xref: /openbmc/linux/drivers/infiniband/hw/hfi1/sdma.c (revision 9a29f5fc)
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2015 - 2018 Intel Corporation.
4  */
5 
6 #include <linux/spinlock.h>
7 #include <linux/seqlock.h>
8 #include <linux/netdevice.h>
9 #include <linux/moduleparam.h>
10 #include <linux/bitops.h>
11 #include <linux/timer.h>
12 #include <linux/vmalloc.h>
13 #include <linux/highmem.h>
14 
15 #include "hfi.h"
16 #include "common.h"
17 #include "qp.h"
18 #include "sdma.h"
19 #include "iowait.h"
20 #include "trace.h"
21 
22 /* must be a power of 2 >= 64 <= 32768 */
23 #define SDMA_DESCQ_CNT 2048
24 #define SDMA_DESC_INTR 64
25 #define INVALID_TAIL 0xffff
26 #define SDMA_PAD max_t(size_t, MAX_16B_PADDING, sizeof(u32))
27 
28 static uint sdma_descq_cnt = SDMA_DESCQ_CNT;
29 module_param(sdma_descq_cnt, uint, S_IRUGO);
30 MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries");
31 
32 static uint sdma_idle_cnt = 250;
33 module_param(sdma_idle_cnt, uint, S_IRUGO);
34 MODULE_PARM_DESC(sdma_idle_cnt, "sdma interrupt idle delay (ns,default 250)");
35 
36 uint mod_num_sdma;
37 module_param_named(num_sdma, mod_num_sdma, uint, S_IRUGO);
38 MODULE_PARM_DESC(num_sdma, "Set max number SDMA engines to use");
39 
40 static uint sdma_desct_intr = SDMA_DESC_INTR;
41 module_param_named(desct_intr, sdma_desct_intr, uint, S_IRUGO | S_IWUSR);
42 MODULE_PARM_DESC(desct_intr, "Number of SDMA descriptor before interrupt");
43 
44 #define SDMA_WAIT_BATCH_SIZE 20
45 /* max wait time for a SDMA engine to indicate it has halted */
46 #define SDMA_ERR_HALT_TIMEOUT 10 /* ms */
47 /* all SDMA engine errors that cause a halt */
48 
49 #define SD(name) SEND_DMA_##name
50 #define ALL_SDMA_ENG_HALT_ERRS \
51 	(SD(ENG_ERR_STATUS_SDMA_WRONG_DW_ERR_SMASK) \
52 	| SD(ENG_ERR_STATUS_SDMA_GEN_MISMATCH_ERR_SMASK) \
53 	| SD(ENG_ERR_STATUS_SDMA_TOO_LONG_ERR_SMASK) \
54 	| SD(ENG_ERR_STATUS_SDMA_TAIL_OUT_OF_BOUNDS_ERR_SMASK) \
55 	| SD(ENG_ERR_STATUS_SDMA_FIRST_DESC_ERR_SMASK) \
56 	| SD(ENG_ERR_STATUS_SDMA_MEM_READ_ERR_SMASK) \
57 	| SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK) \
58 	| SD(ENG_ERR_STATUS_SDMA_LENGTH_MISMATCH_ERR_SMASK) \
59 	| SD(ENG_ERR_STATUS_SDMA_PACKET_DESC_OVERFLOW_ERR_SMASK) \
60 	| SD(ENG_ERR_STATUS_SDMA_HEADER_SELECT_ERR_SMASK) \
61 	| SD(ENG_ERR_STATUS_SDMA_HEADER_ADDRESS_ERR_SMASK) \
62 	| SD(ENG_ERR_STATUS_SDMA_HEADER_LENGTH_ERR_SMASK) \
63 	| SD(ENG_ERR_STATUS_SDMA_TIMEOUT_ERR_SMASK) \
64 	| SD(ENG_ERR_STATUS_SDMA_DESC_TABLE_UNC_ERR_SMASK) \
65 	| SD(ENG_ERR_STATUS_SDMA_ASSEMBLY_UNC_ERR_SMASK) \
66 	| SD(ENG_ERR_STATUS_SDMA_PACKET_TRACKING_UNC_ERR_SMASK) \
67 	| SD(ENG_ERR_STATUS_SDMA_HEADER_STORAGE_UNC_ERR_SMASK) \
68 	| SD(ENG_ERR_STATUS_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SMASK))
69 
70 /* sdma_sendctrl operations */
71 #define SDMA_SENDCTRL_OP_ENABLE    BIT(0)
72 #define SDMA_SENDCTRL_OP_INTENABLE BIT(1)
73 #define SDMA_SENDCTRL_OP_HALT      BIT(2)
74 #define SDMA_SENDCTRL_OP_CLEANUP   BIT(3)
75 
76 /* handle long defines */
77 #define SDMA_EGRESS_PACKET_OCCUPANCY_SMASK \
78 SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SMASK
79 #define SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT \
80 SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT
81 
82 static const char * const sdma_state_names[] = {
83 	[sdma_state_s00_hw_down]                = "s00_HwDown",
84 	[sdma_state_s10_hw_start_up_halt_wait]  = "s10_HwStartUpHaltWait",
85 	[sdma_state_s15_hw_start_up_clean_wait] = "s15_HwStartUpCleanWait",
86 	[sdma_state_s20_idle]                   = "s20_Idle",
87 	[sdma_state_s30_sw_clean_up_wait]       = "s30_SwCleanUpWait",
88 	[sdma_state_s40_hw_clean_up_wait]       = "s40_HwCleanUpWait",
89 	[sdma_state_s50_hw_halt_wait]           = "s50_HwHaltWait",
90 	[sdma_state_s60_idle_halt_wait]         = "s60_IdleHaltWait",
91 	[sdma_state_s80_hw_freeze]		= "s80_HwFreeze",
92 	[sdma_state_s82_freeze_sw_clean]	= "s82_FreezeSwClean",
93 	[sdma_state_s99_running]                = "s99_Running",
94 };
95 
96 #ifdef CONFIG_SDMA_VERBOSITY
97 static const char * const sdma_event_names[] = {
98 	[sdma_event_e00_go_hw_down]   = "e00_GoHwDown",
99 	[sdma_event_e10_go_hw_start]  = "e10_GoHwStart",
100 	[sdma_event_e15_hw_halt_done] = "e15_HwHaltDone",
101 	[sdma_event_e25_hw_clean_up_done] = "e25_HwCleanUpDone",
102 	[sdma_event_e30_go_running]   = "e30_GoRunning",
103 	[sdma_event_e40_sw_cleaned]   = "e40_SwCleaned",
104 	[sdma_event_e50_hw_cleaned]   = "e50_HwCleaned",
105 	[sdma_event_e60_hw_halted]    = "e60_HwHalted",
106 	[sdma_event_e70_go_idle]      = "e70_GoIdle",
107 	[sdma_event_e80_hw_freeze]    = "e80_HwFreeze",
108 	[sdma_event_e81_hw_frozen]    = "e81_HwFrozen",
109 	[sdma_event_e82_hw_unfreeze]  = "e82_HwUnfreeze",
110 	[sdma_event_e85_link_down]    = "e85_LinkDown",
111 	[sdma_event_e90_sw_halted]    = "e90_SwHalted",
112 };
113 #endif
114 
115 static const struct sdma_set_state_action sdma_action_table[] = {
116 	[sdma_state_s00_hw_down] = {
117 		.go_s99_running_tofalse = 1,
118 		.op_enable = 0,
119 		.op_intenable = 0,
120 		.op_halt = 0,
121 		.op_cleanup = 0,
122 	},
123 	[sdma_state_s10_hw_start_up_halt_wait] = {
124 		.op_enable = 0,
125 		.op_intenable = 0,
126 		.op_halt = 1,
127 		.op_cleanup = 0,
128 	},
129 	[sdma_state_s15_hw_start_up_clean_wait] = {
130 		.op_enable = 0,
131 		.op_intenable = 1,
132 		.op_halt = 0,
133 		.op_cleanup = 1,
134 	},
135 	[sdma_state_s20_idle] = {
136 		.op_enable = 0,
137 		.op_intenable = 1,
138 		.op_halt = 0,
139 		.op_cleanup = 0,
140 	},
141 	[sdma_state_s30_sw_clean_up_wait] = {
142 		.op_enable = 0,
143 		.op_intenable = 0,
144 		.op_halt = 0,
145 		.op_cleanup = 0,
146 	},
147 	[sdma_state_s40_hw_clean_up_wait] = {
148 		.op_enable = 0,
149 		.op_intenable = 0,
150 		.op_halt = 0,
151 		.op_cleanup = 1,
152 	},
153 	[sdma_state_s50_hw_halt_wait] = {
154 		.op_enable = 0,
155 		.op_intenable = 0,
156 		.op_halt = 0,
157 		.op_cleanup = 0,
158 	},
159 	[sdma_state_s60_idle_halt_wait] = {
160 		.go_s99_running_tofalse = 1,
161 		.op_enable = 0,
162 		.op_intenable = 0,
163 		.op_halt = 1,
164 		.op_cleanup = 0,
165 	},
166 	[sdma_state_s80_hw_freeze] = {
167 		.op_enable = 0,
168 		.op_intenable = 0,
169 		.op_halt = 0,
170 		.op_cleanup = 0,
171 	},
172 	[sdma_state_s82_freeze_sw_clean] = {
173 		.op_enable = 0,
174 		.op_intenable = 0,
175 		.op_halt = 0,
176 		.op_cleanup = 0,
177 	},
178 	[sdma_state_s99_running] = {
179 		.op_enable = 1,
180 		.op_intenable = 1,
181 		.op_halt = 0,
182 		.op_cleanup = 0,
183 		.go_s99_running_totrue = 1,
184 	},
185 };
186 
187 #define SDMA_TAIL_UPDATE_THRESH 0x1F
188 
189 /* declare all statics here rather than keep sorting */
190 static void sdma_complete(struct kref *);
191 static void sdma_finalput(struct sdma_state *);
192 static void sdma_get(struct sdma_state *);
193 static void sdma_hw_clean_up_task(struct tasklet_struct *);
194 static void sdma_put(struct sdma_state *);
195 static void sdma_set_state(struct sdma_engine *, enum sdma_states);
196 static void sdma_start_hw_clean_up(struct sdma_engine *);
197 static void sdma_sw_clean_up_task(struct tasklet_struct *);
198 static void sdma_sendctrl(struct sdma_engine *, unsigned);
199 static void init_sdma_regs(struct sdma_engine *, u32, uint);
200 static void sdma_process_event(
201 	struct sdma_engine *sde,
202 	enum sdma_events event);
203 static void __sdma_process_event(
204 	struct sdma_engine *sde,
205 	enum sdma_events event);
206 static void dump_sdma_state(struct sdma_engine *sde);
207 static void sdma_make_progress(struct sdma_engine *sde, u64 status);
208 static void sdma_desc_avail(struct sdma_engine *sde, uint avail);
209 static void sdma_flush_descq(struct sdma_engine *sde);
210 
211 /**
212  * sdma_state_name() - return state string from enum
213  * @state: state
214  */
215 static const char *sdma_state_name(enum sdma_states state)
216 {
217 	return sdma_state_names[state];
218 }
219 
220 static void sdma_get(struct sdma_state *ss)
221 {
222 	kref_get(&ss->kref);
223 }
224 
225 static void sdma_complete(struct kref *kref)
226 {
227 	struct sdma_state *ss =
228 		container_of(kref, struct sdma_state, kref);
229 
230 	complete(&ss->comp);
231 }
232 
233 static void sdma_put(struct sdma_state *ss)
234 {
235 	kref_put(&ss->kref, sdma_complete);
236 }
237 
238 static void sdma_finalput(struct sdma_state *ss)
239 {
240 	sdma_put(ss);
241 	wait_for_completion(&ss->comp);
242 }
243 
244 static inline void write_sde_csr(
245 	struct sdma_engine *sde,
246 	u32 offset0,
247 	u64 value)
248 {
249 	write_kctxt_csr(sde->dd, sde->this_idx, offset0, value);
250 }
251 
252 static inline u64 read_sde_csr(
253 	struct sdma_engine *sde,
254 	u32 offset0)
255 {
256 	return read_kctxt_csr(sde->dd, sde->this_idx, offset0);
257 }
258 
259 /*
260  * sdma_wait_for_packet_egress() - wait for the VL FIFO occupancy for
261  * sdma engine 'sde' to drop to 0.
262  */
263 static void sdma_wait_for_packet_egress(struct sdma_engine *sde,
264 					int pause)
265 {
266 	u64 off = 8 * sde->this_idx;
267 	struct hfi1_devdata *dd = sde->dd;
268 	int lcnt = 0;
269 	u64 reg_prev;
270 	u64 reg = 0;
271 
272 	while (1) {
273 		reg_prev = reg;
274 		reg = read_csr(dd, off + SEND_EGRESS_SEND_DMA_STATUS);
275 
276 		reg &= SDMA_EGRESS_PACKET_OCCUPANCY_SMASK;
277 		reg >>= SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT;
278 		if (reg == 0)
279 			break;
280 		/* counter is reest if accupancy count changes */
281 		if (reg != reg_prev)
282 			lcnt = 0;
283 		if (lcnt++ > 500) {
284 			/* timed out - bounce the link */
285 			dd_dev_err(dd, "%s: engine %u timeout waiting for packets to egress, remaining count %u, bouncing link\n",
286 				   __func__, sde->this_idx, (u32)reg);
287 			queue_work(dd->pport->link_wq,
288 				   &dd->pport->link_bounce_work);
289 			break;
290 		}
291 		udelay(1);
292 	}
293 }
294 
295 /*
296  * sdma_wait() - wait for packet egress to complete for all SDMA engines,
297  * and pause for credit return.
298  */
299 void sdma_wait(struct hfi1_devdata *dd)
300 {
301 	int i;
302 
303 	for (i = 0; i < dd->num_sdma; i++) {
304 		struct sdma_engine *sde = &dd->per_sdma[i];
305 
306 		sdma_wait_for_packet_egress(sde, 0);
307 	}
308 }
309 
310 static inline void sdma_set_desc_cnt(struct sdma_engine *sde, unsigned cnt)
311 {
312 	u64 reg;
313 
314 	if (!(sde->dd->flags & HFI1_HAS_SDMA_TIMEOUT))
315 		return;
316 	reg = cnt;
317 	reg &= SD(DESC_CNT_CNT_MASK);
318 	reg <<= SD(DESC_CNT_CNT_SHIFT);
319 	write_sde_csr(sde, SD(DESC_CNT), reg);
320 }
321 
322 static inline void complete_tx(struct sdma_engine *sde,
323 			       struct sdma_txreq *tx,
324 			       int res)
325 {
326 	/* protect against complete modifying */
327 	struct iowait *wait = tx->wait;
328 	callback_t complete = tx->complete;
329 
330 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
331 	trace_hfi1_sdma_out_sn(sde, tx->sn);
332 	if (WARN_ON_ONCE(sde->head_sn != tx->sn))
333 		dd_dev_err(sde->dd, "expected %llu got %llu\n",
334 			   sde->head_sn, tx->sn);
335 	sde->head_sn++;
336 #endif
337 	__sdma_txclean(sde->dd, tx);
338 	if (complete)
339 		(*complete)(tx, res);
340 	if (iowait_sdma_dec(wait))
341 		iowait_drain_wakeup(wait);
342 }
343 
344 /*
345  * Complete all the sdma requests with a SDMA_TXREQ_S_ABORTED status
346  *
347  * Depending on timing there can be txreqs in two places:
348  * - in the descq ring
349  * - in the flush list
350  *
351  * To avoid ordering issues the descq ring needs to be flushed
352  * first followed by the flush list.
353  *
354  * This routine is called from two places
355  * - From a work queue item
356  * - Directly from the state machine just before setting the
357  *   state to running
358  *
359  * Must be called with head_lock held
360  *
361  */
362 static void sdma_flush(struct sdma_engine *sde)
363 {
364 	struct sdma_txreq *txp, *txp_next;
365 	LIST_HEAD(flushlist);
366 	unsigned long flags;
367 	uint seq;
368 
369 	/* flush from head to tail */
370 	sdma_flush_descq(sde);
371 	spin_lock_irqsave(&sde->flushlist_lock, flags);
372 	/* copy flush list */
373 	list_splice_init(&sde->flushlist, &flushlist);
374 	spin_unlock_irqrestore(&sde->flushlist_lock, flags);
375 	/* flush from flush list */
376 	list_for_each_entry_safe(txp, txp_next, &flushlist, list)
377 		complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED);
378 	/* wakeup QPs orphaned on the dmawait list */
379 	do {
380 		struct iowait *w, *nw;
381 
382 		seq = read_seqbegin(&sde->waitlock);
383 		if (!list_empty(&sde->dmawait)) {
384 			write_seqlock(&sde->waitlock);
385 			list_for_each_entry_safe(w, nw, &sde->dmawait, list) {
386 				if (w->wakeup) {
387 					w->wakeup(w, SDMA_AVAIL_REASON);
388 					list_del_init(&w->list);
389 				}
390 			}
391 			write_sequnlock(&sde->waitlock);
392 		}
393 	} while (read_seqretry(&sde->waitlock, seq));
394 }
395 
396 /*
397  * Fields a work request for flushing the descq ring
398  * and the flush list
399  *
400  * If the engine has been brought to running during
401  * the scheduling delay, the flush is ignored, assuming
402  * that the process of bringing the engine to running
403  * would have done this flush prior to going to running.
