xref: /linux/drivers/infiniband/hw/hfi1/pio.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
1 /*
2  * Copyright(c) 2015-2017 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/delay.h>
49 #include "hfi.h"
50 #include "qp.h"
51 #include "trace.h"
52 
53 #define SC_CTXT_PACKET_EGRESS_TIMEOUT 350 /* in chip cycles */
54 
55 #define SC(name) SEND_CTXT_##name
56 /*
57  * Send Context functions
58  */
59 static void sc_wait_for_packet_egress(struct send_context *sc, int pause);
60 
61 /*
62  * Set the CM reset bit and wait for it to clear.  Use the provided
63  * sendctrl register.  This routine has no locking.
64  */
65 void __cm_reset(struct hfi1_devdata *dd, u64 sendctrl)
66 {
67 	write_csr(dd, SEND_CTRL, sendctrl | SEND_CTRL_CM_RESET_SMASK);
68 	while (1) {
69 		udelay(1);
70 		sendctrl = read_csr(dd, SEND_CTRL);
71 		if ((sendctrl & SEND_CTRL_CM_RESET_SMASK) == 0)
72 			break;
73 	}
74 }
75 
76 /* defined in header release 48 and higher */
77 #ifndef SEND_CTRL_UNSUPPORTED_VL_SHIFT
78 #define SEND_CTRL_UNSUPPORTED_VL_SHIFT 3
79 #define SEND_CTRL_UNSUPPORTED_VL_MASK 0xffull
80 #define SEND_CTRL_UNSUPPORTED_VL_SMASK (SEND_CTRL_UNSUPPORTED_VL_MASK \
81 		<< SEND_CTRL_UNSUPPORTED_VL_SHIFT)
82 #endif
83 
84 /* global control of PIO send */
85 void pio_send_control(struct hfi1_devdata *dd, int op)
86 {
87 	u64 reg, mask;
88 	unsigned long flags;
89 	int write = 1;	/* write sendctrl back */
90 	int flush = 0;	/* re-read sendctrl to make sure it is flushed */
91 
92 	spin_lock_irqsave(&dd->sendctrl_lock, flags);
93 
94 	reg = read_csr(dd, SEND_CTRL);
95 	switch (op) {
96 	case PSC_GLOBAL_ENABLE:
97 		reg |= SEND_CTRL_SEND_ENABLE_SMASK;
98 	/* Fall through */
99 	case PSC_DATA_VL_ENABLE:
100 		/* Disallow sending on VLs not enabled */
101 		mask = (((~0ull) << num_vls) & SEND_CTRL_UNSUPPORTED_VL_MASK) <<
102 				SEND_CTRL_UNSUPPORTED_VL_SHIFT;
103 		reg = (reg & ~SEND_CTRL_UNSUPPORTED_VL_SMASK) | mask;
104 		break;
105 	case PSC_GLOBAL_DISABLE:
106 		reg &= ~SEND_CTRL_SEND_ENABLE_SMASK;
107 		break;
108 	case PSC_GLOBAL_VLARB_ENABLE:
109 		reg |= SEND_CTRL_VL_ARBITER_ENABLE_SMASK;
110 		break;
111 	case PSC_GLOBAL_VLARB_DISABLE:
112 		reg &= ~SEND_CTRL_VL_ARBITER_ENABLE_SMASK;
113 		break;
114 	case PSC_CM_RESET:
115 		__cm_reset(dd, reg);
116 		write = 0; /* CSR already written (and flushed) */
117 		break;
118 	case PSC_DATA_VL_DISABLE:
119 		reg |= SEND_CTRL_UNSUPPORTED_VL_SMASK;
120 		flush = 1;
121 		break;
122 	default:
123 		dd_dev_err(dd, "%s: invalid control %d\n", __func__, op);
124 		break;
125 	}
126 
127 	if (write) {
128 		write_csr(dd, SEND_CTRL, reg);
129 		if (flush)
130 			(void)read_csr(dd, SEND_CTRL); /* flush write */
131 	}
132 
133 	spin_unlock_irqrestore(&dd->sendctrl_lock, flags);
134 }
135 
136 /* number of send context memory pools */
137 #define NUM_SC_POOLS 2
138 
139 /* Send Context Size (SCS) wildcards */
140 #define SCS_POOL_0 -1
141 #define SCS_POOL_1 -2
142 
143 /* Send Context Count (SCC) wildcards */
144 #define SCC_PER_VL -1
145 #define SCC_PER_CPU  -2
146 #define SCC_PER_KRCVQ  -3
147 
148 /* Send Context Size (SCS) constants */
149 #define SCS_ACK_CREDITS  32
150 #define SCS_VL15_CREDITS 102	/* 3 pkts of 2048B data + 128B header */
151 
152 #define PIO_THRESHOLD_CEILING 4096
153 
154 #define PIO_WAIT_BATCH_SIZE 5
155 
156 /* default send context sizes */
157 static struct sc_config_sizes sc_config_sizes[SC_MAX] = {
158 	[SC_KERNEL] = { .size  = SCS_POOL_0,	/* even divide, pool 0 */
159 			.count = SCC_PER_VL },	/* one per NUMA */
160 	[SC_ACK]    = { .size  = SCS_ACK_CREDITS,
161 			.count = SCC_PER_KRCVQ },
162 	[SC_USER]   = { .size  = SCS_POOL_0,	/* even divide, pool 0 */
163 			.count = SCC_PER_CPU },	/* one per CPU */
164 	[SC_VL15]   = { .size  = SCS_VL15_CREDITS,
165 			.count = 1 },
166 
167 };
168 
169 /* send context memory pool configuration */
170 struct mem_pool_config {
171 	int centipercent;	/* % of memory, in 100ths of 1% */
172 	int absolute_blocks;	/* absolute block count */
173 };
174 
175 /* default memory pool configuration: 100% in pool 0 */
176 static struct mem_pool_config sc_mem_pool_config[NUM_SC_POOLS] = {
177 	/* centi%, abs blocks */
178 	{  10000,     -1 },		/* pool 0 */
179 	{      0,     -1 },		/* pool 1 */
180 };
181 
182 /* memory pool information, used when calculating final sizes */
183 struct mem_pool_info {
184 	int centipercent;	/*
185 				 * 100th of 1% of memory to use, -1 if blocks
186 				 * already set
187 				 */
188 	int count;		/* count of contexts in the pool */
189 	int blocks;		/* block size of the pool */
190 	int size;		/* context size, in blocks */
191 };
192 
193 /*
194  * Convert a pool wildcard to a valid pool index.  The wildcards
195  * start at -1 and increase negatively.  Map them as:
196  *	-1 => 0
197  *	-2 => 1
198  *	etc.
199  *
200  * Return -1 on non-wildcard input, otherwise convert to a pool number.
201  */
202 static int wildcard_to_pool(int wc)
203 {
204 	if (wc >= 0)
205 		return -1;	/* non-wildcard */
206 	return -wc - 1;
207 }
208 
209 static const char *sc_type_names[SC_MAX] = {
210 	"kernel",
211 	"ack",
212 	"user",
213 	"vl15"
214 };
215 
216 static const char *sc_type_name(int index)
217 {
218 	if (index < 0 || index >= SC_MAX)
219 		return "unknown";
220 	return sc_type_names[index];
221 }
222 
223 /*
224  * Read the send context memory pool configuration and send context
225  * size configuration.  Replace any wildcards and come up with final
226  * counts and sizes for the send context types.
227  */
228 int init_sc_pools_and_sizes(struct hfi1_devdata *dd)
229 {
230 	struct mem_pool_info mem_pool_info[NUM_SC_POOLS] = { { 0 } };
231 	int total_blocks = (dd->chip_pio_mem_size / PIO_BLOCK_SIZE) - 1;
232 	int total_contexts = 0;
233 	int fixed_blocks;
234 	int pool_blocks;
235 	int used_blocks;
236 	int cp_total;		/* centipercent total */
237 	int ab_total;		/* absolute block total */
238 	int extra;
239 	int i;
240 
241 	/*
242 	 * When SDMA is enabled, kernel context pio packet size is capped by
243 	 * "piothreshold". Reduce pio buffer allocation for kernel context by
244 	 * setting it to a fixed size. The allocation allows 3-deep buffering
245 	 * of the largest pio packets plus up to 128 bytes header, sufficient
246 	 * to maintain verbs performance.
