdma_event_e10_go_hw_start:
2910 			break;
2911 		case sdma_event_e15_hw_halt_done:
2912 			break;
2913 		case sdma_event_e25_hw_clean_up_done:
2914 			break;
2915 		case sdma_event_e30_go_running:
2916 			ss->go_s99_running = 1;
2917 			break;
2918 		case sdma_event_e40_sw_cleaned:
2919 			break;
2920 		case sdma_event_e50_hw_cleaned:
2921 			break;
2922 		case sdma_event_e60_hw_halted:
2923 			break;
2924 		case sdma_event_e70_go_idle:
2925 			ss->go_s99_running = 0;
2926 			break;
2927 		case sdma_event_e80_hw_freeze:
2928 			break;
2929 		case sdma_event_e81_hw_frozen:
2930 			sdma_set_state(sde, sdma_state_s82_freeze_sw_clean);
2931 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2932 			break;
2933 		case sdma_event_e82_hw_unfreeze:
2934 			break;
2935 		case sdma_event_e85_link_down:
2936 			break;
2937 		case sdma_event_e90_sw_halted:
2938 			break;
2939 		}
2940 		break;
2941 
2942 	case sdma_state_s82_freeze_sw_clean:
2943 		switch (event) {
2944 		case sdma_event_e00_go_hw_down:
2945 			sdma_set_state(sde, sdma_state_s00_hw_down);
2946 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2947 			break;
2948 		case sdma_event_e10_go_hw_start:
2949 			break;
2950 		case sdma_event_e15_hw_halt_done:
2951 			break;
2952 		case sdma_event_e25_hw_clean_up_done:
2953 			break;
2954 		case sdma_event_e30_go_running:
2955 			ss->go_s99_running = 1;
2956 			break;
2957 		case sdma_event_e40_sw_cleaned:
2958 			/* notify caller this engine is done cleaning */
2959 			atomic_dec(&sde->dd->sdma_unfreeze_count);
2960 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2961 			break;
2962 		case sdma_event_e50_hw_cleaned:
2963 			break;
2964 		case sdma_event_e60_hw_halted:
2965 			break;
2966 		case sdma_event_e70_go_idle:
2967 			ss->go_s99_running = 0;
2968 			break;
2969 		case sdma_event_e80_hw_freeze:
2970 			break;
2971 		case sdma_event_e81_hw_frozen:
2972 			break;
2973 		case sdma_event_e82_hw_unfreeze:
2974 			sdma_hw_start_up(sde);
2975 			sdma_set_state(sde, ss->go_s99_running ?
2976 				       sdma_state_s99_running :
2977 				       sdma_state_s20_idle);
2978 			break;
2979 		case sdma_event_e85_link_down:
2980 			break;
2981 		case sdma_event_e90_sw_halted:
2982 			break;
2983 		}
2984 		break;
2985 
2986 	case sdma_state_s99_running:
2987 		switch (event) {
2988 		case sdma_event_e00_go_hw_down:
2989 			sdma_set_state(sde, sdma_state_s00_hw_down);
2990 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2991 			break;
2992 		case sdma_event_e10_go_hw_start:
2993 			break;
2994 		case sdma_event_e15_hw_halt_done:
2995 			break;
2996 		case sdma_event_e25_hw_clean_up_done:
2997 			break;
2998 		case sdma_event_e30_go_running:
2999 			break;
3000 		case sdma_event_e40_sw_cleaned:
3001 			break;
3002 		case sdma_event_e50_hw_cleaned:
3003 			break;
3004 		case sdma_event_e60_hw_halted:
3005 			need_progress = 1;
3006 			sdma_err_progress_check_schedule(sde);
3007 		case sdma_event_e90_sw_halted:
3008 			/*
3009 			* SW initiated halt does not perform engines
3010 			* progress check
3011 			*/
3012 			sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
3013 			schedule_work(&sde->err_halt_worker);
3014 			break;
3015 		case sdma_event_e70_go_idle:
3016 			sdma_set_state(sde, sdma_state_s60_idle_halt_wait);
3017 			break;
3018 		case sdma_event_e85_link_down:
3019 			ss->go_s99_running = 0;
3020 			/* fall through */
3021 		case sdma_event_e80_hw_freeze:
3022 			sdma_set_state(sde, sdma_state_s80_hw_freeze);
3023 			atomic_dec(&sde->dd->sdma_unfreeze_count);
3024 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
3025 			break;
3026 		case sdma_event_e81_hw_frozen:
3027 			break;
3028 		case sdma_event_e82_hw_unfreeze:
3029 			break;
3030 		}
3031 		break;
3032 	}
3033 
3034 	ss->last_event = event;
3035 	if (need_progress)
3036 		sdma_make_progress(sde, 0);
3037 }
3038 
3039 /*
3040  * _extend_sdma_tx_descs() - helper to extend txreq
3041  *
3042  * This is called once the initial nominal allocation
3043  * of descriptors in the sdma_txreq is exhausted.