404  *
405  */
406 static void sdma_field_flush(struct work_struct *work)
407 {
408 	unsigned long flags;
409 	struct sdma_engine *sde =
410 		container_of(work, struct sdma_engine, flush_worker);
411 
412 	write_seqlock_irqsave(&sde->head_lock, flags);
413 	if (!__sdma_running(sde))
414 		sdma_flush(sde);
415 	write_sequnlock_irqrestore(&sde->head_lock, flags);
416 }
417 
418 static void sdma_err_halt_wait(struct work_struct *work)
419 {
420 	struct sdma_engine *sde = container_of(work, struct sdma_engine,
421 						err_halt_worker);
422 	u64 statuscsr;
423 	unsigned long timeout;
424 
425 	timeout = jiffies + msecs_to_jiffies(SDMA_ERR_HALT_TIMEOUT);
426 	while (1) {
427 		statuscsr = read_sde_csr(sde, SD(STATUS));
428 		statuscsr &= SD(STATUS_ENG_HALTED_SMASK);
429 		if (statuscsr)
430 			break;
431 		if (time_after(jiffies, timeout)) {
432 			dd_dev_err(sde->dd,
433 				   "SDMA engine %d - timeout waiting for engine to halt\n",
434 				   sde->this_idx);
435 			/*
436 			 * Continue anyway.  This could happen if there was
437 			 * an uncorrectable error in the wrong spot.
438 			 */
439 			break;
440 		}
441 		usleep_range(80, 120);
442 	}
443 
444 	sdma_process_event(sde, sdma_event_e15_hw_halt_done);
445 }
446 
447 static void sdma_err_progress_check_schedule(struct sdma_engine *sde)
448 {
449 	if (!is_bx(sde->dd) && HFI1_CAP_IS_KSET(SDMA_AHG)) {
450 		unsigned index;
451 		struct hfi1_devdata *dd = sde->dd;
452 
453 		for (index = 0; index < dd->num_sdma; index++) {
454 			struct sdma_engine *curr_sdma = &dd->per_sdma[index];
455 
456 			if (curr_sdma != sde)
457 				curr_sdma->progress_check_head =
458 							curr_sdma->descq_head;
459 		}
460 		dd_dev_err(sde->dd,
461 			   "SDMA engine %d - check scheduled\n",
462 				sde->this_idx);
463 		mod_timer(&sde->err_progress_check_timer, jiffies + 10);
464 	}
465 }
466 
467 static void sdma_err_progress_check(struct timer_list *t)
468 {
469 	unsigned index;
470 	struct sdma_engine *sde = from_timer(sde, t, err_progress_check_timer);
471 
472 	dd_dev_err(sde->dd, "SDE progress check event\n");
473 	for (index = 0; index < sde->dd->num_sdma; index++) {
474 		struct sdma_engine *curr_sde = &sde->dd->per_sdma[index];
475 		unsigned long flags;
476 
477 		/* check progress on each engine except the current one */
478 		if (curr_sde == sde)
479 			continue;
480 		/*
481 		 * We must lock interrupts when acquiring sde->lock,
482 		 * to avoid a deadlock if interrupt triggers and spins on
483 		 * the same lock on same CPU
484 		 */
485 		spin_lock_irqsave(&curr_sde->tail_lock, flags);
486 		write_seqlock(&curr_sde->head_lock);
487 
488 		/* skip non-running queues */
489 		if (curr_sde->state.current_state != sdma_state_s99_running) {
490 			write_sequnlock(&curr_sde->head_lock);
491 			spin_unlock_irqrestore(&curr_sde->tail_lock, flags);
492 			continue;
493 		}
494 
495 		if ((curr_sde->descq_head != curr_sde->descq_tail) &&
496 		    (curr_sde->descq_head ==
497 				curr_sde->progress_check_head))
498 			__sdma_process_event(curr_sde,
499 					     sdma_event_e90_sw_halted);
500 		write_sequnlock(&curr_sde->head_lock);
501 		spin_unlock_irqrestore(&curr_sde->tail_lock, flags);
502 	}
503 	schedule_work(&sde->err_halt_worker);
504 }
505 
506 static void sdma_hw_clean_up_task(struct tasklet_struct *t)
507 {
508 	struct sdma_engine *sde = from_tasklet(sde, t,
509 					       sdma_hw_clean_up_task);
510 	u64 statuscsr;
511 
512 	while (1) {
513 #ifdef CONFIG_SDMA_VERBOSITY
514 		dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
515 			   sde->this_idx, slashstrip(__FILE__), __LINE__,
516 			__func__);
517 #endif
518 		statuscsr = read_sde_csr(sde, SD(STATUS));
519 		statuscsr &= SD(STATUS_ENG_CLEANED_UP_SMASK);
520 		if (statuscsr)
521 			break;
522 		udelay(10);
523 	}
524 
525 	sdma_process_event(sde, sdma_event_e25_hw_clean_up_done);
526 }
527 
528 static inline struct sdma_txreq *get_txhead(struct sdma_engine *sde)
529 {
530 	return sde->tx_ring[sde->tx_head & sde->sdma_mask];
531 }
532 
533 /*
534  * flush ring for recovery
535  */
536 static void sdma_flush_descq(struct sdma_engine *sde)
537 {
538 	u16 head, tail;
539 	int progress = 0;
540 	struct sdma_txreq *txp = get_txhead(sde);
541 
542 	/* The reason for some of the complexity of this code is that
543 	 * not all descriptors have corresponding txps.  So, we have to
544 	 * be able to skip over descs until we wander into the range of
545 	 * the next txp on the list.
546 	 */
547 	head = sde->descq_head & sde->sdma_mask;
548 	tail = sde->descq_tail & sde->sdma_mask;
549 	while (head != tail) {
550 		/* advance head, wrap if needed */
551 		head = ++sde->descq_head & sde->sdma_mask;
552 		/* if now past this txp's descs, do the callback */
553 		if (txp && txp->next_descq_idx == head) {
554 			/* remove from list */
555 			sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
556 			complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED);
557 			trace_hfi1_sdma_progress(sde, head, tail, txp);
558 			txp = get_txhead(sde);
559 		}
560 		progress++;
561 	}
562 	if (progress)
563 		sdma_desc_avail(sde, sdma_descq_freecnt(sde));
564 }
565 
566 static void sdma_sw_clean_up_task(struct tasklet_struct *t)
567 {
568 	struct sdma_engine *sde = from_tasklet(sde, t, sdma_sw_clean_up_task);
569 	unsigned long flags;
570 
571 	spin_lock_irqsave(&sde->tail_lock, flags);
572 	write_seqlock(&sde->head_lock);
573 
574 	/*
575 	 * At this point, the following should always be true:
576 	 * - We are halted, so no more descriptors are getting retired.
577 	 * - We are not running, so no one is submitting new work.
578 	 * - Only we can send the e40_sw_cleaned, so we can't start
579 	 *   running again until we say so.  So, the active list and
580 	 *   descq are ours to play with.
581 	 */
582 
583 	/*
584 	 * In the error clean up sequence, software clean must be called
585 	 * before the hardware clean so we can use the hardware head in
586 	 * the progress routine.  A hardware clean or SPC unfreeze will
587 	 * reset the hardware head.
588 	 *
589 	 * Process all retired requests. The progress routine will use the
590 	 * latest physical hardware head - we are not running so speed does
591 	 * not matter.
592 	 */
593 	sdma_make_progress(sde, 0);
594 
595 	sdma_flush(sde);
596 
597 	/*
598 	 * Reset our notion of head and tail.
599 	 * Note that the HW registers have been reset via an earlier
600 	 * clean up.
601 	 */
602 	sde->descq_tail = 0;
603 	sde->descq_head = 0;
604 	sde->desc_avail = sdma_descq_freecnt(sde);
605 	*sde->head_dma = 0;
606 
607 	__sdma_process_event(sde, sdma_event_e40_sw_cleaned);
608 
609 	write_sequnlock(&sde->head_lock);
610 	spin_unlock_irqrestore(&sde->tail_lock, flags);
611 }
612 
613 static void sdma_sw_tear_down(struct sdma_engine *sde)
614 {
615 	struct sdma_state *ss = &sde->state;
616 
617 	/* Releasing this reference means the state machine has stopped. */
618 	sdma_put(ss);
619 
620 	/* stop waiting for all unfreeze events to complete */
621 	atomic_set(&sde->dd->sdma_unfreeze_count, -1);
622 	wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
623 }
624 
625 static void sdma_start_hw_clean_up(struct sdma_engine *sde)
626 {
627 	tasklet_hi_schedule(&sde->sdma_hw_clean_up_task);
628 }
629 
630 static void sdma_set_state(struct sdma_engine *sde,
631 			   enum sdma_states next_state)
632 {
633 	struct sdma_state *ss = &sde->state;
634 	const struct sdma_set_state_action *action = sdma_action_table;
635 	unsigned op = 0;
636 
637 	trace_hfi1_sdma_state(
638 		sde,
639 		sdma_state_names[ss->current_state],
640 		sdma_state_names[next_state]);
641 
642 	/* debugging bookkeeping */
643 	ss->previous_state = ss->current_state;
644 	ss->previous_op = ss->current_op;
645 	ss->current_state = next_state;
646 
647 	if (ss->previous_state != sdma_state_s99_running &&
648 	    next_state == sdma_state_s99_running)
649 		sdma_flush(sde);
650 
651 	if (action[next_state].op_enable)
652 		op |= SDMA_SENDCTRL_OP_ENABLE;
653 
654 	if (action[next_state].op_intenable)
655 		op |= SDMA_SENDCTRL_OP_INTENABLE;
656 
657 	if (action[next_state].op_halt)
658 		op |= SDMA_SENDCTRL_OP_HALT;
659 
660 	if (action[next_state].op_cleanup)
661 		op |= SDMA_SENDCTRL_OP_CLEANUP;
662 
663 	if (action[next_state].go_s99_running_tofalse)
664 		ss->go_s99_running = 0;
665 
666 	if (action[next_state].go_s99_running_totrue)
667 		ss->go_s99_running = 1;
668 
669 	ss->current_op = op;
670 	sdma_sendctrl(sde, ss->current_op);
671 }
672 
673 /**
674  * sdma_get_descq_cnt() - called when device probed
675  *
676  * Return a validated descq count.
677  *
678  * This is currently only used in the verbs initialization to build the tx
679  * list.
680  *
681  * This will probably be deleted in favor of a more scalable approach to
682  * alloc tx's.
683  *
684  */
685 u16 sdma_get_descq_cnt(void)
686 {
687 	u16 count = sdma_descq_cnt;
688 
689 	if (!count)
690 		return SDMA_DESCQ_CNT;
691 	/* count must be a power of 2 greater than 64 and less than
692 	 * 32768.   Otherwise return default.
693 	 */
694 	if (!is_power_of_2(count))
695 		return SDMA_DESCQ_CNT;
696 	if (count < 64 || count > 32768)
697 		return SDMA_DESCQ_CNT;
698 	return count;
699 }
700 
701 /**
702  * sdma_engine_get_vl() - return vl for a given sdma engine
703  * @sde: sdma engine
704  *
705  * This function returns the vl mapped to a given engine, or an error if
706  * the mapping can't be found. The mapping fields are protected by RCU.
707  */
708 int sdma_engine_get_vl(struct sdma_engine *sde)
709 {
710 	struct hfi1_devdata *dd = sde->dd;
711 	struct sdma_vl_map *m;
712 	u8 vl;
713 
714 	if (sde->this_idx >= TXE_NUM_SDMA_ENGINES)
715 		return -EINVAL;
716 
717 	rcu_read_lock();
718 	m = rcu_dereference(dd->sdma_map);
719 	if (unlikely(!m)) {
720 		rcu_read_unlock();
721 		return -EINVAL;
722 	}
723 	vl = m->engine_to_vl[sde->this_idx];
724 	rcu_read_unlock();
725 
726 	return vl;
727 }
728 
729 /**
730  * sdma_select_engine_vl() - select sdma engine
731  * @dd: devdata
732  * @selector: a spreading factor
733  * @vl: this vl
734  *
735  *
736  * This function returns an engine based on the selector and a vl.  The
737  * mapping fields are protected by RCU.
738  */
739 struct sdma_engine *sdma_select_engine_vl(
740 	struct hfi1_devdata *dd,
741 	u32 selector,
742 	u8 vl)
743 {
744 	struct sdma_vl_map *m;
745 	struct sdma_map_elem *e;
746 	struct sdma_engine *rval;
747 
748 	/* NOTE This should only happen if SC->VL changed after the initial
749 	 *      checks on the QP/AH
750 	 *      Default will return engine 0 below
751 	 */
752 	if (vl >= num_vls) {
753 		rval = NULL;
754 		goto done;
755 	}
756 
757 	rcu_read_lock();
758 	m = rcu_dereference(dd->sdma_map);
759 	if (unlikely(!m)) {
760 		rcu_read_unlock();
761 		return &dd->per_sdma[0];
762 	}
763 	e = m->map[vl & m->mask];
764 	rval = e->sde[selector & e->mask];
765 	rcu_read_unlock();
766 
767 done:
768 	rval =  !rval ? &dd->per_sdma[0] : rval;
769 	trace_hfi1_sdma_engine_select(dd, selector, vl, rval->this_idx);
770 	return rval;
771 }
772 
773 /**
774  * sdma_select_engine_sc() - select sdma engine
775  * @dd: devdata
776  * @selector: a spreading factor
777  * @sc5: the 5 bit sc
778  *
779  *
780  * This function returns an engine based on the selector and an sc.
781  */
782 struct sdma_engine *sdma_select_engine_sc(
783 	struct hfi1_devdata *dd,
784 	u32 selector,
785 	u8 sc5)
786 {
787 	u8 vl = sc_to_vlt(dd, sc5);
788 
789 	return sdma_select_engine_vl(dd, selector, vl);
790 }
791 
792 struct sdma_rht_map_elem {
793 	u32 mask;
794 	u8 ctr;
795 	struct sdma_engine *sde[];
796 };
797 
798 struct sdma_rht_node {
799 	unsigned long cpu_id;
800 	struct sdma_rht_map_elem *map[HFI1_MAX_VLS_SUPPORTED];
801 	struct rhash_head node;
802 };
803 
804 #define NR_CPUS_HINT 192
805 
806 static const struct rhashtable_params sdma_rht_params = {
807 	.nelem_hint = NR_CPUS_HINT,
808 	.head_offset = offsetof(struct sdma_rht_node, node),
809 	.key_offset = offsetof(struct sdma_rht_node, cpu_id),
810 	.key_len = sizeof_field(struct sdma_rht_node, cpu_id),
811 	.max_size = NR_CPUS,
812 	.min_size = 8,
813 	.automatic_shrinking = true,
814 };
815 
816 /*
817  * sdma_select_user_engine() - select sdma engine based on user setup
818  * @dd: devdata
819  * @selector: a spreading factor
820  * @vl: this vl
821  *
822  * This function returns an sdma engine for a user sdma request.
823  * User defined sdma engine affinity setting is honored when applicable,
824  * otherwise system default sdma engine mapping is used. To ensure correct
825  * ordering, the mapping from <selector, vl> to sde must remain unchanged.
826  */
827 struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd,
828 					    u32 selector, u8 vl)
829 {
830 	struct sdma_rht_node *rht_node;
831 	struct sdma_engine *sde = NULL;
832 	unsigned long cpu_id;
833 
834 	/*
835 	 * To ensure that always the same sdma engine(s) will be
836 	 * selected make sure the process is pinned to this CPU only.