247 	 *
248 	 * When SDMA is disabled, keep the default pooling allocation.
249 	 */
250 	if (HFI1_CAP_IS_KSET(SDMA)) {
251 		u16 max_pkt_size = (piothreshold < PIO_THRESHOLD_CEILING) ?
252 					 piothreshold : PIO_THRESHOLD_CEILING;
253 		sc_config_sizes[SC_KERNEL].size =
254 			3 * (max_pkt_size + 128) / PIO_BLOCK_SIZE;
255 	}
256 
257 	/*
258 	 * Step 0:
259 	 *	- copy the centipercents/absolute sizes from the pool config
260 	 *	- sanity check these values
261 	 *	- add up centipercents, then later check for full value
262 	 *	- add up absolute blocks, then later check for over-commit
263 	 */
264 	cp_total = 0;
265 	ab_total = 0;
266 	for (i = 0; i < NUM_SC_POOLS; i++) {
267 		int cp = sc_mem_pool_config[i].centipercent;
268 		int ab = sc_mem_pool_config[i].absolute_blocks;
269 
270 		/*
271 		 * A negative value is "unused" or "invalid".  Both *can*
272 		 * be valid, but centipercent wins, so check that first
273 		 */
274 		if (cp >= 0) {			/* centipercent valid */
275 			cp_total += cp;
276 		} else if (ab >= 0) {		/* absolute blocks valid */
277 			ab_total += ab;
278 		} else {			/* neither valid */
279 			dd_dev_err(
280 				dd,
281 				"Send context memory pool %d: both the block count and centipercent are invalid\n",
282 				i);
283 			return -EINVAL;
284 		}
285 
286 		mem_pool_info[i].centipercent = cp;
287 		mem_pool_info[i].blocks = ab;
288 	}
289 
290 	/* do not use both % and absolute blocks for different pools */
291 	if (cp_total != 0 && ab_total != 0) {
292 		dd_dev_err(
293 			dd,
294 			"All send context memory pools must be described as either centipercent or blocks, no mixing between pools\n");
295 		return -EINVAL;
296 	}
297 
298 	/* if any percentages are present, they must add up to 100% x 100 */
299 	if (cp_total != 0 && cp_total != 10000) {
300 		dd_dev_err(
301 			dd,
302 			"Send context memory pool centipercent is %d, expecting 10000\n",
303 			cp_total);
304 		return -EINVAL;
305 	}
306 
307 	/* the absolute pool total cannot be more than the mem total */
308 	if (ab_total > total_blocks) {
309 		dd_dev_err(
310 			dd,
311 			"Send context memory pool absolute block count %d is larger than the memory size %d\n",
312 			ab_total, total_blocks);
313 		return -EINVAL;
314 	}
315 
316 	/*
317 	 * Step 2:
318 	 *	- copy from the context size config
319 	 *	- replace context type wildcard counts with real values
320 	 *	- add up non-memory pool block sizes
321 	 *	- add up memory pool user counts
322 	 */
323 	fixed_blocks = 0;
324 	for (i = 0; i < SC_MAX; i++) {
325 		int count = sc_config_sizes[i].count;
326 		int size = sc_config_sizes[i].size;
327 		int pool;
328 
329 		/*
330 		 * Sanity check count: Either a positive value or
331 		 * one of the expected wildcards is valid.  The positive
332 		 * value is checked later when we compare against total
333 		 * memory available.
334 		 */
335 		if (i == SC_ACK) {
336 			count = dd->n_krcv_queues;
337 		} else if (i == SC_KERNEL) {
338 			count = INIT_SC_PER_VL * num_vls;
339 		} else if (count == SCC_PER_CPU) {
340 			count = dd->num_rcv_contexts - dd->n_krcv_queues;
341 		} else if (count < 0) {
342 			dd_dev_err(
343 				dd,
344 				"%s send context invalid count wildcard %d\n",
345 				sc_type_name(i), count);
346 			return -EINVAL;
347 		}
348 		if (total_contexts + count > dd->chip_send_contexts)
349 			count = dd->chip_send_contexts - total_contexts;
350 
351 		total_contexts += count;
352 
353 		/*
354 		 * Sanity check pool: The conversion will return a pool
355 		 * number or -1 if a fixed (non-negative) value.  The fixed
356 		 * value is checked later when we compare against
357 		 * total memory available.
358 		 */
359 		pool = wildcard_to_pool(size);
360 		if (pool == -1) {			/* non-wildcard */
361 			fixed_blocks += size * count;
362 		} else if (pool < NUM_SC_POOLS) {	/* valid wildcard */
363 			mem_pool_info[pool].count += count;
364 		} else {				/* invalid wildcard */
365 			dd_dev_err(
366 				dd,
367 				"%s send context invalid pool wildcard %d\n",
368 				sc_type_name(i), size);
369 			return -EINVAL;
370 		}
371 
372 		dd->sc_sizes[i].count = count;
373 		dd->sc_sizes[i].size = size;
374 	}
375 	if (fixed_blocks > total_blocks) {
376 		dd_dev_err(
377 			dd,
378 			"Send context fixed block count, %u, larger than total block count %u\n",
379 			fixed_blocks, total_blocks);
380 		return -EINVAL;
381 	}
382 
383 	/* step 3: calculate the blocks in the pools, and pool context sizes */
384 	pool_blocks = total_blocks - fixed_blocks;
385 	if (ab_total > pool_blocks) {
386 		dd_dev_err(
387 			dd,
388 			"Send context fixed pool sizes, %u, larger than pool block count %u\n",
389 			ab_total, pool_blocks);
390 		return -EINVAL;
391 	}
392 	/* subtract off the fixed pool blocks */
393 	pool_blocks -= ab_total;
394 
395 	for (i = 0; i < NUM_SC_POOLS; i++) {
396 		struct mem_pool_info *pi = &mem_pool_info[i];
397 
398 		/* % beats absolute blocks */
399 		if (pi->centipercent >= 0)
400 			pi->blocks = (pool_blocks * pi->centipercent) / 10000;
401 
402 		if (pi->blocks == 0 && pi->count != 0) {
403 			dd_dev_err(
404 				dd,
405 				"Send context memory pool %d has %u contexts, but no blocks\n",
406 				i, pi->count);
407 			return -EINVAL;
408 		}
409 		if (pi->count == 0) {
410 			/* warn about wasted blocks */
411 			if (pi->blocks != 0)
412 				dd_dev_err(
413 					dd,
414 					"Send context memory pool %d has %u blocks, but zero contexts\n",
415 					i, pi->blocks);
416 			pi->size = 0;
417 		} else {
418 			pi->size = pi->blocks / pi->count;
419 		}
420 	}
421 
422 	/* step 4: fill in the context type sizes from the pool sizes */
423 	used_blocks = 0;
424 	for (i = 0; i < SC_MAX; i++) {
425 		if (dd->sc_sizes[i].size < 0) {
426 			unsigned pool = wildcard_to_pool(dd->sc_sizes[i].size);
427 
428 			WARN_ON_ONCE(pool >= NUM_SC_POOLS);
429 			dd->sc_sizes[i].size = mem_pool_info[pool].size;
430 		}
431 		/* make sure we are not larger than what is allowed by the HW */
432 #define PIO_MAX_BLOCKS 1024
433 		if (dd->sc_sizes[i].size > PIO_MAX_BLOCKS)
434 			dd->sc_sizes[i].size = PIO_MAX_BLOCKS;
435 
436 		/* calculate our total usage */
437 		used_blocks += dd->sc_sizes[i].size * dd->sc_sizes[i].count;
438 	}
439 	extra = total_blocks - used_blocks;
440 	if (extra != 0)
441 		dd_dev_info(dd, "unused send context blocks: %d\n", extra);
442 
443 	return total_contexts;
444 }
445 
446 int init_send_contexts(struct hfi1_devdata *dd)
447 {
448 	u16 base;
449 	int ret, i, j, context;
450 
451 	ret = init_credit_return(dd);
452 	if (ret)
453 		return ret;
454 
455 	dd->hw_to_sw = kmalloc_array(TXE_NUM_CONTEXTS, sizeof(u8),
456 					GFP_KERNEL);
457 	dd->send_contexts = kcalloc(dd->num_send_contexts,
458 					sizeof(struct send_context_info),
459 					GFP_KERNEL);
460 	if (!dd->send_contexts || !dd->hw_to_sw) {
461 		kfree(dd->hw_to_sw);
462 		kfree(dd->send_contexts);
463 		free_credit_return(dd);
464 		return -ENOMEM;
465 	}
466 
467 	/* hardware context map starts with invalid send context indices */
468 	for (i = 0; i < TXE_NUM_CONTEXTS; i++)
469 		dd->hw_to_sw[i] = INVALID_SCI;
470 
471 	/*
472 	 * All send contexts have their credit sizes.  Allocate credits
473 	 * for each context one after another from the global space.