3044  *
3045  * The code will bump the allocation up to the max
3046  * of MAX_DESC (64) descriptors. There doesn't seem
3047  * much point in an interim step. The last descriptor
3048  * is reserved for coalesce buffer in order to support
3049  * cases where input packet has >MAX_DESC iovecs.
3050  *
3051  */
3052 static int _extend_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3053 {
3054 	int i;
3055 
3056 	/* Handle last descriptor */
3057 	if (unlikely((tx->num_desc == (MAX_DESC - 1)))) {
3058 		/* if tlen is 0, it is for padding, release last descriptor */
3059 		if (!tx->tlen) {
3060 			tx->desc_limit = MAX_DESC;
3061 		} else if (!tx->coalesce_buf) {
3062 			/* allocate coalesce buffer with space for padding */
3063 			tx->coalesce_buf = kmalloc(tx->tlen + sizeof(u32),
3064 						   GFP_ATOMIC);
3065 			if (!tx->coalesce_buf)
3066 				goto enomem;
3067 			tx->coalesce_idx = 0;
3068 		}
3069 		return 0;
3070 	}
3071 
3072 	if (unlikely(tx->num_desc == MAX_DESC))
3073 		goto enomem;
3074 
3075 	tx->descp = kmalloc_array(
3076 			MAX_DESC,
3077 			sizeof(struct sdma_desc),
3078 			GFP_ATOMIC);
3079 	if (!tx->descp)
3080 		goto enomem;
3081 
3082 	/* reserve last descriptor for coalescing */
3083 	tx->desc_limit = MAX_DESC - 1;
3084 	/* copy ones already built */
3085 	for (i = 0; i < tx->num_desc; i++)
3086 		tx->descp[i] = tx->descs[i];
3087 	return 0;
3088 enomem:
3089 	__sdma_txclean(dd, tx);
3090 	return -ENOMEM;
3091 }
3092 
3093 /*
3094  * ext_coal_sdma_tx_descs() - extend or coalesce sdma tx descriptors
3095  *
3096  * This is called once the initial nominal allocation of descriptors
3097  * in the sdma_txreq is exhausted.
3098  *
3099  * This function calls _extend_sdma_tx_descs to extend or allocate
3100  * coalesce buffer. If there is a allocated coalesce buffer, it will
3101  * copy the input packet data into the coalesce buffer. It also adds
3102  * coalesce buffer descriptor once when whole packet is received.