837 	 */
838 	if (current->nr_cpus_allowed != 1)
839 		goto out;
840 
841 	rcu_read_lock();
842 	cpu_id = smp_processor_id();
843 	rht_node = rhashtable_lookup(dd->sdma_rht, &cpu_id,
844 				     sdma_rht_params);
845 
846 	if (rht_node && rht_node->map[vl]) {
847 		struct sdma_rht_map_elem *map = rht_node->map[vl];
848 
849 		sde = map->sde[selector & map->mask];
850 	}
851 	rcu_read_unlock();
852 
853 	if (sde)
854 		return sde;
855 
856 out:
857 	return sdma_select_engine_vl(dd, selector, vl);
858 }
859 
860 static void sdma_populate_sde_map(struct sdma_rht_map_elem *map)
861 {
862 	int i;
863 
864 	for (i = 0; i < roundup_pow_of_two(map->ctr ? : 1) - map->ctr; i++)
865 		map->sde[map->ctr + i] = map->sde[i];
866 }
867 
868 static void sdma_cleanup_sde_map(struct sdma_rht_map_elem *map,
869 				 struct sdma_engine *sde)
870 {
871 	unsigned int i, pow;
872 
873 	/* only need to check the first ctr entries for a match */
874 	for (i = 0; i < map->ctr; i++) {
875 		if (map->sde[i] == sde) {
876 			memmove(&map->sde[i], &map->sde[i + 1],
877 				(map->ctr - i - 1) * sizeof(map->sde[0]));
878 			map->ctr--;
879 			pow = roundup_pow_of_two(map->ctr ? : 1);
880 			map->mask = pow - 1;
881 			sdma_populate_sde_map(map);
882 			break;
883 		}
884 	}
885 }
886 
887 /*
888  * Prevents concurrent reads and writes of the sdma engine cpu_mask
889  */
890 static DEFINE_MUTEX(process_to_sde_mutex);
891 
892 ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf,
893 				size_t count)
894 {
895 	struct hfi1_devdata *dd = sde->dd;
896 	cpumask_var_t mask, new_mask;
897 	unsigned long cpu;
898 	int ret, vl, sz;
899 	struct sdma_rht_node *rht_node;
900 
901 	vl = sdma_engine_get_vl(sde);
902 	if (unlikely(vl < 0 || vl >= ARRAY_SIZE(rht_node->map)))
903 		return -EINVAL;
904 
905 	ret = zalloc_cpumask_var(&mask, GFP_KERNEL);
906 	if (!ret)
907 		return -ENOMEM;
908 
909 	ret = zalloc_cpumask_var(&new_mask, GFP_KERNEL);
910 	if (!ret) {
911 		free_cpumask_var(mask);
912 		return -ENOMEM;
913 	}
914 	ret = cpulist_parse(buf, mask);
915 	if (ret)
916 		goto out_free;
917 
918 	if (!cpumask_subset(mask, cpu_online_mask)) {
919 		dd_dev_warn(sde->dd, "Invalid CPU mask\n");
920 		ret = -EINVAL;
921 		goto out_free;
922 	}
923 
924 	sz = sizeof(struct sdma_rht_map_elem) +
925 			(TXE_NUM_SDMA_ENGINES * sizeof(struct sdma_engine *));
926 
927 	mutex_lock(&process_to_sde_mutex);
928 
929 	for_each_cpu(cpu, mask) {
930 		/* Check if we have this already mapped */
931 		if (cpumask_test_cpu(cpu, &sde->cpu_mask)) {
932 			cpumask_set_cpu(cpu, new_mask);
933 			continue;
934 		}
935 
936 		rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu,
937 						  sdma_rht_params);
938 		if (!rht_node) {
939 			rht_node = kzalloc(sizeof(*rht_node), GFP_KERNEL);
940 			if (!rht_node) {
941 				ret = -ENOMEM;
942 				goto out;
943 			}
944 
945 			rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
946 			if (!rht_node->map[vl]) {
947 				kfree(rht_node);
948 				ret = -ENOMEM;
949 				goto out;
950 			}
951 			rht_node->cpu_id = cpu;
952 			rht_node->map[vl]->mask = 0;
953 			rht_node->map[vl]->ctr = 1;
954 			rht_node->map[vl]->sde[0] = sde;
955 
956 			ret = rhashtable_insert_fast(dd->sdma_rht,
957 						     &rht_node->node,
958 						     sdma_rht_params);
959 			if (ret) {
960 				kfree(rht_node->map[vl]);
961 				kfree(rht_node);
962 				dd_dev_err(sde->dd, "Failed to set process to sde affinity for cpu %lu\n",
963 					   cpu);
964 				goto out;
965 			}
966 
967 		} else {
968 			int ctr, pow;
969 
970 			/* Add new user mappings */
971 			if (!rht_node->map[vl])
972 				rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
973 
974 			if (!rht_node->map[vl]) {
975 				ret = -ENOMEM;
976 				goto out;
977 			}
978 
979 			rht_node->map[vl]->ctr++;
980 			ctr = rht_node->map[vl]->ctr;
981 			rht_node->map[vl]->sde[ctr - 1] = sde;
982 			pow = roundup_pow_of_two(ctr);
983 			rht_node->map[vl]->mask = pow - 1;
984 
985 			/* Populate the sde map table */
986 			sdma_populate_sde_map(rht_node->map[vl]);
987 		}
988 		cpumask_set_cpu(cpu, new_mask);
989 	}
990 
991 	/* Clean up old mappings */
992 	for_each_cpu(cpu, cpu_online_mask) {
993 		struct sdma_rht_node *rht_node;
994 
995 		/* Don't cleanup sdes that are set in the new mask */
996 		if (cpumask_test_cpu(cpu, mask))
997 			continue;
998 
999 		rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu,
1000 						  sdma_rht_params);
1001 		if (rht_node) {
1002 			bool empty = true;
1003 			int i;
1004 
1005 			/* Remove mappings for old sde */
1006 			for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1007 				if (rht_node->map[i])
1008 					sdma_cleanup_sde_map(rht_node->map[i],
1009 							     sde);
1010 
1011 			/* Free empty hash table entries */
1012 			for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1013 				if (!rht_node->map[i])
1014 					continue;
1015 
1016 				if (rht_node->map[i]->ctr) {
1017 					empty = false;
1018 					break;
1019 				}
1020 			}
1021 
1022 			if (empty) {
1023 				ret = rhashtable_remove_fast(dd->sdma_rht,
1024 							     &rht_node->node,
1025 							     sdma_rht_params);
1026 				WARN_ON(ret);
1027 
1028 				for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1029 					kfree(rht_node->map[i]);
1030 
1031 				kfree(rht_node);
1032 			}
1033 		}
1034 	}
1035 
1036 	cpumask_copy(&sde->cpu_mask, new_mask);
1037 out:
1038 	mutex_unlock(&process_to_sde_mutex);
1039 out_free:
1040 	free_cpumask_var(mask);
1041 	free_cpumask_var(new_mask);
1042 	return ret ? : strnlen(buf, PAGE_SIZE);
1043 }
1044 
1045 ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf)
1046 {
1047 	mutex_lock(&process_to_sde_mutex);
1048 	if (cpumask_empty(&sde->cpu_mask))
1049 		snprintf(buf, PAGE_SIZE, "%s\n", "empty");
1050 	else
1051 		cpumap_print_to_pagebuf(true, buf, &sde->cpu_mask);
1052 	mutex_unlock(&process_to_sde_mutex);
1053 	return strnlen(buf, PAGE_SIZE);
1054 }
1055 
1056 static void sdma_rht_free(void *ptr, void *arg)
1057 {
1058 	struct sdma_rht_node *rht_node = ptr;
1059 	int i;
1060 
1061 	for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1062 		kfree(rht_node->map[i]);
1063 
1064 	kfree(rht_node);
1065 }
1066 
1067 /**
1068  * sdma_seqfile_dump_cpu_list() - debugfs dump the cpu to sdma mappings
1069  * @s: seq file
1070  * @dd: hfi1_devdata
1071  * @cpuid: cpu id
1072  *
1073  * This routine dumps the process to sde mappings per cpu
1074  */
1075 void sdma_seqfile_dump_cpu_list(struct seq_file *s,
1076 				struct hfi1_devdata *dd,
1077 				unsigned long cpuid)
1078 {
1079 	struct sdma_rht_node *rht_node;
1080 	int i, j;
1081 
1082 	rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpuid,
1083 					  sdma_rht_params);
1084 	if (!rht_node)
1085 		return;
1086 
1087 	seq_printf(s, "cpu%3lu: ", cpuid);
1088 	for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1089 		if (!rht_node->map[i] || !rht_node->map[i]->ctr)
1090 			continue;
1091 
1092 		seq_printf(s, " vl%d: [", i);
1093 
1094 		for (j = 0; j < rht_node->map[i]->ctr; j++) {
1095 			if (!rht_node->map[i]->sde[j])
1096 				continue;
1097 
1098 			if (j > 0)
1099 				seq_puts(s, ",");
1100 
1101 			seq_printf(s, " sdma%2d",
1102 				   rht_node->map[i]->sde[j]->this_idx);
1103 		}
1104 		seq_puts(s, " ]");
1105 	}
1106 
1107 	seq_puts(s, "\n");
1108 }
1109 
1110 /*
1111  * Free the indicated map struct
1112  */
1113 static void sdma_map_free(struct sdma_vl_map *m)
1114 {
1115 	int i;
1116 
1117 	for (i = 0; m && i < m->actual_vls; i++)
1118 		kfree(m->map[i]);
1119 	kfree(m);
1120 }
1121 
1122 /*
1123  * Handle RCU callback
1124  */
1125 static void sdma_map_rcu_callback(struct rcu_head *list)
1126 {
1127 	struct sdma_vl_map *m = container_of(list, struct sdma_vl_map, list);
1128 
1129 	sdma_map_free(m);
1130 }
1131 
1132 /**
1133  * sdma_map_init - called when # vls change
1134  * @dd: hfi1_devdata
1135  * @port: port number
1136  * @num_vls: number of vls
1137  * @vl_engines: per vl engine mapping (optional)
1138  *
1139  * This routine changes the mapping based on the number of vls.
1140  *
1141  * vl_engines is used to specify a non-uniform vl/engine loading. NULL
1142  * implies auto computing the loading and giving each VLs a uniform
1143  * distribution of engines per VL.
1144  *
1145  * The auto algorithm computes the sde_per_vl and the number of extra
1146  * engines.  Any extra engines are added from the last VL on down.
1147  *
1148  * rcu locking is used here to control access to the mapping fields.
1149  *
1150  * If either the num_vls or num_sdma are non-power of 2, the array sizes
1151  * in the struct sdma_vl_map and the struct sdma_map_elem are rounded
1152  * up to the next highest power of 2 and the first entry is reused
1153  * in a round robin fashion.
1154  *
1155  * If an error occurs the map change is not done and the mapping is
1156  * not changed.
1157  *
1158  */
1159 int sdma_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_engines)
1160 {
1161 	int i, j;
1162 	int extra, sde_per_vl;
1163 	int engine = 0;
1164 	u8 lvl_engines[OPA_MAX_VLS];
1165 	struct sdma_vl_map *oldmap, *newmap;
1166 
1167 	if (!(dd->flags & HFI1_HAS_SEND_DMA))
1168 		return 0;
1169 
1170 	if (!vl_engines) {
1171 		/* truncate divide */
1172 		sde_per_vl = dd->num_sdma / num_vls;
1173 		/* extras */
1174 		extra = dd->num_sdma % num_vls;
1175 		vl_engines = lvl_engines;
1176 		/* add extras from last vl down */
1177 		for (i = num_vls - 1; i >= 0; i--, extra--)
1178 			vl_engines[i] = sde_per_vl + (extra > 0 ? 1 : 0);
1179 	}
1180 	/* build new map */
1181 	newmap = kzalloc(
1182 		sizeof(struct sdma_vl_map) +
1183 			roundup_pow_of_two(num_vls) *
1184 			sizeof(struct sdma_map_elem *),
1185 		GFP_KERNEL);
1186 	if (!newmap)
1187 		goto bail;
1188 	newmap->actual_vls = num_vls;
1189 	newmap->vls = roundup_pow_of_two(num_vls);
1190 	newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1191 	/* initialize back-map */
1192 	for (i = 0; i < TXE_NUM_SDMA_ENGINES; i++)
1193 		newmap->engine_to_vl[i] = -1;
1194 	for (i = 0; i < newmap->vls; i++) {
1195 		/* save for wrap around */
1196 		int first_engine = engine;
1197 
1198 		if (i < newmap->actual_vls) {
1199 			int sz = roundup_pow_of_two(vl_engines[i]);
1200 
1201 			/* only allocate once */
1202 			newmap->map[i] = kzalloc(
1203 				sizeof(struct sdma_map_elem) +
1204 					sz * sizeof(struct sdma_engine *),
1205 				GFP_KERNEL);
1206 			if (!newmap->map[i])
1207 				goto bail;
1208 			newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1209 			/* assign engines */
1210 			for (j = 0; j < sz; j++) {
1211 				newmap->map[i]->sde[j] =
1212 					&dd->per_sdma[engine];
1213 				if (++engine >= first_engine + vl_engines[i])
1214 					/* wrap back to first engine */
1215 					engine = first_engine;
1216 			}
1217 			/* assign back-map */
1218 			for (j = 0; j < vl_engines[i]; j++)
1219 				newmap->engine_to_vl[first_engine + j] = i;
1220 		} else {
1221 			/* just re-use entry without allocating */
1222 			newmap->map[i] = newmap->map[i % num_vls];
1223 		}
1224 		engine = first_engine + vl_engines[i];
1225 	}
1226 	/* newmap in hand, save old map */
1227 	spin_lock_irq(&dd->sde_map_lock);
1228 	oldmap = rcu_dereference_protected(dd->sdma_map,
1229 					   lockdep_is_held(&dd->sde_map_lock));
1230 
1231 	/* publish newmap */
1232 	rcu_assign_pointer(dd->sdma_map, newmap);
1233 
1234 	spin_unlock_irq(&dd->sde_map_lock);
1235 	/* success, free any old map after grace period */
1236 	if (oldmap)
1237 		call_rcu(&oldmap->list, sdma_map_rcu_callback);
1238 	return 0;
1239 bail:
1240 	/* free any partial allocation */
1241 	sdma_map_free(newmap);
1242 	return -ENOMEM;
1243 }
1244 
1245 /**
1246  * sdma_clean - Clean up allocated memory
1247  * @dd:          struct hfi1_devdata
1248  * @num_engines: num sdma engines
1249  *
1250  * This routine can be called regardless of the success of
1251  * sdma_init()
1252  */
1253 void sdma_clean(struct hfi1_devdata *dd, size_t num_engines)
1254 {
1255 	size_t i;
1256 	struct sdma_engine *sde;
1257 
1258 	if (dd->sdma_pad_dma) {
1259 		dma_free_coherent(&dd->pcidev->dev, SDMA_PAD,
1260 				  (void *)dd->sdma_pad_dma,
1261 				  dd->sdma_pad_phys);
1262 		dd->sdma_pad_dma = NULL;
1263 		dd->sdma_pad_phys = 0;
1264 	}
1265 	if (dd->sdma_heads_dma) {
1266 		dma_free_coherent(&dd->pcidev->dev, dd->sdma_heads_size,
1267 				  (void *)dd->sdma_heads_dma,
1268 				  dd->sdma_heads_phys);
1269 		dd->sdma_heads_dma = NULL;
1270 		dd->sdma_heads_phys = 0;
1271 	}
1272 	for (i = 0; dd->per_sdma && i < num_engines; ++i) {
1273 		sde = &dd->per_sdma[i];
1274 
1275 		sde->head_dma = NULL;
1276 		sde->head_phys = 0;
1277 
1278 		if (sde->descq) {
1279 			dma_free_coherent(
1280 				&dd->pcidev->dev,
1281 				sde->descq_cnt * sizeof(u64[2]),
1282 				sde->descq,
1283 				sde->descq_phys
1284 			);
1285 			sde->descq = NULL;
1286 			sde->descq_phys = 0;
1287 		}
1288 		kvfree(sde->tx_ring);
1289 		sde->tx_ring = NULL;
1290 	}
1291 	if (rcu_access_pointer(dd->sdma_map)) {
1292 		spin_lock_irq(&dd->sde_map_lock);
1293 		sdma_map_free(rcu_access_pointer(dd->sdma_map));
1294 		RCU_INIT_POINTER(dd->sdma_map, NULL);
1295 		spin_unlock_irq(&dd->sde_map_lock);
1296 		synchronize_rcu();
1297 	}
1298 	kfree(dd->per_sdma);
1299 	dd->per_sdma = NULL;
1300 
1301 	if (dd->sdma_rht) {
1302 		rhashtable_free_and_destroy(dd->sdma_rht, sdma_rht_free, NULL);
1303 		kfree(dd->sdma_rht);
1304 		dd->sdma_rht = NULL;
1305 	}
1306 }
1307 
1308 /**
1309  * sdma_init() - called when device probed
1310  * @dd: hfi1_devdata
1311  * @port: port number (currently only zero)
1312  *
1313  * Initializes each sde and its csrs.