474 	 */
475 	context = 0;
476 	base = 1;
477 	for (i = 0; i < SC_MAX; i++) {
478 		struct sc_config_sizes *scs = &dd->sc_sizes[i];
479 
480 		for (j = 0; j < scs->count; j++) {
481 			struct send_context_info *sci =
482 						&dd->send_contexts[context];
483 			sci->type = i;
484 			sci->base = base;
485 			sci->credits = scs->size;
486 
487 			context++;
488 			base += scs->size;
489 		}
490 	}
491 
492 	return 0;
493 }
494 
495 /*
496  * Allocate a software index and hardware context of the given type.
497  *
498  * Must be called with dd->sc_lock held.
499  */
500 static int sc_hw_alloc(struct hfi1_devdata *dd, int type, u32 *sw_index,
501 		       u32 *hw_context)
502 {
503 	struct send_context_info *sci;
504 	u32 index;
505 	u32 context;
506 
507 	for (index = 0, sci = &dd->send_contexts[0];
508 			index < dd->num_send_contexts; index++, sci++) {
509 		if (sci->type == type && sci->allocated == 0) {
510 			sci->allocated = 1;
511 			/* use a 1:1 mapping, but make them non-equal */
512 			context = dd->chip_send_contexts - index - 1;
513 			dd->hw_to_sw[context] = index;
514 			*sw_index = index;
515 			*hw_context = context;
516 			return 0; /* success */
517 		}
518 	}
519 	dd_dev_err(dd, "Unable to locate a free type %d send context\n", type);
520 	return -ENOSPC;
521 }
522 
523 /*
524  * Free the send context given by its software index.
525  *
526  * Must be called with dd->sc_lock held.
527  */
528 static void sc_hw_free(struct hfi1_devdata *dd, u32 sw_index, u32 hw_context)
529 {
530 	struct send_context_info *sci;
531 
532 	sci = &dd->send_contexts[sw_index];
533 	if (!sci->allocated) {
534 		dd_dev_err(dd, "%s: sw_index %u not allocated? hw_context %u\n",
535 			   __func__, sw_index, hw_context);
536 	}
537 	sci->allocated = 0;
538 	dd->hw_to_sw[hw_context] = INVALID_SCI;
539 }
540 
541 /* return the base context of a context in a group */
542 static inline u32 group_context(u32 context, u32 group)
543 {
544 	return (context >> group) << group;
545 }
546 
547 /* return the size of a group */
548 static inline u32 group_size(u32 group)
549 {
550 	return 1 << group;
551 }
552 
553 /*
554  * Obtain the credit return addresses, kernel virtual and bus, for the
555  * given sc.
556  *
557  * To understand this routine:
558  * o va and dma are arrays of struct credit_return.  One for each physical
559  *   send context, per NUMA.
560  * o Each send context always looks in its relative location in a struct
561  *   credit_return for its credit return.
562  * o Each send context in a group must have its return address CSR programmed
563  *   with the same value.  Use the address of the first send context in the
564  *   group.
565  */
566 static void cr_group_addresses(struct send_context *sc, dma_addr_t *dma)
567 {
568 	u32 gc = group_context(sc->hw_context, sc->group);
569 	u32 index = sc->hw_context & 0x7;
570 
571 	sc->hw_free = &sc->dd->cr_base[sc->node].va[gc].cr[index];
572 	*dma = (unsigned long)
573 	       &((struct credit_return *)sc->dd->cr_base[sc->node].dma)[gc];
574 }
575 
576 /*
577  * Work queue function triggered in error interrupt routine for
578  * kernel contexts.
579  */
580 static void sc_halted(struct work_struct *work)
581 {
582 	struct send_context *sc;
583 
584 	sc = container_of(work, struct send_context, halt_work);
585 	sc_restart(sc);
586 }
587 
588 /*
589  * Calculate PIO block threshold for this send context using the given MTU.
590  * Trigger a return when one MTU plus optional header of credits remain.
591  *
592  * Parameter mtu is in bytes.
593  * Parameter hdrqentsize is in DWORDs.
594  *
595  * Return value is what to write into the CSR: trigger return when
596  * unreturned credits pass this count.
597  */
598 u32 sc_mtu_to_threshold(struct send_context *sc, u32 mtu, u32 hdrqentsize)
599 {
600 	u32 release_credits;
601 	u32 threshold;
602 
603 	/* add in the header size, then divide by the PIO block size */
604 	mtu += hdrqentsize << 2;
605 	release_credits = DIV_ROUND_UP(mtu, PIO_BLOCK_SIZE);
606 
607 	/* check against this context's credits */
608 	if (sc->credits <= release_credits)
609 		threshold = 1;
610 	else
611 		threshold = sc->credits - release_credits;
612 
613 	return threshold;
614 }
615 
616 /*
617  * Calculate credit threshold in terms of percent of the allocated credits.
618  * Trigger when unreturned credits equal or exceed the percentage of the whole.
619  *
620  * Return value is what to write into the CSR: trigger return when
621  * unreturned credits pass this count.
622  */
623 u32 sc_percent_to_threshold(struct send_context *sc, u32 percent)
624 {
625 	return (sc->credits * percent) / 100;
626 }
627 
628 /*
629  * Set the credit return threshold.
630  */
631 void sc_set_cr_threshold(struct send_context *sc, u32 new_threshold)
632 {
633 	unsigned long flags;
634 	u32 old_threshold;
635 	int force_return = 0;
636 
637 	spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
638 
639 	old_threshold = (sc->credit_ctrl >>
640 				SC(CREDIT_CTRL_THRESHOLD_SHIFT))
641 			 & SC(CREDIT_CTRL_THRESHOLD_MASK);
642 
643 	if (new_threshold != old_threshold) {
644 		sc->credit_ctrl =
645 			(sc->credit_ctrl
646 				& ~SC(CREDIT_CTRL_THRESHOLD_SMASK))
647 			| ((new_threshold
648 				& SC(CREDIT_CTRL_THRESHOLD_MASK))
649 			   << SC(CREDIT_CTRL_THRESHOLD_SHIFT));
650 		write_kctxt_csr(sc->dd, sc->hw_context,
651 				SC(CREDIT_CTRL), sc->credit_ctrl);
652 
653 		/* force a credit return on change to avoid a possible stall */
654 		force_return = 1;
655 	}
656 
657 	spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
658 
659 	if (force_return)
660 		sc_return_credits(sc);
661 }
662 
663 /*
664  * set_pio_integrity
665  *
666  * Set the CHECK_ENABLE register for the send context 'sc'.
667  */
668 void set_pio_integrity(struct send_context *sc)
669 {
670 	struct hfi1_devdata *dd = sc->dd;
671 	u32 hw_context = sc->hw_context;
672 	int type = sc->type;
673 
674 	write_kctxt_csr(dd, hw_context,
675 			SC(CHECK_ENABLE),
676 			hfi1_pkt_default_send_ctxt_mask(dd, type));
677 }
678 
679 static u32 get_buffers_allocated(struct send_context *sc)
680 {
681 	int cpu;
682 	u32 ret = 0;
683 
684 	for_each_possible_cpu(cpu)
685 		ret += *per_cpu_ptr(sc->buffers_allocated, cpu);
686 	return ret;
687 }
688 
689 static void reset_buffers_allocated(struct send_context *sc)
690 {
691 	int cpu;
692 
693 	for_each_possible_cpu(cpu)
694 		(*per_cpu_ptr(sc->buffers_allocated, cpu)) = 0;
695 }
696 
697 /*
698  * Allocate a NUMA relative send context structure of the given type along
699  * with a HW context.