3103  *
3104  * Return:
3105  * <0 - error
3106  * 0 - coalescing, don't populate descriptor
3107  * 1 - continue with populating descriptor
3108  */
3109 int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx,
3110 			   int type, void *kvaddr, struct page *page,
3111 			   unsigned long offset, u16 len)
3112 {
3113 	int pad_len, rval;
3114 	dma_addr_t addr;
3115 
3116 	rval = _extend_sdma_tx_descs(dd, tx);
3117 	if (rval) {
3118 		__sdma_txclean(dd, tx);
3119 		return rval;
3120 	}
3121 
3122 	/* If coalesce buffer is allocated, copy data into it */
3123 	if (tx->coalesce_buf) {
3124 		if (type == SDMA_MAP_NONE) {
3125 			__sdma_txclean(dd, tx);
3126 			return -EINVAL;
3127 		}
3128 
3129 		if (type == SDMA_MAP_PAGE) {
3130 			kvaddr = kmap(page);
3131 			kvaddr += offset;
3132 		} else if (WARN_ON(!kvaddr)) {
3133 			__sdma_txclean(dd, tx);
3134 			return -EINVAL;
3135 		}
3136 
3137 		memcpy(tx->coalesce_buf + tx->coalesce_idx, kvaddr, len);
3138 		tx->coalesce_idx += len;
3139 		if (type == SDMA_MAP_PAGE)
3140 			kunmap(page);
3141 
3142 		/* If there is more data, return */
3143 		if (tx->tlen - tx->coalesce_idx)
3144 			return 0;
3145 
3146 		/* Whole packet is received; add any padding */
3147 		pad_len = tx->packet_len & (sizeof(u32) - 1);
3148 		if (pad_len) {
3149 			pad_len = sizeof(u32) - pad_len;
3150 			memset(tx->coalesce_buf + tx->coalesce_idx, 0, pad_len);
3151 			/* padding is taken care of for coalescing case */
3152 			tx->packet_len += pad_len;
3153 			tx->tlen += pad_len;
3154 		}
3155 
3156 		/* dma map the coalesce buffer */
3157 		addr = dma_map_single(&dd->pcidev->dev,
3158 				      tx->coalesce_buf,
3159 				      tx->tlen,
3160 				      DMA_TO_DEVICE);
3161 
3162 		if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) {
3163 			__sdma_txclean(dd, tx);
3164 			return -ENOSPC;
3165 		}
3166 
3167 		/* Add descriptor for coalesce buffer */
3168 		tx->desc_limit = MAX_DESC;
3169 		return _sdma_txadd_daddr(dd, SDMA_MAP_SINGLE, tx,
3170 					 addr, tx->tlen);
3171 	}
3172 
3173 	return 1;
3174 }
3175 
3176 /* Update sdes when the lmc changes */
3177 void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid)
3178 {
3179 	struct sdma_engine *sde;
3180 	int i;
3181 	u64 sreg;
3182 
3183 	sreg = ((mask & SD(CHECK_SLID_MASK_MASK)) <<
3184 		SD(CHECK_SLID_MASK_SHIFT)) |
3185 		(((lid & mask) & SD(CHECK_SLID_VALUE_MASK)) <<
3186 		SD(CHECK_SLID_VALUE_SHIFT));
3187 
3188 	for (i = 0; i < dd->num_sdma; i++) {
3189 		hfi1_cdbg(LINKVERB, "SendDmaEngine[%d].SLID_CHECK = 0x%x",
3190 			  i, (u32)sreg);
3191 		sde = &dd->per_sdma[i];
3192 		write_sde_csr(sde, SD(CHECK_SLID), sreg);
3193 	}
3194 }
3195 
3196 /* tx not dword sized - pad */
3197 int _pad_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3198 {
3199 	int rval = 0;
3200 
3201 	tx->num_desc++;
3202 	if ((unlikely(tx->num_desc == tx->desc_limit))) {
3203 		rval = _extend_sdma_tx_descs(dd, tx);
3204 		if (rval) {
3205 			__sdma_txclean(dd, tx);
3206 			return rval;
3207 		}
3208 	}
3209 	/* finish the one just added */
3210 	make_tx_sdma_desc(
3211 		tx,
3212 		SDMA_MAP_NONE,
3213 		dd->sdma_pad_phys,
3214 		sizeof(u32) - (tx->packet_len & (sizeof(u32) - 1)));
3215 	_sdma_close_tx(dd, tx);
3216 	return rval;
3217 }
3218 
3219 /*
3220  * Add ahg to the sdma_txreq
3221  *
3222  * The logic will consume up to 3
3223  * descriptors at the beginning of
3224  * sdma_txreq.