1314  * Interrupts are not required to be enabled.
1315  *
1316  * Returns:
1317  * 0 - success, -errno on failure
1318  */
1319 int sdma_init(struct hfi1_devdata *dd, u8 port)
1320 {
1321 	unsigned this_idx;
1322 	struct sdma_engine *sde;
1323 	struct rhashtable *tmp_sdma_rht;
1324 	u16 descq_cnt;
1325 	void *curr_head;
1326 	struct hfi1_pportdata *ppd = dd->pport + port;
1327 	u32 per_sdma_credits;
1328 	uint idle_cnt = sdma_idle_cnt;
1329 	size_t num_engines = chip_sdma_engines(dd);
1330 	int ret = -ENOMEM;
1331 
1332 	if (!HFI1_CAP_IS_KSET(SDMA)) {
1333 		HFI1_CAP_CLEAR(SDMA_AHG);
1334 		return 0;
1335 	}
1336 	if (mod_num_sdma &&
1337 	    /* can't exceed chip support */
1338 	    mod_num_sdma <= chip_sdma_engines(dd) &&
1339 	    /* count must be >= vls */
1340 	    mod_num_sdma >= num_vls)
1341 		num_engines = mod_num_sdma;
1342 
1343 	dd_dev_info(dd, "SDMA mod_num_sdma: %u\n", mod_num_sdma);
1344 	dd_dev_info(dd, "SDMA chip_sdma_engines: %u\n", chip_sdma_engines(dd));
1345 	dd_dev_info(dd, "SDMA chip_sdma_mem_size: %u\n",
1346 		    chip_sdma_mem_size(dd));
1347 
1348 	per_sdma_credits =
1349 		chip_sdma_mem_size(dd) / (num_engines * SDMA_BLOCK_SIZE);
1350 
1351 	/* set up freeze waitqueue */
1352 	init_waitqueue_head(&dd->sdma_unfreeze_wq);
1353 	atomic_set(&dd->sdma_unfreeze_count, 0);
1354 
1355 	descq_cnt = sdma_get_descq_cnt();
1356 	dd_dev_info(dd, "SDMA engines %zu descq_cnt %u\n",
1357 		    num_engines, descq_cnt);
1358 
1359 	/* alloc memory for array of send engines */
1360 	dd->per_sdma = kcalloc_node(num_engines, sizeof(*dd->per_sdma),
1361 				    GFP_KERNEL, dd->node);
1362 	if (!dd->per_sdma)
1363 		return ret;
1364 
1365 	idle_cnt = ns_to_cclock(dd, idle_cnt);
1366 	if (idle_cnt)
1367 		dd->default_desc1 =
1368 			SDMA_DESC1_HEAD_TO_HOST_FLAG;
1369 	else
1370 		dd->default_desc1 =
1371 			SDMA_DESC1_INT_REQ_FLAG;
1372 
1373 	if (!sdma_desct_intr)
1374 		sdma_desct_intr = SDMA_DESC_INTR;
1375 
1376 	/* Allocate memory for SendDMA descriptor FIFOs */
1377 	for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1378 		sde = &dd->per_sdma[this_idx];
1379 		sde->dd = dd;
1380 		sde->ppd = ppd;
1381 		sde->this_idx = this_idx;
1382 		sde->descq_cnt = descq_cnt;
1383 		sde->desc_avail = sdma_descq_freecnt(sde);
1384 		sde->sdma_shift = ilog2(descq_cnt);
1385 		sde->sdma_mask = (1 << sde->sdma_shift) - 1;
1386 
1387 		/* Create a mask specifically for each interrupt source */
1388 		sde->int_mask = (u64)1 << (0 * TXE_NUM_SDMA_ENGINES +
1389 					   this_idx);
1390 		sde->progress_mask = (u64)1 << (1 * TXE_NUM_SDMA_ENGINES +
1391 						this_idx);
1392 		sde->idle_mask = (u64)1 << (2 * TXE_NUM_SDMA_ENGINES +
1393 					    this_idx);
1394 		/* Create a combined mask to cover all 3 interrupt sources */
1395 		sde->imask = sde->int_mask | sde->progress_mask |
1396 			     sde->idle_mask;
1397 
1398 		spin_lock_init(&sde->tail_lock);
1399 		seqlock_init(&sde->head_lock);
1400 		spin_lock_init(&sde->senddmactrl_lock);
1401 		spin_lock_init(&sde->flushlist_lock);
1402 		seqlock_init(&sde->waitlock);
1403 		/* insure there is always a zero bit */
1404 		sde->ahg_bits = 0xfffffffe00000000ULL;
1405 
1406 		sdma_set_state(sde, sdma_state_s00_hw_down);
1407 
1408 		/* set up reference counting */
1409 		kref_init(&sde->state.kref);
1410 		init_completion(&sde->state.comp);
1411 
1412 		INIT_LIST_HEAD(&sde->flushlist);
1413 		INIT_LIST_HEAD(&sde->dmawait);
1414 
1415 		sde->tail_csr =
1416 			get_kctxt_csr_addr(dd, this_idx, SD(TAIL));
1417 
1418 		tasklet_setup(&sde->sdma_hw_clean_up_task,
1419 			      sdma_hw_clean_up_task);
1420 		tasklet_setup(&sde->sdma_sw_clean_up_task,
1421 			      sdma_sw_clean_up_task);
1422 		INIT_WORK(&sde->err_halt_worker, sdma_err_halt_wait);
1423 		INIT_WORK(&sde->flush_worker, sdma_field_flush);
1424 
1425 		sde->progress_check_head = 0;
1426 
1427 		timer_setup(&sde->err_progress_check_timer,
1428 			    sdma_err_progress_check, 0);
1429 
1430 		sde->descq = dma_alloc_coherent(&dd->pcidev->dev,
1431 						descq_cnt * sizeof(u64[2]),
1432 						&sde->descq_phys, GFP_KERNEL);
1433 		if (!sde->descq)
1434 			goto bail;
1435 		sde->tx_ring =
1436 			kvzalloc_node(array_size(descq_cnt,
1437 						 sizeof(struct sdma_txreq *)),
1438 				      GFP_KERNEL, dd->node);
1439 		if (!sde->tx_ring)
1440 			goto bail;
1441 	}
1442 
1443 	dd->sdma_heads_size = L1_CACHE_BYTES * num_engines;
1444 	/* Allocate memory for DMA of head registers to memory */
1445 	dd->sdma_heads_dma = dma_alloc_coherent(&dd->pcidev->dev,
1446 						dd->sdma_heads_size,
1447 						&dd->sdma_heads_phys,
1448 						GFP_KERNEL);
1449 	if (!dd->sdma_heads_dma) {
1450 		dd_dev_err(dd, "failed to allocate SendDMA head memory\n");
1451 		goto bail;
1452 	}
1453 
1454 	/* Allocate memory for pad */
1455 	dd->sdma_pad_dma = dma_alloc_coherent(&dd->pcidev->dev, SDMA_PAD,
1456 					      &dd->sdma_pad_phys, GFP_KERNEL);
1457 	if (!dd->sdma_pad_dma) {
1458 		dd_dev_err(dd, "failed to allocate SendDMA pad memory\n");
1459 		goto bail;
1460 	}
1461 
1462 	/* assign each engine to different cacheline and init registers */
1463 	curr_head = (void *)dd->sdma_heads_dma;
1464 	for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1465 		unsigned long phys_offset;
1466 
1467 		sde = &dd->per_sdma[this_idx];
1468 
1469 		sde->head_dma = curr_head;
1470 		curr_head += L1_CACHE_BYTES;
1471 		phys_offset = (unsigned long)sde->head_dma -
1472 			      (unsigned long)dd->sdma_heads_dma;
1473 		sde->head_phys = dd->sdma_heads_phys + phys_offset;
1474 		init_sdma_regs(sde, per_sdma_credits, idle_cnt);
1475 	}
1476 	dd->flags |= HFI1_HAS_SEND_DMA;
1477 	dd->flags |= idle_cnt ? HFI1_HAS_SDMA_TIMEOUT : 0;
1478 	dd->num_sdma = num_engines;
1479 	ret = sdma_map_init(dd, port, ppd->vls_operational, NULL);
1480 	if (ret < 0)
1481 		goto bail;
1482 
1483 	tmp_sdma_rht = kzalloc(sizeof(*tmp_sdma_rht), GFP_KERNEL);
1484 	if (!tmp_sdma_rht) {
1485 		ret = -ENOMEM;
1486 		goto bail;
1487 	}
1488 
1489 	ret = rhashtable_init(tmp_sdma_rht, &sdma_rht_params);
1490 	if (ret < 0) {
1491 		kfree(tmp_sdma_rht);
1492 		goto bail;
1493 	}
1494 
1495 	dd->sdma_rht = tmp_sdma_rht;
1496 
1497 	dd_dev_info(dd, "SDMA num_sdma: %u\n", dd->num_sdma);
1498 	return 0;
1499 
1500 bail:
1501 	sdma_clean(dd, num_engines);
1502 	return ret;
1503 }
1504 
1505 /**
1506  * sdma_all_running() - called when the link goes up
1507  * @dd: hfi1_devdata
1508  *
1509  * This routine moves all engines to the running state.
1510  */
1511 void sdma_all_running(struct hfi1_devdata *dd)
1512 {
1513 	struct sdma_engine *sde;
1514 	unsigned int i;
1515 
1516 	/* move all engines to running */
1517 	for (i = 0; i < dd->num_sdma; ++i) {
1518 		sde = &dd->per_sdma[i];
1519 		sdma_process_event(sde, sdma_event_e30_go_running);
1520 	}
1521 }
1522 
1523 /**
1524  * sdma_all_idle() - called when the link goes down
1525  * @dd: hfi1_devdata
1526  *
1527  * This routine moves all engines to the idle state.
1528  */
1529 void sdma_all_idle(struct hfi1_devdata *dd)
1530 {
1531 	struct sdma_engine *sde;
1532 	unsigned int i;
1533 
1534 	/* idle all engines */
1535 	for (i = 0; i < dd->num_sdma; ++i) {
1536 		sde = &dd->per_sdma[i];
1537 		sdma_process_event(sde, sdma_event_e70_go_idle);
1538 	}
1539 }
1540 
1541 /**
1542  * sdma_start() - called to kick off state processing for all engines
1543  * @dd: hfi1_devdata
1544  *
1545  * This routine is for kicking off the state processing for all required
1546  * sdma engines.  Interrupts need to be working at this point.
1547  *
1548  */
1549 void sdma_start(struct hfi1_devdata *dd)
1550 {
1551 	unsigned i;
1552 	struct sdma_engine *sde;
1553 
1554 	/* kick off the engines state processing */
1555 	for (i = 0; i < dd->num_sdma; ++i) {
1556 		sde = &dd->per_sdma[i];
1557 		sdma_process_event(sde, sdma_event_e10_go_hw_start);
1558 	}
1559 }
1560 
1561 /**
1562  * sdma_exit() - used when module is removed
1563  * @dd: hfi1_devdata
1564  */
1565 void sdma_exit(struct hfi1_devdata *dd)
1566 {
1567 	unsigned this_idx;
1568 	struct sdma_engine *sde;
1569 
1570 	for (this_idx = 0; dd->per_sdma && this_idx < dd->num_sdma;
1571 			++this_idx) {
1572 		sde = &dd->per_sdma[this_idx];
1573 		if (!list_empty(&sde->dmawait))
1574 			dd_dev_err(dd, "sde %u: dmawait list not empty!\n",
1575 				   sde->this_idx);
1576 		sdma_process_event(sde, sdma_event_e00_go_hw_down);
1577 
1578 		del_timer_sync(&sde->err_progress_check_timer);
1579 
1580 		/*
1581 		 * This waits for the state machine to exit so it is not
1582 		 * necessary to kill the sdma_sw_clean_up_task to make sure
1583 		 * it is not running.
1584 		 */
1585 		sdma_finalput(&sde->state);
1586 	}
1587 }
1588 
1589 /*
1590  * unmap the indicated descriptor
1591  */
1592 static inline void sdma_unmap_desc(
1593 	struct hfi1_devdata *dd,
1594 	struct sdma_desc *descp)
1595 {
1596 	switch (sdma_mapping_type(descp)) {
1597 	case SDMA_MAP_SINGLE:
1598 		dma_unmap_single(
1599 			&dd->pcidev->dev,
1600 			sdma_mapping_addr(descp),
1601 			sdma_mapping_len(descp),
1602 			DMA_TO_DEVICE);
1603 		break;
1604 	case SDMA_MAP_PAGE:
1605 		dma_unmap_page(
1606 			&dd->pcidev->dev,
1607 			sdma_mapping_addr(descp),
1608 			sdma_mapping_len(descp),
1609 			DMA_TO_DEVICE);
1610 		break;
1611 	}
1612 }
1613 
1614 /*
1615  * return the mode as indicated by the first
1616  * descriptor in the tx.
1617  */
1618 static inline u8 ahg_mode(struct sdma_txreq *tx)
1619 {
1620 	return (tx->descp[0].qw[1] & SDMA_DESC1_HEADER_MODE_SMASK)
1621 		>> SDMA_DESC1_HEADER_MODE_SHIFT;
1622 }
1623 
1624 /**
1625  * __sdma_txclean() - clean tx of mappings, descp *kmalloc's
1626  * @dd: hfi1_devdata for unmapping
1627  * @tx: tx request to clean
1628  *
1629  * This is used in the progress routine to clean the tx or
1630  * by the ULP to toss an in-process tx build.
1631  *
1632  * The code can be called multiple times without issue.
1633  *
1634  */
1635 void __sdma_txclean(
1636 	struct hfi1_devdata *dd,
1637 	struct sdma_txreq *tx)
1638 {
1639 	u16 i;
1640 
1641 	if (tx->num_desc) {
1642 		u8 skip = 0, mode = ahg_mode(tx);
1643 
1644 		/* unmap first */
1645 		sdma_unmap_desc(dd, &tx->descp[0]);
1646 		/* determine number of AHG descriptors to skip */
1647 		if (mode > SDMA_AHG_APPLY_UPDATE1)
1648 			skip = mode >> 1;
1649 		for (i = 1 + skip; i < tx->num_desc; i++)
1650 			sdma_unmap_desc(dd, &tx->descp[i]);
1651 		tx->num_desc = 0;
1652 	}
1653 	kfree(tx->coalesce_buf);
1654 	tx->coalesce_buf = NULL;
1655 	/* kmalloc'ed descp */
1656 	if (unlikely(tx->desc_limit > ARRAY_SIZE(tx->descs))) {
1657 		tx->desc_limit = ARRAY_SIZE(tx->descs);
1658 		kfree(tx->descp);
1659 	}
1660 }
1661 
1662 static inline u16 sdma_gethead(struct sdma_engine *sde)
1663 {
1664 	struct hfi1_devdata *dd = sde->dd;
1665 	int use_dmahead;
1666 	u16 hwhead;
1667 
1668 #ifdef CONFIG_SDMA_VERBOSITY
1669 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1670 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1671 #endif
1672 
1673 retry:
1674 	use_dmahead = HFI1_CAP_IS_KSET(USE_SDMA_HEAD) && __sdma_running(sde) &&
1675 					(dd->flags & HFI1_HAS_SDMA_TIMEOUT);
1676 	hwhead = use_dmahead ?
1677 		(u16)le64_to_cpu(*sde->head_dma) :
1678 		(u16)read_sde_csr(sde, SD(HEAD));
1679 
1680 	if (unlikely(HFI1_CAP_IS_KSET(SDMA_HEAD_CHECK))) {
1681 		u16 cnt;
1682 		u16 swtail;
1683 		u16 swhead;
1684 		int sane;
1685 
1686 		swhead = sde->descq_head & sde->sdma_mask;
1687 		/* this code is really bad for cache line trading */
1688 		swtail = READ_ONCE(sde->descq_tail) & sde->sdma_mask;
1689 		cnt = sde->descq_cnt;
1690 
1691 		if (swhead < swtail)
1692 			/* not wrapped */
1693 			sane = (hwhead >= swhead) & (hwhead <= swtail);
1694 		else if (swhead > swtail)
1695 			/* wrapped around */
1696 			sane = ((hwhead >= swhead) && (hwhead < cnt)) ||
1697 				(hwhead <= swtail);
1698 		else
1699 			/* empty */
1700 			sane = (hwhead == swhead);
1701 
1702 		if (unlikely(!sane)) {
1703 			dd_dev_err(dd, "SDMA(%u) bad head (%s) hwhd=%u swhd=%u swtl=%u cnt=%u\n",
1704 				   sde->this_idx,
1705 				   use_dmahead ? "dma" : "kreg",
1706 				   hwhead, swhead, swtail, cnt);
1707 			if (use_dmahead) {
1708 				/* try one more time, using csr */
1709 				use_dmahead = 0;
1710 				goto retry;
1711 			}
1712 			/* proceed as if no progress */
1713 			hwhead = swhead;
1714 		}
1715 	}
1716 	return hwhead;
1717 }
1718 
1719 /*
1720  * This is called when there are send DMA descriptors that might be
1721  * available.