700  */
701 struct send_context *sc_alloc(struct hfi1_devdata *dd, int type,
702 			      uint hdrqentsize, int numa)
703 {
704 	struct send_context_info *sci;
705 	struct send_context *sc = NULL;
706 	int req_type = type;
707 	dma_addr_t dma;
708 	unsigned long flags;
709 	u64 reg;
710 	u32 thresh;
711 	u32 sw_index;
712 	u32 hw_context;
713 	int ret;
714 	u8 opval, opmask;
715 
716 	/* do not allocate while frozen */
717 	if (dd->flags & HFI1_FROZEN)
718 		return NULL;
719 
720 	sc = kzalloc_node(sizeof(*sc), GFP_KERNEL, numa);
721 	if (!sc)
722 		return NULL;
723 
724 	sc->buffers_allocated = alloc_percpu(u32);
725 	if (!sc->buffers_allocated) {
726 		kfree(sc);
727 		dd_dev_err(dd,
728 			   "Cannot allocate buffers_allocated per cpu counters\n"
729 			  );
730 		return NULL;
731 	}
732 
733 	/*
734 	 * VNIC contexts are dynamically allocated.
735 	 * Hence, pick a user context for VNIC.
736 	 */
737 	if (type == SC_VNIC)
738 		type = SC_USER;
739 
740 	spin_lock_irqsave(&dd->sc_lock, flags);
741 	ret = sc_hw_alloc(dd, type, &sw_index, &hw_context);
742 	if (ret) {
743 		spin_unlock_irqrestore(&dd->sc_lock, flags);
744 		free_percpu(sc->buffers_allocated);
745 		kfree(sc);
746 		return NULL;
747 	}
748 
749 	/*
750 	 * VNIC contexts are used by kernel driver.
751 	 * Hence, mark them as kernel contexts.
752 	 */
753 	if (req_type == SC_VNIC) {
754 		dd->send_contexts[sw_index].type = SC_KERNEL;
755 		type = SC_KERNEL;
756 	}
757 
758 	sci = &dd->send_contexts[sw_index];
759 	sci->sc = sc;
760 
761 	sc->dd = dd;
762 	sc->node = numa;
763 	sc->type = type;
764 	spin_lock_init(&sc->alloc_lock);
765 	spin_lock_init(&sc->release_lock);
766 	spin_lock_init(&sc->credit_ctrl_lock);
767 	INIT_LIST_HEAD(&sc->piowait);
768 	INIT_WORK(&sc->halt_work, sc_halted);
769 	init_waitqueue_head(&sc->halt_wait);
770 
771 	/* grouping is always single context for now */
772 	sc->group = 0;
773 
774 	sc->sw_index = sw_index;
775 	sc->hw_context = hw_context;
776 	cr_group_addresses(sc, &dma);
777 	sc->credits = sci->credits;
778 	sc->size = sc->credits * PIO_BLOCK_SIZE;
779 
780 /* PIO Send Memory Address details */
781 #define PIO_ADDR_CONTEXT_MASK 0xfful
782 #define PIO_ADDR_CONTEXT_SHIFT 16
783 	sc->base_addr = dd->piobase + ((hw_context & PIO_ADDR_CONTEXT_MASK)
784 					<< PIO_ADDR_CONTEXT_SHIFT);
785 
786 	/* set base and credits */
787 	reg = ((sci->credits & SC(CTRL_CTXT_DEPTH_MASK))
788 					<< SC(CTRL_CTXT_DEPTH_SHIFT))
789 		| ((sci->base & SC(CTRL_CTXT_BASE_MASK))
790 					<< SC(CTRL_CTXT_BASE_SHIFT));
791 	write_kctxt_csr(dd, hw_context, SC(CTRL), reg);
792 
793 	set_pio_integrity(sc);
794 
795 	/* unmask all errors */
796 	write_kctxt_csr(dd, hw_context, SC(ERR_MASK), (u64)-1);
797 
798 	/* set the default partition key */
799 	write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY),
800 			(SC(CHECK_PARTITION_KEY_VALUE_MASK) &
801 			 DEFAULT_PKEY) <<
802 			SC(CHECK_PARTITION_KEY_VALUE_SHIFT));
803 
804 	/* per context type checks */
805 	if (type == SC_USER) {
806 		opval = USER_OPCODE_CHECK_VAL;
807 		opmask = USER_OPCODE_CHECK_MASK;
808 	} else {
809 		opval = OPCODE_CHECK_VAL_DISABLED;
810 		opmask = OPCODE_CHECK_MASK_DISABLED;
811 	}
812 
813 	/* set the send context check opcode mask and value */
814 	write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE),
815 			((u64)opmask << SC(CHECK_OPCODE_MASK_SHIFT)) |
816 			((u64)opval << SC(CHECK_OPCODE_VALUE_SHIFT)));
817 
818 	/* set up credit return */
819 	reg = dma & SC(CREDIT_RETURN_ADDR_ADDRESS_SMASK);
820 	write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), reg);
821 
822 	/*
823 	 * Calculate the initial credit return threshold.
824 	 *
825 	 * For Ack contexts, set a threshold for half the credits.
826 	 * For User contexts use the given percentage.  This has been
827 	 * sanitized on driver start-up.
828 	 * For Kernel contexts, use the default MTU plus a header
829 	 * or half the credits, whichever is smaller. This should
830 	 * work for both the 3-deep buffering allocation and the
831 	 * pooling allocation.
832 	 */
833 	if (type == SC_ACK) {
834 		thresh = sc_percent_to_threshold(sc, 50);
835 	} else if (type == SC_USER) {
836 		thresh = sc_percent_to_threshold(sc,
837 						 user_credit_return_threshold);
838 	} else { /* kernel */
839 		thresh = min(sc_percent_to_threshold(sc, 50),
840 			     sc_mtu_to_threshold(sc, hfi1_max_mtu,
841 						 hdrqentsize));
842 	}
843 	reg = thresh << SC(CREDIT_CTRL_THRESHOLD_SHIFT);
844 	/* add in early return */
845 	if (type == SC_USER && HFI1_CAP_IS_USET(EARLY_CREDIT_RETURN))
846 		reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK);
847 	else if (HFI1_CAP_IS_KSET(EARLY_CREDIT_RETURN)) /* kernel, ack */
848 		reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK);
849 
850 	/* set up write-through credit_ctrl */
851 	sc->credit_ctrl = reg;
852 	write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), reg);
853 
854 	/* User send contexts should not allow sending on VL15 */
855 	if (type == SC_USER) {
856 		reg = 1ULL << 15;
857 		write_kctxt_csr(dd, hw_context, SC(CHECK_VL), reg);
858 	}
859 
860 	spin_unlock_irqrestore(&dd->sc_lock, flags);
861 
862 	/*
863 	 * Allocate shadow ring to track outstanding PIO buffers _after_
864 	 * unlocking.  We don't know the size until the lock is held and
865 	 * we can't allocate while the lock is held.  No one is using
866 	 * the context yet, so allocate it now.
867 	 *
868 	 * User contexts do not get a shadow ring.
869 	 */
870 	if (type != SC_USER) {
871 		/*
872 		 * Size the shadow ring 1 larger than the number of credits
873 		 * so head == tail can mean empty.