3225  */
3226 void _sdma_txreq_ahgadd(
3227 	struct sdma_txreq *tx,
3228 	u8 num_ahg,
3229 	u8 ahg_entry,
3230 	u32 *ahg,
3231 	u8 ahg_hlen)
3232 {
3233 	u32 i, shift = 0, desc = 0;
3234 	u8 mode;
3235 
3236 	WARN_ON_ONCE(num_ahg > 9 || (ahg_hlen & 3) || ahg_hlen == 4);
3237 	/* compute mode */
3238 	if (num_ahg == 1)
3239 		mode = SDMA_AHG_APPLY_UPDATE1;
3240 	else if (num_ahg <= 5)
3241 		mode = SDMA_AHG_APPLY_UPDATE2;
3242 	else
3243 		mode = SDMA_AHG_APPLY_UPDATE3;
3244 	tx->num_desc++;
3245 	/* initialize to consumed descriptors to zero */
3246 	switch (mode) {
3247 	case SDMA_AHG_APPLY_UPDATE3:
3248 		tx->num_desc++;
3249 		tx->descs[2].qw[0] = 0;
3250 		tx->descs[2].qw[1] = 0;
3251 		/* FALLTHROUGH */
3252 	case SDMA_AHG_APPLY_UPDATE2:
3253 		tx->num_desc++;
3254 		tx->descs[1].qw[0] = 0;
3255 		tx->descs[1].qw[1] = 0;
3256 		break;
3257 	}
3258 	ahg_hlen >>= 2;
3259 	tx->descs[0].qw[1] |=
3260 		(((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK)
3261 			<< SDMA_DESC1_HEADER_INDEX_SHIFT) |
3262 		(((u64)ahg_hlen & SDMA_DESC1_HEADER_DWS_MASK)
3263 			<< SDMA_DESC1_HEADER_DWS_SHIFT) |
3264 		(((u64)mode & SDMA_DESC1_HEADER_MODE_MASK)
3265 			<< SDMA_DESC1_HEADER_MODE_SHIFT) |
3266 		(((u64)ahg[0] & SDMA_DESC1_HEADER_UPDATE1_MASK)
3267 			<< SDMA_DESC1_HEADER_UPDATE1_SHIFT);
3268 	for (i = 0; i < (num_ahg - 1); i++) {
3269 		if (!shift && !(i & 2))
3270 			desc++;
3271 		tx->descs[desc].qw[!!(i & 2)] |=
3272 			(((u64)ahg[i + 1])
3273 				<< shift);
3274 		shift = (shift + 32) & 63;
3275 	}
3276 }
3277 
3278 /**
3279  * sdma_ahg_alloc - allocate an AHG entry
3280  * @sde: engine to allocate from
3281  *
3282  * Return:
3283  * 0-31 when successful, -EOPNOTSUPP if AHG is not enabled,
3284  * -ENOSPC if an entry is not available
3285  */
3286 int sdma_ahg_alloc(struct sdma_engine *sde)
3287 {
3288 	int nr;
3289 	int oldbit;
3290 
3291 	if (!sde) {
3292 		trace_hfi1_ahg_allocate(sde, -EINVAL);
3293 		return -EINVAL;
3294 	}
3295 	while (1) {
3296 		nr = ffz(ACCESS_ONCE(sde->ahg_bits));
3297 		if (nr > 31) {
3298 			trace_hfi1_ahg_allocate(sde, -ENOSPC);
3299 			return -ENOSPC;
3300 		}
3301 		oldbit = test_and_set_bit(nr, &sde->ahg_bits);
3302 		if (!oldbit)
3303 			break;
3304 		cpu_relax();
3305 	}
3306 	trace_hfi1_ahg_allocate(sde, nr);
3307 	return nr;
3308 }
3309 
3310 /**
3311  * sdma_ahg_free - free an AHG entry
3312  * @sde: engine to return AHG entry
3313  * @ahg_index: index to free
3314  *
3315  * This routine frees the indicate AHG entry.