1722  *
1723  * This is called with head_lock held.
1724  */
1725 static void sdma_desc_avail(struct sdma_engine *sde, uint avail)
1726 {
1727 	struct iowait *wait, *nw, *twait;
1728 	struct iowait *waits[SDMA_WAIT_BATCH_SIZE];
1729 	uint i, n = 0, seq, tidx = 0;
1730 
1731 #ifdef CONFIG_SDMA_VERBOSITY
1732 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", sde->this_idx,
1733 		   slashstrip(__FILE__), __LINE__, __func__);
1734 	dd_dev_err(sde->dd, "avail: %u\n", avail);
1735 #endif
1736 
1737 	do {
1738 		seq = read_seqbegin(&sde->waitlock);
1739 		if (!list_empty(&sde->dmawait)) {
1740 			/* at least one item */
1741 			write_seqlock(&sde->waitlock);
1742 			/* Harvest waiters wanting DMA descriptors */
1743 			list_for_each_entry_safe(
1744 					wait,
1745 					nw,
1746 					&sde->dmawait,
1747 					list) {
1748 				u32 num_desc;
1749 
1750 				if (!wait->wakeup)
1751 					continue;
1752 				if (n == ARRAY_SIZE(waits))
1753 					break;
1754 				iowait_init_priority(wait);
1755 				num_desc = iowait_get_all_desc(wait);
1756 				if (num_desc > avail)
1757 					break;
1758 				avail -= num_desc;
1759 				/* Find the top-priority wait memeber */
1760 				if (n) {
1761 					twait = waits[tidx];
1762 					tidx =
1763 					    iowait_priority_update_top(wait,
1764 								       twait,
1765 								       n,
1766 								       tidx);
1767 				}
1768 				list_del_init(&wait->list);
1769 				waits[n++] = wait;
1770 			}
1771 			write_sequnlock(&sde->waitlock);
1772 			break;
1773 		}
1774 	} while (read_seqretry(&sde->waitlock, seq));
1775 
1776 	/* Schedule the top-priority entry first */
1777 	if (n)
1778 		waits[tidx]->wakeup(waits[tidx], SDMA_AVAIL_REASON);
1779 
1780 	for (i = 0; i < n; i++)
1781 		if (i != tidx)
1782 			waits[i]->wakeup(waits[i], SDMA_AVAIL_REASON);
1783 }
1784 
1785 /* head_lock must be held */
1786 static void sdma_make_progress(struct sdma_engine *sde, u64 status)
1787 {
1788 	struct sdma_txreq *txp = NULL;
1789 	int progress = 0;
1790 	u16 hwhead, swhead;
1791 	int idle_check_done = 0;
1792 
1793 	hwhead = sdma_gethead(sde);
1794 
1795 	/* The reason for some of the complexity of this code is that
1796 	 * not all descriptors have corresponding txps.  So, we have to
1797 	 * be able to skip over descs until we wander into the range of
1798 	 * the next txp on the list.
1799 	 */
1800 
1801 retry:
1802 	txp = get_txhead(sde);
1803 	swhead = sde->descq_head & sde->sdma_mask;
1804 	trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1805 	while (swhead != hwhead) {
1806 		/* advance head, wrap if needed */
1807 		swhead = ++sde->descq_head & sde->sdma_mask;
1808 
1809 		/* if now past this txp's descs, do the callback */
1810 		if (txp && txp->next_descq_idx == swhead) {
1811 			/* remove from list */
1812 			sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
1813 			complete_tx(sde, txp, SDMA_TXREQ_S_OK);
1814 			/* see if there is another txp */
1815 			txp = get_txhead(sde);
1816 		}
1817 		trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1818 		progress++;
1819 	}
1820 
1821 	/*
1822 	 * The SDMA idle interrupt is not guaranteed to be ordered with respect
1823 	 * to updates to the dma_head location in host memory. The head
1824 	 * value read might not be fully up to date. If there are pending
1825 	 * descriptors and the SDMA idle interrupt fired then read from the
1826 	 * CSR SDMA head instead to get the latest value from the hardware.
1827 	 * The hardware SDMA head should be read at most once in this invocation
1828 	 * of sdma_make_progress(..) which is ensured by idle_check_done flag
1829 	 */
1830 	if ((status & sde->idle_mask) && !idle_check_done) {
1831 		u16 swtail;
1832 
1833 		swtail = READ_ONCE(sde->descq_tail) & sde->sdma_mask;
1834 		if (swtail != hwhead) {
1835 			hwhead = (u16)read_sde_csr(sde, SD(HEAD));
1836 			idle_check_done = 1;
1837 			goto retry;
1838 		}
1839 	}
1840 
1841 	sde->last_status = status;
1842 	if (progress)
1843 		sdma_desc_avail(sde, sdma_descq_freecnt(sde));
1844 }
1845 
1846 /*
1847  * sdma_engine_interrupt() - interrupt handler for engine
1848  * @sde: sdma engine
1849  * @status: sdma interrupt reason
1850  *
1851  * Status is a mask of the 3 possible interrupts for this engine.  It will
1852  * contain bits _only_ for this SDMA engine.  It will contain at least one
1853  * bit, it may contain more.
1854  */
1855 void sdma_engine_interrupt(struct sdma_engine *sde, u64 status)
1856 {
1857 	trace_hfi1_sdma_engine_interrupt(sde, status);
1858 	write_seqlock(&sde->head_lock);
1859 	sdma_set_desc_cnt(sde, sdma_desct_intr);
1860 	if (status & sde->idle_mask)
1861 		sde->idle_int_cnt++;
1862 	else if (status & sde->progress_mask)
1863 		sde->progress_int_cnt++;
1864 	else if (status & sde->int_mask)
1865 		sde->sdma_int_cnt++;
1866 	sdma_make_progress(sde, status);
1867 	write_sequnlock(&sde->head_lock);
1868 }
1869 
1870 /**
1871  * sdma_engine_error() - error handler for engine
1872  * @sde: sdma engine
1873  * @status: sdma interrupt reason
1874  */
1875 void sdma_engine_error(struct sdma_engine *sde, u64 status)
1876 {
1877 	unsigned long flags;
1878 
1879 #ifdef CONFIG_SDMA_VERBOSITY
1880 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) error status 0x%llx state %s\n",
1881 		   sde->this_idx,
1882 		   (unsigned long long)status,
1883 		   sdma_state_names[sde->state.current_state]);
1884 #endif
1885 	spin_lock_irqsave(&sde->tail_lock, flags);
1886 	write_seqlock(&sde->head_lock);
1887 	if (status & ALL_SDMA_ENG_HALT_ERRS)
1888 		__sdma_process_event(sde, sdma_event_e60_hw_halted);
1889 	if (status & ~SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK)) {
1890 		dd_dev_err(sde->dd,
1891 			   "SDMA (%u) engine error: 0x%llx state %s\n",
1892 			   sde->this_idx,
1893 			   (unsigned long long)status,
1894 			   sdma_state_names[sde->state.current_state]);
1895 		dump_sdma_state(sde);
1896 	}
1897 	write_sequnlock(&sde->head_lock);
1898 	spin_unlock_irqrestore(&sde->tail_lock, flags);
1899 }
1900 
1901 static void sdma_sendctrl(struct sdma_engine *sde, unsigned op)
1902 {
1903 	u64 set_senddmactrl = 0;
1904 	u64 clr_senddmactrl = 0;
1905 	unsigned long flags;
1906 
1907 #ifdef CONFIG_SDMA_VERBOSITY
1908 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) senddmactrl E=%d I=%d H=%d C=%d\n",
1909 		   sde->this_idx,
1910 		   (op & SDMA_SENDCTRL_OP_ENABLE) ? 1 : 0,
1911 		   (op & SDMA_SENDCTRL_OP_INTENABLE) ? 1 : 0,
1912 		   (op & SDMA_SENDCTRL_OP_HALT) ? 1 : 0,
1913 		   (op & SDMA_SENDCTRL_OP_CLEANUP) ? 1 : 0);
1914 #endif
1915 
1916 	if (op & SDMA_SENDCTRL_OP_ENABLE)
1917 		set_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1918 	else
1919 		clr_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1920 
1921 	if (op & SDMA_SENDCTRL_OP_INTENABLE)
1922 		set_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1923 	else
1924 		clr_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1925 
1926 	if (op & SDMA_SENDCTRL_OP_HALT)
1927 		set_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1928 	else
1929 		clr_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1930 
1931 	spin_lock_irqsave(&sde->senddmactrl_lock, flags);
1932 
1933 	sde->p_senddmactrl |= set_senddmactrl;
1934 	sde->p_senddmactrl &= ~clr_senddmactrl;
1935 
1936 	if (op & SDMA_SENDCTRL_OP_CLEANUP)
1937 		write_sde_csr(sde, SD(CTRL),
1938 			      sde->p_senddmactrl |
1939 			      SD(CTRL_SDMA_CLEANUP_SMASK));
1940 	else
1941 		write_sde_csr(sde, SD(CTRL), sde->p_senddmactrl);
1942 
1943 	spin_unlock_irqrestore(&sde->senddmactrl_lock, flags);
1944 
1945 #ifdef CONFIG_SDMA_VERBOSITY
1946 	sdma_dumpstate(sde);
1947 #endif
1948 }
1949 
1950 static void sdma_setlengen(struct sdma_engine *sde)
1951 {
1952 #ifdef CONFIG_SDMA_VERBOSITY
1953 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1954 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1955 #endif
1956 
1957 	/*
1958 	 * Set SendDmaLenGen and clear-then-set the MSB of the generation
1959 	 * count to enable generation checking and load the internal
1960 	 * generation counter.
1961 	 */
1962 	write_sde_csr(sde, SD(LEN_GEN),
1963 		      (sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT));
1964 	write_sde_csr(sde, SD(LEN_GEN),
1965 		      ((sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT)) |
1966 		      (4ULL << SD(LEN_GEN_GENERATION_SHIFT)));
1967 }
1968 
1969 static inline void sdma_update_tail(struct sdma_engine *sde, u16 tail)
1970 {
1971 	/* Commit writes to memory and advance the tail on the chip */
1972 	smp_wmb(); /* see get_txhead() */
1973 	writeq(tail, sde->tail_csr);
1974 }
1975 
1976 /*
1977  * This is called when changing to state s10_hw_start_up_halt_wait as
1978  * a result of send buffer errors or send DMA descriptor errors.
1979  */
1980 static void sdma_hw_start_up(struct sdma_engine *sde)
1981 {
1982 	u64 reg;
1983 
1984 #ifdef CONFIG_SDMA_VERBOSITY
1985 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1986 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1987 #endif
1988 
1989 	sdma_setlengen(sde);
1990 	sdma_update_tail(sde, 0); /* Set SendDmaTail */
1991 	*sde->head_dma = 0;
1992 
1993 	reg = SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_MASK) <<
1994 	      SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SHIFT);
1995 	write_sde_csr(sde, SD(ENG_ERR_CLEAR), reg);
1996 }
1997 
1998 /*
1999  * set_sdma_integrity
2000  *
2001  * Set the SEND_DMA_CHECK_ENABLE register for send DMA engine 'sde'.
2002  */
2003 static void set_sdma_integrity(struct sdma_engine *sde)
2004 {
2005 	struct hfi1_devdata *dd = sde->dd;
2006 
2007 	write_sde_csr(sde, SD(CHECK_ENABLE),
2008 		      hfi1_pkt_base_sdma_integrity(dd));
2009 }
2010 
2011 static void init_sdma_regs(
2012 	struct sdma_engine *sde,
2013 	u32 credits,
2014 	uint idle_cnt)
2015 {
2016 	u8 opval, opmask;
2017 #ifdef CONFIG_SDMA_VERBOSITY
2018 	struct hfi1_devdata *dd = sde->dd;
2019 
2020 	dd_dev_err(dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2021 		   sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2022 #endif
2023 
2024 	write_sde_csr(sde, SD(BASE_ADDR), sde->descq_phys);
2025 	sdma_setlengen(sde);
2026 	sdma_update_tail(sde, 0); /* Set SendDmaTail */
2027 	write_sde_csr(sde, SD(RELOAD_CNT), idle_cnt);
2028 	write_sde_csr(sde, SD(DESC_CNT), 0);
2029 	write_sde_csr(sde, SD(HEAD_ADDR), sde->head_phys);
2030 	write_sde_csr(sde, SD(MEMORY),
2031 		      ((u64)credits << SD(MEMORY_SDMA_MEMORY_CNT_SHIFT)) |
2032 		      ((u64)(credits * sde->this_idx) <<
2033 		       SD(MEMORY_SDMA_MEMORY_INDEX_SHIFT)));
2034 	write_sde_csr(sde, SD(ENG_ERR_MASK), ~0ull);
2035 	set_sdma_integrity(sde);
2036 	opmask = OPCODE_CHECK_MASK_DISABLED;
2037 	opval = OPCODE_CHECK_VAL_DISABLED;
2038 	write_sde_csr(sde, SD(CHECK_OPCODE),
2039 		      (opmask << SEND_CTXT_CHECK_OPCODE_MASK_SHIFT) |
2040 		      (opval << SEND_CTXT_CHECK_OPCODE_VALUE_SHIFT));
2041 }
2042 
2043 #ifdef CONFIG_SDMA_VERBOSITY
2044 
2045 #define sdma_dumpstate_helper0(reg) do { \
2046 		csr = read_csr(sde->dd, reg); \
2047 		dd_dev_err(sde->dd, "%36s     0x%016llx\n", #reg, csr); \
2048 	} while (0)
2049 
2050 #define sdma_dumpstate_helper(reg) do { \
2051 		csr = read_sde_csr(sde, reg); \
2052 		dd_dev_err(sde->dd, "%36s[%02u] 0x%016llx\n", \
2053 			#reg, sde->this_idx, csr); \
2054 	} while (0)
2055 
2056 #define sdma_dumpstate_helper2(reg) do { \
2057 		csr = read_csr(sde->dd, reg + (8 * i)); \
2058 		dd_dev_err(sde->dd, "%33s_%02u     0x%016llx\n", \
2059 				#reg, i, csr); \
2060 	} while (0)
2061 
2062 void sdma_dumpstate(struct sdma_engine *sde)
2063 {
2064 	u64 csr;
2065 	unsigned i;
2066 
2067 	sdma_dumpstate_helper(SD(CTRL));
2068 	sdma_dumpstate_helper(SD(STATUS));
2069 	sdma_dumpstate_helper0(SD(ERR_STATUS));
2070 	sdma_dumpstate_helper0(SD(ERR_MASK));
2071 	sdma_dumpstate_helper(SD(ENG_ERR_STATUS));
2072 	sdma_dumpstate_helper(SD(ENG_ERR_MASK));
2073 
2074 	for (i = 0; i < CCE_NUM_INT_CSRS; ++i) {
2075 		sdma_dumpstate_helper2(CCE_INT_STATUS);
2076 		sdma_dumpstate_helper2(CCE_INT_MASK);
2077 		sdma_dumpstate_helper2(CCE_INT_BLOCKED);
2078 	}
2079 
2080 	sdma_dumpstate_helper(SD(TAIL));
2081 	sdma_dumpstate_helper(SD(HEAD));
2082 	sdma_dumpstate_helper(SD(PRIORITY_THLD));
2083 	sdma_dumpstate_helper(SD(IDLE_CNT));
2084 	sdma_dumpstate_helper(SD(RELOAD_CNT));
2085 	sdma_dumpstate_helper(SD(DESC_CNT));
2086 	sdma_dumpstate_helper(SD(DESC_FETCHED_CNT));
2087 	sdma_dumpstate_helper(SD(MEMORY));
2088 	sdma_dumpstate_helper0(SD(ENGINES));
2089 	sdma_dumpstate_helper0(SD(MEM_SIZE));
2090 	/* sdma_dumpstate_helper(SEND_EGRESS_SEND_DMA_STATUS);  */
2091 	sdma_dumpstate_helper(SD(BASE_ADDR));
2092 	sdma_dumpstate_helper(SD(LEN_GEN));
2093 	sdma_dumpstate_helper(SD(HEAD_ADDR));
2094 	sdma_dumpstate_helper(SD(CHECK_ENABLE));
2095 	sdma_dumpstate_helper(SD(CHECK_VL));
2096 	sdma_dumpstate_helper(SD(CHECK_JOB_KEY));
2097 	sdma_dumpstate_helper(SD(CHECK_PARTITION_KEY));
2098 	sdma_dumpstate_helper(SD(CHECK_SLID));
2099 	sdma_dumpstate_helper(SD(CHECK_OPCODE));
2100 }
2101 #endif
2102 
2103 static void dump_sdma_state(struct sdma_engine *sde)
2104 {
2105 	struct hw_sdma_desc *descqp;
2106 	u64 desc[2];
2107 	u64 addr;
2108 	u8 gen;
2109 	u16 len;
2110 	u16 head, tail, cnt;
2111 
2112 	head = sde->descq_head & sde->sdma_mask;
2113 	tail = sde->descq_tail & sde->sdma_mask;
2114 	cnt = sdma_descq_freecnt(sde);
2115 
2116 	dd_dev_err(sde->dd,
2117 		   "SDMA (%u) descq_head: %u descq_tail: %u freecnt: %u FLE %d\n",
2118 		   sde->this_idx, head, tail, cnt,
2119 		   !list_empty(&sde->flushlist));
2120 
2121 	/* print info for each entry in the descriptor queue */
2122 	while (head != tail) {
2123 		char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2124 
2125 		descqp = &sde->descq[head];
2126 		desc[0] = le64_to_cpu(descqp->qw[0]);
2127 		desc[1] = le64_to_cpu(descqp->qw[1]);
2128 		flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2129 		flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2130 				'H' : '-';
2131 		flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2132 		flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2133 		addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2134 			& SDMA_DESC0_PHY_ADDR_MASK;
2135 		gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2136 			& SDMA_DESC1_GENERATION_MASK;
2137 		len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2138 			& SDMA_DESC0_BYTE_COUNT_MASK;
2139 		dd_dev_err(sde->dd,
2140 			   "SDMA sdmadesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2141 			   head, flags, addr, gen, len);
2142 		dd_dev_err(sde->dd,
2143 			   "\tdesc0:0x%016llx desc1 0x%016llx\n",
2144 			   desc[0], desc[1]);
2145 		if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2146 			dd_dev_err(sde->dd,
2147 				   "\taidx: %u amode: %u alen: %u\n",
2148 				   (u8)((desc[1] &
2149 					 SDMA_DESC1_HEADER_INDEX_SMASK) >>
2150 					SDMA_DESC1_HEADER_INDEX_SHIFT),
2151 				   (u8)((desc[1] &
2152 					 SDMA_DESC1_HEADER_MODE_SMASK) >>
2153 					SDMA_DESC1_HEADER_MODE_SHIFT),
2154 				   (u8)((desc[1] &
2155 					 SDMA_DESC1_HEADER_DWS_SMASK) >>
2156 					SDMA_DESC1_HEADER_DWS_SHIFT));
2157 		head++;
2158 		head &= sde->sdma_mask;
2159 	}
2160 }
2161 
2162 #define SDE_FMT \
2163 	"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"
2164 /**
2165  * sdma_seqfile_dump_sde() - debugfs dump of sde
2166  * @s: seq file
2167  * @sde: send dma engine to dump
2168  *
2169  * This routine dumps the sde to the indicated seq file.