874 		 */
875 		sc->sr_size = sci->credits + 1;
876 		sc->sr = kzalloc_node(sizeof(union pio_shadow_ring) *
877 				sc->sr_size, GFP_KERNEL, numa);
878 		if (!sc->sr) {
879 			sc_free(sc);
880 			return NULL;
881 		}
882 	}
883 
884 	hfi1_cdbg(PIO,
885 		  "Send context %u(%u) %s group %u credits %u credit_ctrl 0x%llx threshold %u\n",
886 		  sw_index,
887 		  hw_context,
888 		  sc_type_name(type),
889 		  sc->group,
890 		  sc->credits,
891 		  sc->credit_ctrl,
892 		  thresh);
893 
894 	return sc;
895 }
896 
897 /* free a per-NUMA send context structure */
898 void sc_free(struct send_context *sc)
899 {
900 	struct hfi1_devdata *dd;
901 	unsigned long flags;
902 	u32 sw_index;
903 	u32 hw_context;
904 
905 	if (!sc)
906 		return;
907 
908 	sc->flags |= SCF_IN_FREE;	/* ensure no restarts */
909 	dd = sc->dd;
910 	if (!list_empty(&sc->piowait))
911 		dd_dev_err(dd, "piowait list not empty!\n");
912 	sw_index = sc->sw_index;
913 	hw_context = sc->hw_context;
914 	sc_disable(sc);	/* make sure the HW is disabled */
915 	flush_work(&sc->halt_work);
916 
917 	spin_lock_irqsave(&dd->sc_lock, flags);
918 	dd->send_contexts[sw_index].sc = NULL;
919 
920 	/* clear/disable all registers set in sc_alloc */
921 	write_kctxt_csr(dd, hw_context, SC(CTRL), 0);
922 	write_kctxt_csr(dd, hw_context, SC(CHECK_ENABLE), 0);
923 	write_kctxt_csr(dd, hw_context, SC(ERR_MASK), 0);
924 	write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY), 0);
925 	write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE), 0);
926 	write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), 0);
927 	write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), 0);
928 
929 	/* release the index and context for re-use */
930 	sc_hw_free(dd, sw_index, hw_context);
931 	spin_unlock_irqrestore(&dd->sc_lock, flags);
932 
933 	kfree(sc->sr);
934 	free_percpu(sc->buffers_allocated);
935 	kfree(sc);
936 }
937 
938 /* disable the context */
939 void sc_disable(struct send_context *sc)
940 {
941 	u64 reg;
942 	unsigned long flags;
943 	struct pio_buf *pbuf;
944 
945 	if (!sc)
946 		return;
947 
948 	/* do all steps, even if already disabled */
949 	spin_lock_irqsave(&sc->alloc_lock, flags);
950 	reg = read_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL));
951 	reg &= ~SC(CTRL_CTXT_ENABLE_SMASK);
952 	sc->flags &= ~SCF_ENABLED;
953 	sc_wait_for_packet_egress(sc, 1);
954 	write_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL), reg);
955 	spin_unlock_irqrestore(&sc->alloc_lock, flags);
956 
957 	/*
958 	 * Flush any waiters.  Once the context is disabled,
959 	 * credit return interrupts are stopped (although there
960 	 * could be one in-process when the context is disabled).
961 	 * Wait one microsecond for any lingering interrupts, then
962 	 * proceed with the flush.
963 	 */
964 	udelay(1);
965 	spin_lock_irqsave(&sc->release_lock, flags);
966 	if (sc->sr) {	/* this context has a shadow ring */
967 		while (sc->sr_tail != sc->sr_head) {
968 			pbuf = &sc->sr[sc->sr_tail].pbuf;
969 			if (pbuf->cb)
970 				(*pbuf->cb)(pbuf->arg, PRC_SC_DISABLE);
971 			sc->sr_tail++;
972 			if (sc->sr_tail >= sc->sr_size)
973 				sc->sr_tail = 0;
974 		}
975 	}
976 	spin_unlock_irqrestore(&sc->release_lock, flags);
977 }
978 
979 /* return SendEgressCtxtStatus.PacketOccupancy */
980 #define packet_occupancy(r) \
981 	(((r) & SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SMASK)\
982 	>> SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SHIFT)
983 
984 /* is egress halted on the context? */
985 #define egress_halted(r) \
986 	((r) & SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_HALT_STATUS_SMASK)
987 
988 /* wait for packet egress, optionally pause for credit return  */
989 static void sc_wait_for_packet_egress(struct send_context *sc, int pause)
990 {
991 	struct hfi1_devdata *dd = sc->dd;
992 	u64 reg = 0;
993 	u64 reg_prev;
994 	u32 loop = 0;
995 
996 	while (1) {
997 		reg_prev = reg;
998 		reg = read_csr(dd, sc->hw_context * 8 +
999 			       SEND_EGRESS_CTXT_STATUS);
1000 		/* done if egress is stopped */
1001 		if (egress_halted(reg))
1002 			break;
1003 		reg = packet_occupancy(reg);
1004 		if (reg == 0)
1005 			break;
1006 		/* counter is reset if occupancy count changes */
1007 		if (reg != reg_prev)
1008 			loop = 0;
1009 		if (loop > 50000) {
1010 			/* timed out - bounce the link */
1011 			dd_dev_err(dd,
1012 				   "%s: context %u(%u) timeout waiting for packets to egress, remaining count %u, bouncing link\n",
1013 				   __func__, sc->sw_index,
1014 				   sc->hw_context, (u32)reg);
1015 			queue_work(dd->pport->hfi1_wq,
1016 				   &dd->pport->link_bounce_work);
1017 			break;
1018 		}
1019 		loop++;
1020 		udelay(1);
1021 	}
1022 
1023 	if (pause)
1024 		/* Add additional delay to ensure chip returns all credits */
1025 		pause_for_credit_return(dd);
1026 }
1027 
1028 void sc_wait(struct hfi1_devdata *dd)
1029 {
1030 	int i;
1031 
1032 	for (i = 0; i < dd->num_send_contexts; i++) {
1033 		struct send_context *sc = dd->send_contexts[i].sc;
1034 
1035 		if (!sc)
1036 			continue;
1037 		sc_wait_for_packet_egress(sc, 0);
1038 	}
1039 }
1040 
1041 /*
1042  * Restart a context after it has been halted due to error.
1043  *
1044  * If the first step fails - wait for the halt to be asserted, return early.
1045  * Otherwise complain about timeouts but keep going.
1046  *
1047  * It is expected that allocations (enabled flag bit) have been shut off
1048  * already (only applies to kernel contexts).
1049  */
1050 int sc_restart(struct send_context *sc)
1051 {
1052 	struct hfi1_devdata *dd = sc->dd;
1053 	u64 reg;
1054 	u32 loop;
1055 	int count;
1056 
1057 	/* bounce off if not halted, or being free'd */
1058 	if (!(sc->flags & SCF_HALTED) || (sc->flags & SCF_IN_FREE))
1059 		return -EINVAL;
1060 
1061 	dd_dev_info(dd, "restarting send context %u(%u)\n", sc->sw_index,
1062 		    sc->hw_context);
1063 
1064 	/*
1065 	 * Step 1: Wait for the context to actually halt.
1066 	 *
1067 	 * The error interrupt is asynchronous to actually setting halt
1068 	 * on the context.
1069 	 */
1070 	loop = 0;
1071 	while (1) {
1072 		reg = read_kctxt_csr(dd, sc->hw_context, SC(STATUS));
1073 		if (reg & SC(STATUS_CTXT_HALTED_SMASK))
1074 			break;
1075 		if (loop > 100) {
1076 			dd_dev_err(dd, "%s: context %u(%u) not halting, skipping\n",
1077 				   __func__, sc->sw_index, sc->hw_context);
1078 			return -ETIME;
1079 		}
1080 		loop++;
1081 		udelay(1);
1082 	}
1083 
1084 	/*
1085 	 * Step 2: Ensure no users are still trying to write to PIO.
1086 	 *
1087 	 * For kernel contexts, we have already turned off buffer allocation.
1088 	 * Now wait for the buffer count to go to zero.
1089 	 *
1090 	 * For user contexts, the user handling code has cut off write access
1091 	 * to the context's PIO pages before calling this routine and will
1092 	 * restore write access after this routine returns.
1093 	 */
1094 	if (sc->type != SC_USER) {
1095 		/* kernel context */
1096 		loop = 0;
1097 		while (1) {
1098 			count = get_buffers_allocated(sc);
1099 			if (count == 0)
1100 				break;
1101 			if (loop > 100) {
1102 				dd_dev_err(dd,
1103 					   "%s: context %u(%u) timeout waiting for PIO buffers to zero, remaining %d\n",
1104 					   __func__, sc->sw_index,
1105 					   sc->hw_context, count);
1106 			}
1107 			loop++;
1108 			udelay(1);
1109 		}
1110 	}
1111 
1112 	/*
1113 	 * Step 3: Wait for all packets to egress.