3316  */
3317 void sdma_ahg_free(struct sdma_engine *sde, int ahg_index)
3318 {
3319 	if (!sde)
3320 		return;
3321 	trace_hfi1_ahg_deallocate(sde, ahg_index);
3322 	if (ahg_index < 0 || ahg_index > 31)
3323 		return;
3324 	clear_bit(ahg_index, &sde->ahg_bits);
3325 }
3326 
3327 /*
3328  * SPC freeze handling for SDMA engines.  Called when the driver knows
3329  * the SPC is going into a freeze but before the freeze is fully
3330  * settled.  Generally an error interrupt.
3331  *
3332  * This event will pull the engine out of running so no more entries can be
3333  * added to the engine's queue.
3334  */
3335 void sdma_freeze_notify(struct hfi1_devdata *dd, int link_down)
3336 {
3337 	int i;
3338 	enum sdma_events event = link_down ? sdma_event_e85_link_down :
3339 					     sdma_event_e80_hw_freeze;
3340 
3341 	/* set up the wait but do not wait here */
3342 	atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3343 
3344 	/* tell all engines to stop running and wait */
3345 	for (i = 0; i < dd->num_sdma; i++)
3346 		sdma_process_event(&dd->per_sdma[i], event);
3347 
3348 	/* sdma_freeze() will wait for all engines to have stopped */
3349 }
3350 
3351 /*
3352  * SPC freeze handling for SDMA engines.  Called when the driver knows
3353  * the SPC is fully frozen.
3354  */
3355 void sdma_freeze(struct hfi1_devdata *dd)
3356 {
3357 	int i;
3358 	int ret;
3359 
3360 	/*
3361 	 * Make sure all engines have moved out of the running state before
3362 	 * continuing.
3363 	 */
3364 	ret = wait_event_interruptible(dd->sdma_unfreeze_wq,
3365 				       atomic_read(&dd->sdma_unfreeze_count) <=
3366 				       0);
3367 	/* interrupted or count is negative, then unloading - just exit */
3368 	if (ret || atomic_read(&dd->sdma_unfreeze_count) < 0)
3369 		return;
3370 
3371 	/* set up the count for the next wait */
3372 	atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3373 
3374 	/* tell all engines that the SPC is frozen, they can start cleaning */
3375 	for (i = 0; i < dd->num_sdma; i++)
3376 		sdma_process_event(&dd->per_sdma[i], sdma_event_e81_hw_frozen);
3377 
3378 	/*
3379 	 * Wait for everyone to finish software clean before exiting.  The
3380 	 * software clean will read engine CSRs, so must be completed before
3381 	 * the next step, which will clear the engine CSRs.
3382 	 */
3383 	(void)wait_event_interruptible(dd->sdma_unfreeze_wq,
3384 				atomic_read(&dd->sdma_unfreeze_count) <= 0);
3385 	/* no need to check results - done no matter what */
3386 }
3387 
3388 /*
3389  * SPC freeze handling for the SDMA engines.  Called after the SPC is unfrozen.
3390  *
3391  * The SPC freeze acts like a SDMA halt and a hardware clean combined.  All
3392  * that is left is a software clean.  We could do it after the SPC is fully
3393  * frozen, but then we'd have to add another state to wait for the unfreeze.
3394  * Instead, just defer the software clean until the unfreeze step.
3395  */
3396 void sdma_unfreeze(struct hfi1_devdata *dd)
3397 {
3398 	int i;
3399 
3400 	/* tell all engines start freeze clean up */
3401 	for (i = 0; i < dd->num_sdma; i++)
3402 		sdma_process_event(&dd->per_sdma[i],
3403 				   sdma_event_e82_hw_unfreeze);
3404 }
3405 
3406 /**
3407  * _sdma_engine_progress_schedule() - schedule progress on engine
3408  * @sde: sdma_engine to schedule progress
3409  *
3410  */
3411 void _sdma_engine_progress_schedule(
3412 	struct sdma_engine *sde)
3413 {
3414 	trace_hfi1_sdma_engine_progress(sde, sde->progress_mask);
3415 	/* assume we have selected a good cpu */
3416 	write_csr(sde->dd,
3417 		  CCE_INT_FORCE + (8 * (IS_SDMA_START / 64)),
3418 		  sde->progress_mask);
3419 }
3420