2170  */
2171 void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *sde)
2172 {
2173 	u16 head, tail;
2174 	struct hw_sdma_desc *descqp;
2175 	u64 desc[2];
2176 	u64 addr;
2177 	u8 gen;
2178 	u16 len;
2179 
2180 	head = sde->descq_head & sde->sdma_mask;
2181 	tail = READ_ONCE(sde->descq_tail) & sde->sdma_mask;
2182 	seq_printf(s, SDE_FMT, sde->this_idx,
2183 		   sde->cpu,
2184 		   sdma_state_name(sde->state.current_state),
2185 		   (unsigned long long)read_sde_csr(sde, SD(CTRL)),
2186 		   (unsigned long long)read_sde_csr(sde, SD(STATUS)),
2187 		   (unsigned long long)read_sde_csr(sde, SD(ENG_ERR_STATUS)),
2188 		   (unsigned long long)read_sde_csr(sde, SD(TAIL)), tail,
2189 		   (unsigned long long)read_sde_csr(sde, SD(HEAD)), head,
2190 		   (unsigned long long)le64_to_cpu(*sde->head_dma),
2191 		   (unsigned long long)read_sde_csr(sde, SD(MEMORY)),
2192 		   (unsigned long long)read_sde_csr(sde, SD(LEN_GEN)),
2193 		   (unsigned long long)read_sde_csr(sde, SD(RELOAD_CNT)),
2194 		   (unsigned long long)sde->last_status,
2195 		   (unsigned long long)sde->ahg_bits,
2196 		   sde->tx_tail,
2197 		   sde->tx_head,
2198 		   sde->descq_tail,
2199 		   sde->descq_head,
2200 		   !list_empty(&sde->flushlist),
2201 		   sde->descq_full_count,
2202 		   (unsigned long long)read_sde_csr(sde, SEND_DMA_CHECK_SLID));
2203 
2204 	/* print info for each entry in the descriptor queue */
2205 	while (head != tail) {
2206 		char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2207 
2208 		descqp = &sde->descq[head];
2209 		desc[0] = le64_to_cpu(descqp->qw[0]);
2210 		desc[1] = le64_to_cpu(descqp->qw[1]);
2211 		flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2212 		flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2213 				'H' : '-';
2214 		flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2215 		flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2216 		addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2217 			& SDMA_DESC0_PHY_ADDR_MASK;
2218 		gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2219 			& SDMA_DESC1_GENERATION_MASK;
2220 		len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2221 			& SDMA_DESC0_BYTE_COUNT_MASK;
2222 		seq_printf(s,
2223 			   "\tdesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2224 			   head, flags, addr, gen, len);
2225 		if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2226 			seq_printf(s, "\t\tahgidx: %u ahgmode: %u\n",
2227 				   (u8)((desc[1] &
2228 					 SDMA_DESC1_HEADER_INDEX_SMASK) >>
2229 					SDMA_DESC1_HEADER_INDEX_SHIFT),
2230 				   (u8)((desc[1] &
2231 					 SDMA_DESC1_HEADER_MODE_SMASK) >>
2232 					SDMA_DESC1_HEADER_MODE_SHIFT));
2233 		head = (head + 1) & sde->sdma_mask;
2234 	}
2235 }
2236 
2237 /*
2238  * add the generation number into
2239  * the qw1 and return
2240  */
2241 static inline u64 add_gen(struct sdma_engine *sde, u64 qw1)
2242 {
2243 	u8 generation = (sde->descq_tail >> sde->sdma_shift) & 3;
2244 
2245 	qw1 &= ~SDMA_DESC1_GENERATION_SMASK;
2246 	qw1 |= ((u64)generation & SDMA_DESC1_GENERATION_MASK)
2247 			<< SDMA_DESC1_GENERATION_SHIFT;
2248 	return qw1;
2249 }
2250 
2251 /*
2252  * This routine submits the indicated tx
2253  *
2254  * Space has already been guaranteed and
2255  * tail side of ring is locked.
2256  *
2257  * The hardware tail update is done
2258  * in the caller and that is facilitated
2259  * by returning the new tail.
2260  *
2261  * There is special case logic for ahg
2262  * to not add the generation number for
2263  * up to 2 descriptors that follow the
2264  * first descriptor.
2265  *
2266  */
2267 static inline u16 submit_tx(struct sdma_engine *sde, struct sdma_txreq *tx)
2268 {
2269 	int i;
2270 	u16 tail;
2271 	struct sdma_desc *descp = tx->descp;
2272 	u8 skip = 0, mode = ahg_mode(tx);
2273 
2274 	tail = sde->descq_tail & sde->sdma_mask;
2275 	sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2276 	sde->descq[tail].qw[1] = cpu_to_le64(add_gen(sde, descp->qw[1]));
2277 	trace_hfi1_sdma_descriptor(sde, descp->qw[0], descp->qw[1],
2278 				   tail, &sde->descq[tail]);
2279 	tail = ++sde->descq_tail & sde->sdma_mask;
2280 	descp++;
2281 	if (mode > SDMA_AHG_APPLY_UPDATE1)
2282 		skip = mode >> 1;
2283 	for (i = 1; i < tx->num_desc; i++, descp++) {
2284 		u64 qw1;
2285 
2286 		sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2287 		if (skip) {
2288 			/* edits don't have generation */
2289 			qw1 = descp->qw[1];
2290 			skip--;
2291 		} else {
2292 			/* replace generation with real one for non-edits */
2293 			qw1 = add_gen(sde, descp->qw[1]);
2294 		}
2295 		sde->descq[tail].qw[1] = cpu_to_le64(qw1);
2296 		trace_hfi1_sdma_descriptor(sde, descp->qw[0], qw1,
2297 					   tail, &sde->descq[tail]);
2298 		tail = ++sde->descq_tail & sde->sdma_mask;
2299 	}
2300 	tx->next_descq_idx = tail;
2301 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2302 	tx->sn = sde->tail_sn++;
2303 	trace_hfi1_sdma_in_sn(sde, tx->sn);
2304 	WARN_ON_ONCE(sde->tx_ring[sde->tx_tail & sde->sdma_mask]);
2305 #endif
2306 	sde->tx_ring[sde->tx_tail++ & sde->sdma_mask] = tx;
2307 	sde->desc_avail -= tx->num_desc;
2308 	return tail;
2309 }
2310 
2311 /*
2312  * Check for progress
2313  */
2314 static int sdma_check_progress(
2315 	struct sdma_engine *sde,
2316 	struct iowait_work *wait,
2317 	struct sdma_txreq *tx,
2318 	bool pkts_sent)
2319 {
2320 	int ret;
2321 
2322 	sde->desc_avail = sdma_descq_freecnt(sde);
2323 	if (tx->num_desc <= sde->desc_avail)
2324 		return -EAGAIN;
2325 	/* pulse the head_lock */
2326 	if (wait && iowait_ioww_to_iow(wait)->sleep) {
2327 		unsigned seq;
2328 
2329 		seq = raw_seqcount_begin(
2330 			(const seqcount_t *)&sde->head_lock.seqcount);
2331 		ret = wait->iow->sleep(sde, wait, tx, seq, pkts_sent);
2332 		if (ret == -EAGAIN)
2333 			sde->desc_avail = sdma_descq_freecnt(sde);
2334 	} else {
2335 		ret = -EBUSY;
2336 	}
2337 	return ret;
2338 }
2339 
2340 /**
2341  * sdma_send_txreq() - submit a tx req to ring
2342  * @sde: sdma engine to use
2343  * @wait: SE wait structure to use when full (may be NULL)
2344  * @tx: sdma_txreq to submit
2345  * @pkts_sent: has any packet been sent yet?
2346  *
2347  * The call submits the tx into the ring.  If a iowait structure is non-NULL
2348  * the packet will be queued to the list in wait.
2349  *
2350  * Return:
2351  * 0 - Success, -EINVAL - sdma_txreq incomplete, -EBUSY - no space in
2352  * ring (wait == NULL)
2353  * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2354  */
2355 int sdma_send_txreq(struct sdma_engine *sde,
2356 		    struct iowait_work *wait,
2357 		    struct sdma_txreq *tx,
2358 		    bool pkts_sent)
2359 {
2360 	int ret = 0;
2361 	u16 tail;
2362 	unsigned long flags;
2363 
2364 	/* user should have supplied entire packet */
2365 	if (unlikely(tx->tlen))
2366 		return -EINVAL;
2367 	tx->wait = iowait_ioww_to_iow(wait);
2368 	spin_lock_irqsave(&sde->tail_lock, flags);
2369 retry:
2370 	if (unlikely(!__sdma_running(sde)))
2371 		goto unlock_noconn;
2372 	if (unlikely(tx->num_desc > sde->desc_avail))
2373 		goto nodesc;
2374 	tail = submit_tx(sde, tx);
2375 	if (wait)
2376 		iowait_sdma_inc(iowait_ioww_to_iow(wait));
2377 	sdma_update_tail(sde, tail);
2378 unlock:
2379 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2380 	return ret;
2381 unlock_noconn:
2382 	if (wait)
2383 		iowait_sdma_inc(iowait_ioww_to_iow(wait));
2384 	tx->next_descq_idx = 0;
2385 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2386 	tx->sn = sde->tail_sn++;
2387 	trace_hfi1_sdma_in_sn(sde, tx->sn);
2388 #endif
2389 	spin_lock(&sde->flushlist_lock);
2390 	list_add_tail(&tx->list, &sde->flushlist);
2391 	spin_unlock(&sde->flushlist_lock);
2392 	iowait_inc_wait_count(wait, tx->num_desc);
2393 	queue_work_on(sde->cpu, system_highpri_wq, &sde->flush_worker);
2394 	ret = -ECOMM;
2395 	goto unlock;
2396 nodesc:
2397 	ret = sdma_check_progress(sde, wait, tx, pkts_sent);
2398 	if (ret == -EAGAIN) {
2399 		ret = 0;
2400 		goto retry;
2401 	}
2402 	sde->descq_full_count++;
2403 	goto unlock;
2404 }
2405 
2406 /**
2407  * sdma_send_txlist() - submit a list of tx req to ring
2408  * @sde: sdma engine to use
2409  * @wait: SE wait structure to use when full (may be NULL)
2410  * @tx_list: list of sdma_txreqs to submit
2411  * @count_out: pointer to a u16 which, after return will contain the total number of
2412  *             sdma_txreqs removed from the tx_list. This will include sdma_txreqs
2413  *             whose SDMA descriptors are submitted to the ring and the sdma_txreqs
2414  *             which are added to SDMA engine flush list if the SDMA engine state is
2415  *             not running.
2416  *
2417  * The call submits the list into the ring.
2418  *
2419  * If the iowait structure is non-NULL and not equal to the iowait list
2420  * the unprocessed part of the list  will be appended to the list in wait.
2421  *
2422  * In all cases, the tx_list will be updated so the head of the tx_list is
2423  * the list of descriptors that have yet to be transmitted.
2424  *
2425  * The intent of this call is to provide a more efficient
2426  * way of submitting multiple packets to SDMA while holding the tail
2427  * side locking.
2428  *
2429  * Return:
2430  * 0 - Success,
2431  * -EINVAL - sdma_txreq incomplete, -EBUSY - no space in ring (wait == NULL)
2432  * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2433  */
2434 int sdma_send_txlist(struct sdma_engine *sde, struct iowait_work *wait,
2435 		     struct list_head *tx_list, u16 *count_out)
2436 {
2437 	struct sdma_txreq *tx, *tx_next;
2438 	int ret = 0;
2439 	unsigned long flags;
2440 	u16 tail = INVALID_TAIL;
2441 	u32 submit_count = 0, flush_count = 0, total_count;
2442 
2443 	spin_lock_irqsave(&sde->tail_lock, flags);
2444 retry:
2445 	list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2446 		tx->wait = iowait_ioww_to_iow(wait);
2447 		if (unlikely(!__sdma_running(sde)))
2448 			goto unlock_noconn;
2449 		if (unlikely(tx->num_desc > sde->desc_avail))
2450 			goto nodesc;
2451 		if (unlikely(tx->tlen)) {
2452 			ret = -EINVAL;
2453 			goto update_tail;
2454 		}
2455 		list_del_init(&tx->list);
2456 		tail = submit_tx(sde, tx);
2457 		submit_count++;
2458 		if (tail != INVALID_TAIL &&
2459 		    (submit_count & SDMA_TAIL_UPDATE_THRESH) == 0) {
2460 			sdma_update_tail(sde, tail);
2461 			tail = INVALID_TAIL;
2462 		}
2463 	}
2464 update_tail:
2465 	total_count = submit_count + flush_count;
2466 	if (wait) {
2467 		iowait_sdma_add(iowait_ioww_to_iow(wait), total_count);
2468 		iowait_starve_clear(submit_count > 0,
2469 				    iowait_ioww_to_iow(wait));
2470 	}
2471 	if (tail != INVALID_TAIL)
2472 		sdma_update_tail(sde, tail);
2473 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2474 	*count_out = total_count;
2475 	return ret;
2476 unlock_noconn:
2477 	spin_lock(&sde->flushlist_lock);
2478 	list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2479 		tx->wait = iowait_ioww_to_iow(wait);
2480 		list_del_init(&tx->list);
2481 		tx->next_descq_idx = 0;
2482 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2483 		tx->sn = sde->tail_sn++;
2484 		trace_hfi1_sdma_in_sn(sde, tx->sn);
2485 #endif
2486 		list_add_tail(&tx->list, &sde->flushlist);
2487 		flush_count++;
2488 		iowait_inc_wait_count(wait, tx->num_desc);
2489 	}
2490 	spin_unlock(&sde->flushlist_lock);
2491 	queue_work_on(sde->cpu, system_highpri_wq, &sde->flush_worker);
2492 	ret = -ECOMM;
2493 	goto update_tail;
2494 nodesc:
2495 	ret = sdma_check_progress(sde, wait, tx, submit_count > 0);
2496 	if (ret == -EAGAIN) {
2497 		ret = 0;
2498 		goto retry;
2499 	}
2500 	sde->descq_full_count++;
2501 	goto update_tail;
2502 }
2503 
2504 static void sdma_process_event(struct sdma_engine *sde, enum sdma_events event)
2505 {
2506 	unsigned long flags;
2507 
2508 	spin_lock_irqsave(&sde->tail_lock, flags);
2509 	write_seqlock(&sde->head_lock);
2510 
2511 	__sdma_process_event(sde, event);
2512 
2513 	if (sde->state.current_state == sdma_state_s99_running)
2514 		sdma_desc_avail(sde, sdma_descq_freecnt(sde));
2515 
2516 	write_sequnlock(&sde->head_lock);
2517 	spin_unlock_irqrestore(&sde->tail_lock, flags);
2518 }
2519 
2520 static void __sdma_process_event(struct sdma_engine *sde,
2521 				 enum sdma_events event)
2522 {
2523 	struct sdma_state *ss = &sde->state;
2524 	int need_progress = 0;
2525 
2526 	/* CONFIG SDMA temporary */
2527 #ifdef CONFIG_SDMA_VERBOSITY
2528 	dd_dev_err(sde->dd, "CONFIG SDMA(%u) [%s] %s\n", sde->this_idx,
2529 		   sdma_state_names[ss->current_state],
2530 		   sdma_event_names[event]);
2531 #endif
2532 
2533 	switch (ss->current_state) {
2534 	case sdma_state_s00_hw_down:
2535 		switch (event) {
2536 		case sdma_event_e00_go_hw_down:
2537 			break;
2538 		case sdma_event_e30_go_running:
2539 			/*
2540 			 * If down, but running requested (usually result
2541 			 * of link up, then we need to start up.