1114 	 * This is done while disabling the send context
1115 	 *
1116 	 * Step 4: Disable the context
1117 	 *
1118 	 * This is a superset of the halt.  After the disable, the
1119 	 * errors can be cleared.
1120 	 */
1121 	sc_disable(sc);
1122 
1123 	/*
1124 	 * Step 5: Enable the context
1125 	 *
1126 	 * This enable will clear the halted flag and per-send context
1127 	 * error flags.
1128 	 */
1129 	return sc_enable(sc);
1130 }
1131 
1132 /*
1133  * PIO freeze processing.  To be called after the TXE block is fully frozen.
1134  * Go through all frozen send contexts and disable them.  The contexts are
1135  * already stopped by the freeze.
1136  */
1137 void pio_freeze(struct hfi1_devdata *dd)
1138 {
1139 	struct send_context *sc;
1140 	int i;
1141 
1142 	for (i = 0; i < dd->num_send_contexts; i++) {
1143 		sc = dd->send_contexts[i].sc;
1144 		/*
1145 		 * Don't disable unallocated, unfrozen, or user send contexts.
1146 		 * User send contexts will be disabled when the process
1147 		 * calls into the driver to reset its context.
1148 		 */
1149 		if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER)
1150 			continue;
1151 
1152 		/* only need to disable, the context is already stopped */
1153 		sc_disable(sc);
1154 	}
1155 }
1156 
1157 /*
1158  * Unfreeze PIO for kernel send contexts.  The precondition for calling this
1159  * is that all PIO send contexts have been disabled and the SPC freeze has
1160  * been cleared.  Now perform the last step and re-enable each kernel context.
1161  * User (PSM) processing will occur when PSM calls into the kernel to
1162  * acknowledge the freeze.
1163  */
1164 void pio_kernel_unfreeze(struct hfi1_devdata *dd)
1165 {
1166 	struct send_context *sc;
1167 	int i;
1168 
1169 	for (i = 0; i < dd->num_send_contexts; i++) {
1170 		sc = dd->send_contexts[i].sc;
1171 		if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER)
1172 			continue;
1173 
1174 		sc_enable(sc);	/* will clear the sc frozen flag */
1175 	}
1176 }
1177 
1178 /*
1179  * Wait for the SendPioInitCtxt.PioInitInProgress bit to clear.
1180  * Returns:
1181  *	-ETIMEDOUT - if we wait too long
1182  *	-EIO	   - if there was an error
1183  */
1184 static int pio_init_wait_progress(struct hfi1_devdata *dd)
1185 {
1186 	u64 reg;
1187 	int max, count = 0;
1188 
1189 	/* max is the longest possible HW init time / delay */
1190 	max = (dd->icode == ICODE_FPGA_EMULATION) ? 120 : 5;
1191 	while (1) {
1192 		reg = read_csr(dd, SEND_PIO_INIT_CTXT);
1193 		if (!(reg & SEND_PIO_INIT_CTXT_PIO_INIT_IN_PROGRESS_SMASK))
1194 			break;
1195 		if (count >= max)
1196 			return -ETIMEDOUT;
1197 		udelay(5);
1198 		count++;
1199 	}
1200 
1201 	return reg & SEND_PIO_INIT_CTXT_PIO_INIT_ERR_SMASK ? -EIO : 0;
1202 }
1203 
1204 /*
1205  * Reset all of the send contexts to their power-on state.  Used
1206  * only during manual init - no lock against sc_enable needed.
1207  */
1208 void pio_reset_all(struct hfi1_devdata *dd)
1209 {
1210 	int ret;
1211 
1212 	/* make sure the init engine is not busy */
1213 	ret = pio_init_wait_progress(dd);
1214 	/* ignore any timeout */
1215 	if (ret == -EIO) {
1216 		/* clear the error */
1217 		write_csr(dd, SEND_PIO_ERR_CLEAR,
1218 			  SEND_PIO_ERR_CLEAR_PIO_INIT_SM_IN_ERR_SMASK);
1219 	}
1220 
1221 	/* reset init all */
1222 	write_csr(dd, SEND_PIO_INIT_CTXT,
1223 		  SEND_PIO_INIT_CTXT_PIO_ALL_CTXT_INIT_SMASK);
1224 	udelay(2);
1225 	ret = pio_init_wait_progress(dd);
1226 	if (ret < 0) {
1227 		dd_dev_err(dd,
1228 			   "PIO send context init %s while initializing all PIO blocks\n",
1229 			   ret == -ETIMEDOUT ? "is stuck" : "had an error");
1230 	}
1231 }
1232 
1233 /* enable the context */
1234 int sc_enable(struct send_context *sc)
1235 {
1236 	u64 sc_ctrl, reg, pio;
1237 	struct hfi1_devdata *dd;
1238 	unsigned long flags;
1239 	int ret = 0;
1240 
1241 	if (!sc)
1242 		return -EINVAL;
1243 	dd = sc->dd;
1244 
1245 	/*
1246 	 * Obtain the allocator lock to guard against any allocation
1247 	 * attempts (which should not happen prior to context being
1248 	 * enabled). On the release/disable side we don't need to
1249 	 * worry about locking since the releaser will not do anything
1250 	 * if the context accounting values have not changed.
1251 	 */
1252 	spin_lock_irqsave(&sc->alloc_lock, flags);
1253 	sc_ctrl = read_kctxt_csr(dd, sc->hw_context, SC(CTRL));
1254 	if ((sc_ctrl & SC(CTRL_CTXT_ENABLE_SMASK)))
1255 		goto unlock; /* already enabled */
1256 
1257 	/* IMPORTANT: only clear free and fill if transitioning 0 -> 1 */
1258 
1259 	*sc->hw_free = 0;
1260 	sc->free = 0;
1261 	sc->alloc_free = 0;
1262 	sc->fill = 0;
1263 	sc->fill_wrap = 0;
1264 	sc->sr_head = 0;
1265 	sc->sr_tail = 0;
1266 	sc->flags = 0;
1267 	/* the alloc lock insures no fast path allocation */
1268 	reset_buffers_allocated(sc);
1269 
1270 	/*
1271 	 * Clear all per-context errors.  Some of these will be set when
1272 	 * we are re-enabling after a context halt.  Now that the context
1273 	 * is disabled, the halt will not clear until after the PIO init
1274 	 * engine runs below.
1275 	 */
1276 	reg = read_kctxt_csr(dd, sc->hw_context, SC(ERR_STATUS));
1277 	if (reg)
1278 		write_kctxt_csr(dd, sc->hw_context, SC(ERR_CLEAR), reg);
1279 
1280 	/*
1281 	 * The HW PIO initialization engine can handle only one init
1282 	 * request at a time. Serialize access to each device's engine.
1283 	 */
1284 	spin_lock(&dd->sc_init_lock);
1285 	/*
1286 	 * Since access to this code block is serialized and
1287 	 * each access waits for the initialization to complete
1288 	 * before releasing the lock, the PIO initialization engine
1289 	 * should not be in use, so we don't have to wait for the
1290 	 * InProgress bit to go down.
1291 	 */
1292 	pio = ((sc->hw_context & SEND_PIO_INIT_CTXT_PIO_CTXT_NUM_MASK) <<
1293 	       SEND_PIO_INIT_CTXT_PIO_CTXT_NUM_SHIFT) |
1294 		SEND_PIO_INIT_CTXT_PIO_SINGLE_CTXT_INIT_SMASK;
1295 	write_csr(dd, SEND_PIO_INIT_CTXT, pio);
1296 	/*
1297 	 * Wait until the engine is done.  Give the chip the required time
1298 	 * so, hopefully, we read the register just once.
1299 	 */
1300 	udelay(2);
1301 	ret = pio_init_wait_progress(dd);
1302 	spin_unlock(&dd->sc_init_lock);
1303 	if (ret) {
1304 		dd_dev_err(dd,
1305 			   "sctxt%u(%u): Context not enabled due to init failure %d\n",
1306 			   sc->sw_index, sc->hw_context, ret);
1307 		goto unlock;
1308 	}
1309 
1310 	/*
1311 	 * All is well. Enable the context.
1312 	 */
1313 	sc_ctrl |= SC(CTRL_CTXT_ENABLE_SMASK);
1314 	write_kctxt_csr(dd, sc->hw_context, SC(CTRL), sc_ctrl);
1315 	/*
1316 	 * Read SendCtxtCtrl to force the write out and prevent a timing
1317 	 * hazard where a PIO write may reach the context before the enable.
1318 	 */
1319 	read_kctxt_csr(dd, sc->hw_context, SC(CTRL));
1320 	sc->flags |= SCF_ENABLED;
1321 
1322 unlock:
1323 	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1324 
1325 	return ret;
1326 }
1327 
1328 /* force a credit return on the context */
1329 void sc_return_credits(struct send_context *sc)
1330 {
1331 	if (!sc)
1332 		return;
1333 
1334 	/* a 0->1 transition schedules a credit return */
1335 	write_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE),
1336 			SC(CREDIT_FORCE_FORCE_RETURN_SMASK));
1337 	/*
1338 	 * Ensure that the write is flushed and the credit return is
1339 	 * scheduled. We care more about the 0 -> 1 transition.
1340 	 */
1341 	read_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE));
1342 	/* set back to 0 for next time */
1343 	write_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE), 0);
1344 }
1345 
1346 /* allow all in-flight packets to drain on the context */
1347 void sc_flush(struct send_context *sc)
1348 {
1349 	if (!sc)
1350 		return;
1351 
1352 	sc_wait_for_packet_egress(sc, 1);
1353 }
1354 
1355 /* drop all packets on the context, no waiting until they are sent */
1356 void sc_drop(struct send_context *sc)
1357 {
1358 	if (!sc)
1359 		return;
1360 
1361 	dd_dev_info(sc->dd, "%s: context %u(%u) - not implemented\n",
1362 		    __func__, sc->sw_index, sc->hw_context);
1363 }
1364 
1365 /*
1366  * Start the software reaction to a context halt or SPC freeze:
1367  *	- mark the context as halted or frozen
1368  *	- stop buffer allocations
1369  *
1370  * Called from the error interrupt.  Other work is deferred until
1371  * out of the interrupt.
1372  */
1373 void sc_stop(struct send_context *sc, int flag)
1374 {
1375 	unsigned long flags;
1376 
1377 	/* mark the context */
1378 	sc->flags |= flag;
1379 
1380 	/* stop buffer allocations */
1381 	spin_lock_irqsave(&sc->alloc_lock, flags);
1382 	sc->flags &= ~SCF_ENABLED;
1383 	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1384 	wake_up(&sc->halt_wait);
1385 }
1386 
1387 #define BLOCK_DWORDS (PIO_BLOCK_SIZE / sizeof(u32))
1388 #define dwords_to_blocks(x) DIV_ROUND_UP(x, BLOCK_DWORDS)
1389 
1390 /*
1391  * The send context buffer "allocator".
1392  *
1393  * @sc: the PIO send context we are allocating from
1394  * @len: length of whole packet - including PBC - in dwords
1395  * @cb: optional callback to call when the buffer is finished sending
1396  * @arg: argument for cb
1397  *
1398  * Return a pointer to a PIO buffer if successful, NULL if not enough room.
1399  */
1400 struct pio_buf *sc_buffer_alloc(struct send_context *sc, u32 dw_len,
1401 				pio_release_cb cb, void *arg)
1402 {
1403 	struct pio_buf *pbuf = NULL;
1404 	unsigned long flags;
1405 	unsigned long avail;
1406 	unsigned long blocks = dwords_to_blocks(dw_len);
1407 	u32 fill_wrap;
1408 	int trycount = 0;
1409 	u32 head, next;
1410 
1411 	spin_lock_irqsave(&sc->alloc_lock, flags);
1412 	if (!(sc->flags & SCF_ENABLED)) {
1413 		spin_unlock_irqrestore(&sc->alloc_lock, flags);
1414 		goto done;
1415 	}
1416 
1417 retry:
1418 	avail = (unsigned long)sc->credits - (sc->fill - sc->alloc_free);
1419 	if (blocks > avail) {
1420 		/* not enough room */
1421 		if (unlikely(trycount))	{ /* already tried to get more room */
1422 			spin_unlock_irqrestore(&sc->alloc_lock, flags);
1423 			goto done;
1424 		}
1425 		/* copy from receiver cache line and recalculate */
1426 		sc->alloc_free = ACCESS_ONCE(sc->free);
1427 		avail =
1428 			(unsigned long)sc->credits -
1429 			(sc->fill - sc->alloc_free);
1430 		if (blocks > avail) {
1431 			/* still no room, actively update */
1432 			sc_release_update(sc);
1433 			sc->alloc_free = ACCESS_ONCE(sc->free);
1434 			trycount++;
1435 			goto retry;
1436 		}
1437 	}
1438 
1439 	/* there is enough room */
1440 
1441 	preempt_disable();
1442 	this_cpu_inc(*sc->buffers_allocated);
1443 
1444 	/* read this once */
1445 	head = sc->sr_head;
1446 
1447 	/* "allocate" the buffer */
1448 	sc->fill += blocks;
1449 	fill_wrap = sc->fill_wrap;
1450 	sc->fill_wrap += blocks;
1451 	if (sc->fill_wrap >= sc->credits)
1452 		sc->fill_wrap = sc->fill_wrap - sc->credits;
1453 
1454 	/*
1455 	 * Fill the parts that the releaser looks at before moving the head.
1456 	 * The only necessary piece is the sent_at field.  The credits
1457 	 * we have just allocated cannot have been returned yet, so the
1458 	 * cb and arg will not be looked at for a "while".  Put them
1459 	 * on this side of the memory barrier anyway.
1460 	 */
1461 	pbuf = &sc->sr[head].pbuf;
1462 	pbuf->sent_at = sc->fill;
1463 	pbuf->cb = cb;
1464 	pbuf->arg = arg;
1465 	pbuf->sc = sc;	/* could be filled in at sc->sr init time */
1466 	/* make sure this is in memory before updating the head */
1467 
1468 	/* calculate next head index, do not store */
1469 	next = head + 1;
1470 	if (next >= sc->sr_size)
1471 		next = 0;
1472 	/*
1473 	 * update the head - must be last! - the releaser can look at fields
1474 	 * in pbuf once we move the head
1475 	 */
1476 	smp_wmb();
1477 	sc->sr_head = next;
1478 	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1479 
1480 	/* finish filling in the buffer outside the lock */
1481 	pbuf->start = sc->base_addr + fill_wrap * PIO_BLOCK_SIZE;
1482 	pbuf->end = sc->base_addr + sc->size;
1483 	pbuf->qw_written = 0;
1484 	pbuf->carry_bytes = 0;
1485 	pbuf->carry.val64 = 0;
1486 done:
1487 	return pbuf;
1488 }
1489 
1490 /*
1491  * There are at least two entities that can turn on credit return
1492  * interrupts and they can overlap.  Avoid problems by implementing
1493  * a count scheme that is enforced by a lock.  The lock is needed because
1494  * the count and CSR write must be paired.
1495  */
1496 
1497 /*
1498  * Start credit return interrupts.  This is managed by a count.  If already
1499  * on, just increment the count.
1500  */
1501 void sc_add_credit_return_intr(struct send_context *sc)
1502 {
1503 	unsigned long flags;
1504 
1505 	/* lock must surround both the count change and the CSR update */
1506 	spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
1507 	if (sc->credit_intr_count == 0) {
1508 		sc->credit_ctrl |= SC(CREDIT_CTRL_CREDIT_INTR_SMASK);
1509 		write_kctxt_csr(sc->dd, sc->hw_context,
1510 				SC(CREDIT_CTRL), sc->credit_ctrl);
1511 	}
1512 	sc->credit_intr_count++;
1513 	spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
1514 }
1515 
1516 /*
1517  * Stop credit return interrupts.  This is managed by a count.  Decrement the
1518  * count, if the last user, then turn the credit interrupts off.