2542 			 * This can happen when hw down is requested while
2543 			 * bringing the link up with traffic active on
2544 			 * 7220, e.g.
2545 			 */
2546 			ss->go_s99_running = 1;
2547 			fallthrough;	/* and start dma engine */
2548 		case sdma_event_e10_go_hw_start:
2549 			/* This reference means the state machine is started */
2550 			sdma_get(&sde->state);
2551 			sdma_set_state(sde,
2552 				       sdma_state_s10_hw_start_up_halt_wait);
2553 			break;
2554 		case sdma_event_e15_hw_halt_done:
2555 			break;
2556 		case sdma_event_e25_hw_clean_up_done:
2557 			break;
2558 		case sdma_event_e40_sw_cleaned:
2559 			sdma_sw_tear_down(sde);
2560 			break;
2561 		case sdma_event_e50_hw_cleaned:
2562 			break;
2563 		case sdma_event_e60_hw_halted:
2564 			break;
2565 		case sdma_event_e70_go_idle:
2566 			break;
2567 		case sdma_event_e80_hw_freeze:
2568 			break;
2569 		case sdma_event_e81_hw_frozen:
2570 			break;
2571 		case sdma_event_e82_hw_unfreeze:
2572 			break;
2573 		case sdma_event_e85_link_down:
2574 			break;
2575 		case sdma_event_e90_sw_halted:
2576 			break;
2577 		}
2578 		break;
2579 
2580 	case sdma_state_s10_hw_start_up_halt_wait:
2581 		switch (event) {
2582 		case sdma_event_e00_go_hw_down:
2583 			sdma_set_state(sde, sdma_state_s00_hw_down);
2584 			sdma_sw_tear_down(sde);
2585 			break;
2586 		case sdma_event_e10_go_hw_start:
2587 			break;
2588 		case sdma_event_e15_hw_halt_done:
2589 			sdma_set_state(sde,
2590 				       sdma_state_s15_hw_start_up_clean_wait);
2591 			sdma_start_hw_clean_up(sde);
2592 			break;
2593 		case sdma_event_e25_hw_clean_up_done:
2594 			break;
2595 		case sdma_event_e30_go_running:
2596 			ss->go_s99_running = 1;
2597 			break;
2598 		case sdma_event_e40_sw_cleaned:
2599 			break;
2600 		case sdma_event_e50_hw_cleaned:
2601 			break;
2602 		case sdma_event_e60_hw_halted:
2603 			schedule_work(&sde->err_halt_worker);
2604 			break;
2605 		case sdma_event_e70_go_idle:
2606 			ss->go_s99_running = 0;
2607 			break;
2608 		case sdma_event_e80_hw_freeze:
2609 			break;
2610 		case sdma_event_e81_hw_frozen:
2611 			break;
2612 		case sdma_event_e82_hw_unfreeze:
2613 			break;
2614 		case sdma_event_e85_link_down:
2615 			break;
2616 		case sdma_event_e90_sw_halted:
2617 			break;
2618 		}
2619 		break;
2620 
2621 	case sdma_state_s15_hw_start_up_clean_wait:
2622 		switch (event) {
2623 		case sdma_event_e00_go_hw_down:
2624 			sdma_set_state(sde, sdma_state_s00_hw_down);
2625 			sdma_sw_tear_down(sde);
2626 			break;
2627 		case sdma_event_e10_go_hw_start:
2628 			break;
2629 		case sdma_event_e15_hw_halt_done:
2630 			break;
2631 		case sdma_event_e25_hw_clean_up_done:
2632 			sdma_hw_start_up(sde);
2633 			sdma_set_state(sde, ss->go_s99_running ?
2634 				       sdma_state_s99_running :
2635 				       sdma_state_s20_idle);
2636 			break;
2637 		case sdma_event_e30_go_running:
2638 			ss->go_s99_running = 1;
2639 			break;
2640 		case sdma_event_e40_sw_cleaned:
2641 			break;
2642 		case sdma_event_e50_hw_cleaned:
2643 			break;
2644 		case sdma_event_e60_hw_halted:
2645 			break;
2646 		case sdma_event_e70_go_idle:
2647 			ss->go_s99_running = 0;
2648 			break;
2649 		case sdma_event_e80_hw_freeze:
2650 			break;
2651 		case sdma_event_e81_hw_frozen:
2652 			break;
2653 		case sdma_event_e82_hw_unfreeze:
2654 			break;
2655 		case sdma_event_e85_link_down:
2656 			break;
2657 		case sdma_event_e90_sw_halted:
2658 			break;
2659 		}
2660 		break;
2661 
2662 	case sdma_state_s20_idle:
2663 		switch (event) {
2664 		case sdma_event_e00_go_hw_down:
2665 			sdma_set_state(sde, sdma_state_s00_hw_down);
2666 			sdma_sw_tear_down(sde);
2667 			break;
2668 		case sdma_event_e10_go_hw_start:
2669 			break;
2670 		case sdma_event_e15_hw_halt_done:
2671 			break;
2672 		case sdma_event_e25_hw_clean_up_done:
2673 			break;
2674 		case sdma_event_e30_go_running:
2675 			sdma_set_state(sde, sdma_state_s99_running);
2676 			ss->go_s99_running = 1;
2677 			break;
2678 		case sdma_event_e40_sw_cleaned:
2679 			break;
2680 		case sdma_event_e50_hw_cleaned:
2681 			break;
2682 		case sdma_event_e60_hw_halted:
2683 			sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
2684 			schedule_work(&sde->err_halt_worker);
2685 			break;
2686 		case sdma_event_e70_go_idle:
2687 			break;
2688 		case sdma_event_e85_link_down:
2689 		case sdma_event_e80_hw_freeze:
2690 			sdma_set_state(sde, sdma_state_s80_hw_freeze);
2691 			atomic_dec(&sde->dd->sdma_unfreeze_count);
2692 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2693 			break;
2694 		case sdma_event_e81_hw_frozen:
2695 			break;
2696 		case sdma_event_e82_hw_unfreeze:
2697 			break;
2698 		case sdma_event_e90_sw_halted:
2699 			break;
2700 		}
2701 		break;
2702 
2703 	case sdma_state_s30_sw_clean_up_wait:
2704 		switch (event) {
2705 		case sdma_event_e00_go_hw_down:
2706 			sdma_set_state(sde, sdma_state_s00_hw_down);
2707 			break;
2708 		case sdma_event_e10_go_hw_start:
2709 			break;
2710 		case sdma_event_e15_hw_halt_done:
2711 			break;
2712 		case sdma_event_e25_hw_clean_up_done:
2713 			break;
2714 		case sdma_event_e30_go_running:
2715 			ss->go_s99_running = 1;
2716 			break;
2717 		case sdma_event_e40_sw_cleaned:
2718 			sdma_set_state(sde, sdma_state_s40_hw_clean_up_wait);
2719 			sdma_start_hw_clean_up(sde);
2720 			break;
2721 		case sdma_event_e50_hw_cleaned:
2722 			break;
2723 		case sdma_event_e60_hw_halted:
2724 			break;
2725 		case sdma_event_e70_go_idle:
2726 			ss->go_s99_running = 0;
2727 			break;
2728 		case sdma_event_e80_hw_freeze:
2729 			break;
2730 		case sdma_event_e81_hw_frozen:
2731 			break;
2732 		case sdma_event_e82_hw_unfreeze:
2733 			break;
2734 		case sdma_event_e85_link_down:
2735 			ss->go_s99_running = 0;
2736 			break;
2737 		case sdma_event_e90_sw_halted:
2738 			break;
2739 		}
2740 		break;
2741 
2742 	case sdma_state_s40_hw_clean_up_wait:
2743 		switch (event) {
2744 		case sdma_event_e00_go_hw_down:
2745 			sdma_set_state(sde, sdma_state_s00_hw_down);
2746 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2747 			break;
2748 		case sdma_event_e10_go_hw_start:
2749 			break;
2750 		case sdma_event_e15_hw_halt_done:
2751 			break;
2752 		case sdma_event_e25_hw_clean_up_done:
2753 			sdma_hw_start_up(sde);
2754 			sdma_set_state(sde, ss->go_s99_running ?
2755 				       sdma_state_s99_running :
2756 				       sdma_state_s20_idle);
2757 			break;
2758 		case sdma_event_e30_go_running:
2759 			ss->go_s99_running = 1;
2760 			break;
2761 		case sdma_event_e40_sw_cleaned:
2762 			break;
2763 		case sdma_event_e50_hw_cleaned:
2764 			break;
2765 		case sdma_event_e60_hw_halted:
2766 			break;
2767 		case sdma_event_e70_go_idle:
2768 			ss->go_s99_running = 0;
2769 			break;
2770 		case sdma_event_e80_hw_freeze:
2771 			break;
2772 		case sdma_event_e81_hw_frozen:
2773 			break;
2774 		case sdma_event_e82_hw_unfreeze:
2775 			break;
2776 		case sdma_event_e85_link_down:
2777 			ss->go_s99_running = 0;
2778 			break;
2779 		case sdma_event_e90_sw_halted:
2780 			break;
2781 		}
2782 		break;
2783 
2784 	case sdma_state_s50_hw_halt_wait:
2785 		switch (event) {
2786 		case sdma_event_e00_go_hw_down:
2787 			sdma_set_state(sde, sdma_state_s00_hw_down);
2788 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2789 			break;
2790 		case sdma_event_e10_go_hw_start:
2791 			break;
2792 		case sdma_event_e15_hw_halt_done:
2793 			sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2794 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2795 			break;
2796 		case sdma_event_e25_hw_clean_up_done:
2797 			break;
2798 		case sdma_event_e30_go_running:
2799 			ss->go_s99_running = 1;
2800 			break;
2801 		case sdma_event_e40_sw_cleaned:
2802 			break;
2803 		case sdma_event_e50_hw_cleaned:
2804 			break;
2805 		case sdma_event_e60_hw_halted:
2806 			schedule_work(&sde->err_halt_worker);
2807 			break;
2808 		case sdma_event_e70_go_idle:
2809 			ss->go_s99_running = 0;
2810 			break;
2811 		case sdma_event_e80_hw_freeze:
2812 			break;
2813 		case sdma_event_e81_hw_frozen:
2814 			break;
2815 		case sdma_event_e82_hw_unfreeze:
2816 			break;
2817 		case sdma_event_e85_link_down:
2818 			ss->go_s99_running = 0;
2819 			break;
2820 		case sdma_event_e90_sw_halted:
2821 			break;
2822 		}
2823 		break;
2824 
2825 	case sdma_state_s60_idle_halt_wait:
2826 		switch (event) {
2827 		case sdma_event_e00_go_hw_down:
2828 			sdma_set_state(sde, sdma_state_s00_hw_down);
2829 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2830 			break;
2831 		case sdma_event_e10_go_hw_start:
2832 			break;
2833 		case sdma_event_e15_hw_halt_done:
2834 			sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2835 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2836 			break;
2837 		case sdma_event_e25_hw_clean_up_done:
2838 			break;
2839 		case sdma_event_e30_go_running:
2840 			ss->go_s99_running = 1;
2841 			break;
2842 		case sdma_event_e40_sw_cleaned:
2843 			break;
2844 		case sdma_event_e50_hw_cleaned:
2845 			break;
2846 		case sdma_event_e60_hw_halted:
2847 			schedule_work(&sde->err_halt_worker);
2848 			break;
2849 		case sdma_event_e70_go_idle:
2850 			ss->go_s99_running = 0;
2851 			break;
2852 		case sdma_event_e80_hw_freeze:
2853 			break;
2854 		case sdma_event_e81_hw_frozen:
2855 			break;
2856 		case sdma_event_e82_hw_unfreeze:
2857 			break;
2858 		case sdma_event_e85_link_down:
2859 			break;
2860 		case sdma_event_e90_sw_halted:
2861 			break;
2862 		}
2863 		break;
2864 
2865 	case sdma_state_s80_hw_freeze:
2866 		switch (event) {
2867 		case sdma_event_e00_go_hw_down:
2868 			sdma_set_state(sde, sdma_state_s00_hw_down);
2869 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2870 			break;
2871 		case sdma_event_e10_go_hw_start:
2872 			break;
2873 		case sdma_event_e15_hw_halt_done:
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 			break;
2886 		case sdma_event_e70_go_idle:
2887 			ss->go_s99_running = 0;
2888 			break;
2889 		case sdma_event_e80_hw_freeze:
2890 			break;
2891 		case sdma_event_e81_hw_frozen:
2892 			sdma_set_state(sde, sdma_state_s82_freeze_sw_clean);
2893 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2894 			break;
2895 		case sdma_event_e82_hw_unfreeze:
2896 			break;
2897 		case sdma_event_e85_link_down:
2898 			break;
2899 		case sdma_event_e90_sw_halted:
2900 			break;
2901 		}
2902 		break;
2903 
2904 	case sdma_state_s82_freeze_sw_clean:
2905 		switch (event) {
2906 		case sdma_event_e00_go_hw_down:
2907 			sdma_set_state(sde, sdma_state_s00_hw_down);
2908 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2909 			break;
2910 		case sdma_event_e10_go_hw_start:
2911 			break;
2912 		case sdma_event_e15_hw_halt_done:
2913 			break;
2914 		case sdma_event_e25_hw_clean_up_done:
2915 			break;
2916 		case sdma_event_e30_go_running:
2917 			ss->go_s99_running = 1;
2918 			break;
2919 		case sdma_event_e40_sw_cleaned:
2920 			/* notify caller this engine is done cleaning */
2921 			atomic_dec(&sde->dd->sdma_unfreeze_count);
2922 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2923 			break;
2924 		case sdma_event_e50_hw_cleaned:
2925 			break;
2926 		case sdma_event_e60_hw_halted:
2927 			break;
2928 		case sdma_event_e70_go_idle:
2929 			ss->go_s99_running = 0;
2930 			break;
2931 		case sdma_event_e80_hw_freeze:
2932 			break;
2933 		case sdma_event_e81_hw_frozen:
2934 			break;
2935 		case sdma_event_e82_hw_unfreeze:
2936 			sdma_hw_start_up(sde);
2937 			sdma_set_state(sde, ss->go_s99_running ?
2938 				       sdma_state_s99_running :
2939 				       sdma_state_s20_idle);
2940 			break;
2941 		case sdma_event_e85_link_down:
2942 			break;
2943 		case sdma_event_e90_sw_halted:
2944 			break;
2945 		}
2946 		break;
2947 
2948 	case sdma_state_s99_running:
2949 		switch (event) {
2950 		case sdma_event_e00_go_hw_down:
2951 			sdma_set_state(sde, sdma_state_s00_hw_down);
2952 			tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2953 			break;
2954 		case sdma_event_e10_go_hw_start:
2955 			break;
2956 		case sdma_event_e15_hw_halt_done:
2957 			break;
2958 		case sdma_event_e25_hw_clean_up_done:
2959 			break;
2960 		case sdma_event_e30_go_running:
2961 			break;
2962 		case sdma_event_e40_sw_cleaned:
2963 			break;
2964 		case sdma_event_e50_hw_cleaned:
2965 			break;
2966 		case sdma_event_e60_hw_halted:
2967 			need_progress = 1;
2968 			sdma_err_progress_check_schedule(sde);
2969 			fallthrough;
2970 		case sdma_event_e90_sw_halted:
2971 			/*
2972 			* SW initiated halt does not perform engines
2973 			* progress check
2974 			*/
2975 			sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
2976 			schedule_work(&sde->err_halt_worker);
2977 			break;
2978 		case sdma_event_e70_go_idle:
2979 			sdma_set_state(sde, sdma_state_s60_idle_halt_wait);
2980 			break;
2981 		case sdma_event_e85_link_down:
2982 			ss->go_s99_running = 0;
2983 			fallthrough;
2984 		case sdma_event_e80_hw_freeze:
2985 			sdma_set_state(sde, sdma_state_s80_hw_freeze);
2986 			atomic_dec(&sde->dd->sdma_unfreeze_count);
2987 			wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2988 			break;
2989 		case sdma_event_e81_hw_frozen:
2990 			break;
2991 		case sdma_event_e82_hw_unfreeze:
2992 			break;
2993 		}
2994 		break;
2995 	}
2996 
2997 	ss->last_event = event;
2998 	if (need_progress)
2999 		sdma_make_progress(sde, 0);
3000 }
3001 
3002 /*
3003  * _extend_sdma_tx_descs() - helper to extend txreq
3004  *
3005  * This is called once the initial nominal allocation
3006  * of descriptors in the sdma_txreq is exhausted.
3007  *
3008  * The code will bump the allocation up to the max
3009  * of MAX_DESC (64) descriptors. There doesn't seem
3010  * much point in an interim step. The last descriptor
3011  * is reserved for coalesce buffer in order to support
3012  * cases where input packet has >MAX_DESC iovecs.
3013  *
3014  */
3015 static int _extend_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3016 {
3017 	int i;
3018 	struct sdma_desc *descp;
3019 
3020 	/* Handle last descriptor */
3021 	if (unlikely((tx->num_desc == (MAX_DESC - 1)))) {
3022 		/* if tlen is 0, it is for padding, release last descriptor */
3023 		if (!tx->tlen) {
3024 			tx->desc_limit = MAX_DESC;
3025 		} else if (!tx->coalesce_buf) {
3026 			/* allocate coalesce buffer with space for padding */
3027 			tx->coalesce_buf = kmalloc(tx->tlen + sizeof(u32),
3028 						   GFP_ATOMIC);
3029 			if (!tx->coalesce_buf)
3030 				goto enomem;
3031 			tx->coalesce_idx = 0;
3032 		}
3033 		return 0;
3034 	}
3035 
3036 	if (unlikely(tx->num_desc == MAX_DESC))
3037 		goto enomem;
3038 
3039 	descp = kmalloc_array(MAX_DESC, sizeof(struct sdma_desc), GFP_ATOMIC);
3040 	if (!descp)
3041 		goto enomem;
3042 	tx->descp = descp;
3043 
3044 	/* reserve last descriptor for coalescing */
3045 	tx->desc_limit = MAX_DESC - 1;
3046 	/* copy ones already built */
3047 	for (i = 0; i < tx->num_desc; i++)
3048 		tx->descp[i] = tx->descs[i];
3049 	return 0;
3050 enomem:
3051 	__sdma_txclean(dd, tx);
3052 	return -ENOMEM;
3053 }
3054 
3055 /*
3056  * ext_coal_sdma_tx_descs() - extend or coalesce sdma tx descriptors
3057  *
3058  * This is called once the initial nominal allocation of descriptors
3059  * in the sdma_txreq is exhausted.
3060  *
3061  * This function calls _extend_sdma_tx_descs to extend or allocate
3062  * coalesce buffer. If there is a allocated coalesce buffer, it will
3063  * copy the input packet data into the coalesce buffer. It also adds
3064  * coalesce buffer descriptor once when whole packet is received.
3065  *
3066  * Return:
3067  * <0 - error
3068  * 0 - coalescing, don't populate descriptor
3069  * 1 - continue with populating descriptor
3070  */
3071 int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx,
3072 			   int type, void *kvaddr, struct page *page,
3073 			   unsigned long offset, u16 len)
3074 {
3075 	int pad_len, rval;
3076 	dma_addr_t addr;
3077 
3078 	rval = _extend_sdma_tx_descs(dd, tx);
3079 	if (rval) {
3080 		__sdma_txclean(dd, tx);
3081 		return rval;
3082 	}
3083 
3084 	/* If coalesce buffer is allocated, copy data into it */
3085 	if (tx->coalesce_buf) {
3086 		if (type == SDMA_MAP_NONE) {
3087 			__sdma_txclean(dd, tx);
3088 			return -EINVAL;
3089 		}
3090 
3091 		if (type == SDMA_MAP_PAGE) {
3092 			kvaddr = kmap_local_page(page);
3093 			kvaddr += offset;
3094 		} else if (WARN_ON(!kvaddr)) {
3095 			__sdma_txclean(dd, tx);
3096 			return -EINVAL;
3097 		}
3098 
3099 		memcpy(tx->coalesce_buf + tx->coalesce_idx, kvaddr, len);
3100 		tx->coalesce_idx += len;
3101 		if (type == SDMA_MAP_PAGE)
3102 			kunmap_local(kvaddr);
3103 
3104 		/* If there is more data, return */
3105 		if (tx->tlen - tx->coalesce_idx)
3106 			return 0;
3107 
3108 		/* Whole packet is received; add any padding */
3109 		pad_len = tx->packet_len & (sizeof(u32) - 1);
3110 		if (pad_len) {
3111 			pad_len = sizeof(u32) - pad_len;
3112 			memset(tx->coalesce_buf + tx->coalesce_idx, 0, pad_len);
3113 			/* padding is taken care of for coalescing case */
3114 			tx->packet_len += pad_len;
3115 			tx->tlen += pad_len;
3116 		}
3117 
3118 		/* dma map the coalesce buffer */
3119 		addr = dma_map_single(&dd->pcidev->dev,
3120 				      tx->coalesce_buf,
3121 				      tx->tlen,
3122 				      DMA_TO_DEVICE);
3123 
3124 		if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) {
3125 			__sdma_txclean(dd, tx);
3126 			return -ENOSPC;
3127 		}
3128 
3129 		/* Add descriptor for coalesce buffer */
3130 		tx->desc_limit = MAX_DESC;
3131 		return _sdma_txadd_daddr(dd, SDMA_MAP_SINGLE, tx,
3132 					 addr, tx->tlen);
3133 	}
3134 
3135 	return 1;
3136 }
3137 
3138 /* Update sdes when the lmc changes */
3139 void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid)
3140 {
3141 	struct sdma_engine *sde;
3142 	int i;
3143 	u64 sreg;
3144 
3145 	sreg = ((mask & SD(CHECK_SLID_MASK_MASK)) <<
3146 		SD(CHECK_SLID_MASK_SHIFT)) |
3147 		(((lid & mask) & SD(CHECK_SLID_VALUE_MASK)) <<
3148 		SD(CHECK_SLID_VALUE_SHIFT));
3149 
3150 	for (i = 0; i < dd->num_sdma; i++) {
3151 		hfi1_cdbg(LINKVERB, "SendDmaEngine[%d].SLID_CHECK = 0x%x",
3152 			  i, (u32)sreg);
3153 		sde = &dd->per_sdma[i];
3154 		write_sde_csr(sde, SD(CHECK_SLID), sreg);
3155 	}
3156 }
3157 
3158 /* tx not dword sized - pad */
3159 int _pad_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3160 {
3161 	int rval = 0;
3162 
3163 	tx->num_desc++;
3164 	if ((unlikely(tx->num_desc == tx->desc_limit))) {
3165 		rval = _extend_sdma_tx_descs(dd, tx);
3166 		if (rval) {
3167 			__sdma_txclean(dd, tx);
3168 			return rval;
3169 		}
3170 	}
3171 	/* finish the one just added */
3172 	make_tx_sdma_desc(
3173 		tx,
3174 		SDMA_MAP_NONE,
3175 		dd->sdma_pad_phys,
3176 		sizeof(u32) - (tx->packet_len & (sizeof(u32) - 1)));
3177 	_sdma_close_tx(dd, tx);
3178 	return rval;
3179 }
3180 
3181 /*
3182  * Add ahg to the sdma_txreq
3183  *
3184  * The logic will consume up to 3
3185  * descriptors at the beginning of
3186  * sdma_txreq.
3187  */
3188 void _sdma_txreq_ahgadd(
3189 	struct sdma_txreq *tx,
3190 	u8 num_ahg,
3191 	u8 ahg_entry,
3192 	u32 *ahg,
3193 	u8 ahg_hlen)
3194 {
3195 	u32 i, shift = 0, desc = 0;
3196 	u8 mode;
3197 
3198 	WARN_ON_ONCE(num_ahg > 9 || (ahg_hlen & 3) || ahg_hlen == 4);
3199 	/* compute mode */
3200 	if (num_ahg == 1)
3201 		mode = SDMA_AHG_APPLY_UPDATE1;
3202 	else if (num_ahg <= 5)
3203 		mode = SDMA_AHG_APPLY_UPDATE2;
3204 	else
3205 		mode = SDMA_AHG_APPLY_UPDATE3;
3206 	tx->num_desc++;
3207 	/* initialize to consumed descriptors to zero */
3208 	switch (mode) {
3209 	case SDMA_AHG_APPLY_UPDATE3:
3210 		tx->num_desc++;
3211 		tx->descs[2].qw[0] = 0;
3212 		tx->descs[2].qw[1] = 0;
3213 		fallthrough;
3214 	case SDMA_AHG_APPLY_UPDATE2:
3215 		tx->num_desc++;
3216 		tx->descs[1].qw[0] = 0;
3217 		tx->descs[1].qw[1] = 0;
3218 		break;
3219 	}
3220 	ahg_hlen >>= 2;
3221 	tx->descs[0].qw[1] |=
3222 		(((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK)
3223 			<< SDMA_DESC1_HEADER_INDEX_SHIFT) |
3224 		(((u64)ahg_hlen & SDMA_DESC1_HEADER_DWS_MASK)
3225 			<< SDMA_DESC1_HEADER_DWS_SHIFT) |
3226 		(((u64)mode & SDMA_DESC1_HEADER_MODE_MASK)
3227 			<< SDMA_DESC1_HEADER_MODE_SHIFT) |
3228 		(((u64)ahg[0] & SDMA_DESC1_HEADER_UPDATE1_MASK)
3229 			<< SDMA_DESC1_HEADER_UPDATE1_SHIFT);
3230 	for (i = 0; i < (num_ahg - 1); i++) {
3231 		if (!shift && !(i & 2))
3232 			desc++;
3233 		tx->descs[desc].qw[!!(i & 2)] |=
3234 			(((u64)ahg[i + 1])
3235 				<< shift);
3236 		shift = (shift + 32) & 63;
3237 	}
3238 }
3239 
3240 /**
3241  * sdma_ahg_alloc - allocate an AHG entry
3242  * @sde: engine to allocate from
3243  *
3244  * Return:
3245  * 0-31 when successful, -EOPNOTSUPP if AHG is not enabled,
3246  * -ENOSPC if an entry is not available
3247  */
3248 int sdma_ahg_alloc(struct sdma_engine *sde)
3249 {
3250 	int nr;
3251 	int oldbit;
3252 
3253 	if (!sde) {
3254 		trace_hfi1_ahg_allocate(sde, -EINVAL);
3255 		return -EINVAL;
3256 	}
3257 	while (1) {
3258 		nr = ffz(READ_ONCE(sde->ahg_bits));
3259 		if (nr > 31) {
3260 			trace_hfi1_ahg_allocate(sde, -ENOSPC);
3261 			return -ENOSPC;
3262 		}
3263 		oldbit = test_and_set_bit(nr, &sde->ahg_bits);
3264 		if (!oldbit)
3265 			break;
3266 		cpu_relax();
3267 	}
3268 	trace_hfi1_ahg_allocate(sde, nr);
3269 	return nr;
3270 }
3271 
3272 /**
3273  * sdma_ahg_free - free an AHG entry
3274  * @sde: engine to return AHG entry
3275  * @ahg_index: index to free
3276  *
3277  * This routine frees the indicate AHG entry.
3278  */
3279 void sdma_ahg_free(struct sdma_engine *sde, int ahg_index)
3280 {
3281 	if (!sde)
3282 		return;
3283 	trace_hfi1_ahg_deallocate(sde, ahg_index);
3284 	if (ahg_index < 0 || ahg_index > 31)
3285 		return;
3286 	clear_bit(ahg_index, &sde->ahg_bits);
3287 }
3288 
3289 /*
3290  * SPC freeze handling for SDMA engines.  Called when the driver knows
3291  * the SPC is going into a freeze but before the freeze is fully
3292  * settled.  Generally an error interrupt.
3293  *
3294  * This event will pull the engine out of running so no more entries can be
3295  * added to the engine's queue.
3296  */
3297 void sdma_freeze_notify(struct hfi1_devdata *dd, int link_down)
3298 {
3299 	int i;
3300 	enum sdma_events event = link_down ? sdma_event_e85_link_down :
3301 					     sdma_event_e80_hw_freeze;
3302 
3303 	/* set up the wait but do not wait here */
3304 	atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3305 
3306 	/* tell all engines to stop running and wait */
3307 	for (i = 0; i < dd->num_sdma; i++)
3308 		sdma_process_event(&dd->per_sdma[i], event);
3309 
3310 	/* sdma_freeze() will wait for all engines to have stopped */
3311 }
3312 
3313 /*
3314  * SPC freeze handling for SDMA engines.  Called when the driver knows
3315  * the SPC is fully frozen.
3316  */
3317 void sdma_freeze(struct hfi1_devdata *dd)
3318 {
3319 	int i;
3320 	int ret;
3321 
3322 	/*
3323 	 * Make sure all engines have moved out of the running state before
3324 	 * continuing.
3325 	 */
3326 	ret = wait_event_interruptible(dd->sdma_unfreeze_wq,
3327 				       atomic_read(&dd->sdma_unfreeze_count) <=
3328 				       0);
3329 	/* interrupted or count is negative, then unloading - just exit */
3330 	if (ret || atomic_read(&dd->sdma_unfreeze_count) < 0)
3331 		return;
3332 
3333 	/* set up the count for the next wait */
3334 	atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3335 
3336 	/* tell all engines that the SPC is frozen, they can start cleaning */
3337 	for (i = 0; i < dd->num_sdma; i++)
3338 		sdma_process_event(&dd->per_sdma[i], sdma_event_e81_hw_frozen);
3339 
3340 	/*
3341 	 * Wait for everyone to finish software clean before exiting.  The
3342 	 * software clean will read engine CSRs, so must be completed before
3343 	 * the next step, which will clear the engine CSRs.
3344 	 */
3345 	(void)wait_event_interruptible(dd->sdma_unfreeze_wq,
3346 				atomic_read(&dd->sdma_unfreeze_count) <= 0);
3347 	/* no need to check results - done no matter what */
3348 }
3349 
3350 /*
3351  * SPC freeze handling for the SDMA engines.  Called after the SPC is unfrozen.
3352  *
3353  * The SPC freeze acts like a SDMA halt and a hardware clean combined.  All
3354  * that is left is a software clean.  We could do it after the SPC is fully
3355  * frozen, but then we'd have to add another state to wait for the unfreeze.
3356  * Instead, just defer the software clean until the unfreeze step.
3357  */
3358 void sdma_unfreeze(struct hfi1_devdata *dd)
3359 {
3360 	int i;
3361 
3362 	/* tell all engines start freeze clean up */
3363 	for (i = 0; i < dd->num_sdma; i++)
3364 		sdma_process_event(&dd->per_sdma[i],
3365 				   sdma_event_e82_hw_unfreeze);
3366 }
3367 
3368 /**
3369  * _sdma_engine_progress_schedule() - schedule progress on engine
3370  * @sde: sdma_engine to schedule progress
3371  *
3372  */
3373 void _sdma_engine_progress_schedule(
3374 	struct sdma_engine *sde)
3375 {
3376 	trace_hfi1_sdma_engine_progress(sde, sde->progress_mask);
3377 	/* assume we have selected a good cpu */
3378 	write_csr(sde->dd,
3379 		  CCE_INT_FORCE + (8 * (IS_SDMA_START / 64)),
3380 		  sde->progress_mask);
3381 }
3382