1519  */
1520 void sc_del_credit_return_intr(struct send_context *sc)
1521 {
1522 	unsigned long flags;
1523 
1524 	WARN_ON(sc->credit_intr_count == 0);
1525 
1526 	/* lock must surround both the count change and the CSR update */
1527 	spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
1528 	sc->credit_intr_count--;
1529 	if (sc->credit_intr_count == 0) {
1530 		sc->credit_ctrl &= ~SC(CREDIT_CTRL_CREDIT_INTR_SMASK);
1531 		write_kctxt_csr(sc->dd, sc->hw_context,
1532 				SC(CREDIT_CTRL), sc->credit_ctrl);
1533 	}
1534 	spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
1535 }
1536 
1537 /*
1538  * The caller must be careful when calling this.  All needint calls
1539  * must be paired with !needint.
1540  */
1541 void hfi1_sc_wantpiobuf_intr(struct send_context *sc, u32 needint)
1542 {
1543 	if (needint)
1544 		sc_add_credit_return_intr(sc);
1545 	else
1546 		sc_del_credit_return_intr(sc);
1547 	trace_hfi1_wantpiointr(sc, needint, sc->credit_ctrl);
1548 	if (needint) {
1549 		mmiowb();
1550 		sc_return_credits(sc);
1551 	}
1552 }
1553 
1554 /**
1555  * sc_piobufavail - callback when a PIO buffer is available
1556  * @sc: the send context
1557  *
1558  * This is called from the interrupt handler when a PIO buffer is
1559  * available after hfi1_verbs_send() returned an error that no buffers were
1560  * available. Disable the interrupt if there are no more QPs waiting.
1561  */
1562 static void sc_piobufavail(struct send_context *sc)
1563 {
1564 	struct hfi1_devdata *dd = sc->dd;
1565 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1566 	struct list_head *list;
1567 	struct rvt_qp *qps[PIO_WAIT_BATCH_SIZE];
1568 	struct rvt_qp *qp;
1569 	struct hfi1_qp_priv *priv;
1570 	unsigned long flags;
1571 	unsigned i, n = 0;
1572 
1573 	if (dd->send_contexts[sc->sw_index].type != SC_KERNEL &&
1574 	    dd->send_contexts[sc->sw_index].type != SC_VL15)
1575 		return;
1576 	list = &sc->piowait;
1577 	/*
1578 	 * Note: checking that the piowait list is empty and clearing
1579 	 * the buffer available interrupt needs to be atomic or we
1580 	 * could end up with QPs on the wait list with the interrupt
1581 	 * disabled.
1582 	 */
1583 	write_seqlock_irqsave(&dev->iowait_lock, flags);
1584 	while (!list_empty(list)) {
1585 		struct iowait *wait;
1586 
1587 		if (n == ARRAY_SIZE(qps))
1588 			break;
1589 		wait = list_first_entry(list, struct iowait, list);
1590 		qp = iowait_to_qp(wait);
1591 		priv = qp->priv;
1592 		list_del_init(&priv->s_iowait.list);
1593 		priv->s_iowait.lock = NULL;
1594 		/* refcount held until actual wake up */
1595 		qps[n++] = qp;
1596 	}
1597 	/*
1598 	 * If there had been waiters and there are more
1599 	 * insure that we redo the force to avoid a potential hang.
1600 	 */
1601 	if (n) {
1602 		hfi1_sc_wantpiobuf_intr(sc, 0);
1603 		if (!list_empty(list))
1604 			hfi1_sc_wantpiobuf_intr(sc, 1);
1605 	}
1606 	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
1607 
1608 	for (i = 0; i < n; i++)
1609 		hfi1_qp_wakeup(qps[i],
1610 			       RVT_S_WAIT_PIO | RVT_S_WAIT_PIO_DRAIN);
1611 }
1612 
1613 /* translate a send credit update to a bit code of reasons */
1614 static inline int fill_code(u64 hw_free)
1615 {
1616 	int code = 0;
1617 
1618 	if (hw_free & CR_STATUS_SMASK)
1619 		code |= PRC_STATUS_ERR;
1620 	if (hw_free & CR_CREDIT_RETURN_DUE_TO_PBC_SMASK)
1621 		code |= PRC_PBC;
1622 	if (hw_free & CR_CREDIT_RETURN_DUE_TO_THRESHOLD_SMASK)
1623 		code |= PRC_THRESHOLD;
1624 	if (hw_free & CR_CREDIT_RETURN_DUE_TO_ERR_SMASK)
1625 		code |= PRC_FILL_ERR;
1626 	if (hw_free & CR_CREDIT_RETURN_DUE_TO_FORCE_SMASK)
1627 		code |= PRC_SC_DISABLE;
1628 	return code;
1629 }
1630 
1631 /* use the jiffies compare to get the wrap right */
1632 #define sent_before(a, b) time_before(a, b)	/* a < b */
1633 
1634 /*
1635  * The send context buffer "releaser".
1636  */
1637 void sc_release_update(struct send_context *sc)
1638 {
1639 	struct pio_buf *pbuf;
1640 	u64 hw_free;
1641 	u32 head, tail;
1642 	unsigned long old_free;
1643 	unsigned long free;
1644 	unsigned long extra;
1645 	unsigned long flags;
1646 	int code;
1647 
1648 	if (!sc)
1649 		return;
1650 
1651 	spin_lock_irqsave(&sc->release_lock, flags);
1652 	/* update free */
1653 	hw_free = le64_to_cpu(*sc->hw_free);		/* volatile read */
1654 	old_free = sc->free;
1655 	extra = (((hw_free & CR_COUNTER_SMASK) >> CR_COUNTER_SHIFT)
1656 			- (old_free & CR_COUNTER_MASK))
1657 				& CR_COUNTER_MASK;
1658 	free = old_free + extra;
1659 	trace_hfi1_piofree(sc, extra);
1660 
1661 	/* call sent buffer callbacks */
1662 	code = -1;				/* code not yet set */
1663 	head = ACCESS_ONCE(sc->sr_head);	/* snapshot the head */
1664 	tail = sc->sr_tail;
1665 	while (head != tail) {
1666 		pbuf = &sc->sr[tail].pbuf;
1667 
1668 		if (sent_before(free, pbuf->sent_at)) {
1669 			/* not sent yet */
1670 			break;
1671 		}
1672 		if (pbuf->cb) {
1673 			if (code < 0) /* fill in code on first user */
1674 				code = fill_code(hw_free);
1675 			(*pbuf->cb)(pbuf->arg, code);
1676 		}
1677 
1678 		tail++;
1679 		if (tail >= sc->sr_size)
1680 			tail = 0;
1681 	}
1682 	sc->sr_tail = tail;
1683 	/* make sure tail is updated before free */
1684 	smp_wmb();
1685 	sc->free = free;
1686 	spin_unlock_irqrestore(&sc->release_lock, flags);
1687 	sc_piobufavail(sc);
1688 }
1689 
1690 /*
1691  * Send context group releaser.  Argument is the send context that caused
1692  * the interrupt.  Called from the send context interrupt handler.
1693  *
1694  * Call release on all contexts in the group.
1695  *
1696  * This routine takes the sc_lock without an irqsave because it is only
1697  * called from an interrupt handler.  Adjust if that changes.
1698  */
1699 void sc_group_release_update(struct hfi1_devdata *dd, u32 hw_context)
1700 {
1701 	struct send_context *sc;
1702 	u32 sw_index;
1703 	u32 gc, gc_end;
1704 
1705 	spin_lock(&dd->sc_lock);
1706 	sw_index = dd->hw_to_sw[hw_context];
1707 	if (unlikely(sw_index >= dd->num_send_contexts)) {
1708 		dd_dev_err(dd, "%s: invalid hw (%u) to sw (%u) mapping\n",
1709 			   __func__, hw_context, sw_index);
1710 		goto done;
1711 	}
1712 	sc = dd->send_contexts[sw_index].sc;
1713 	if (unlikely(!sc))
1714 		goto done;
1715 
1716 	gc = group_context(hw_context, sc->group);
1717 	gc_end = gc + group_size(sc->group);
1718 	for (; gc < gc_end; gc++) {
1719 		sw_index = dd->hw_to_sw[gc];
1720 		if (unlikely(sw_index >= dd->num_send_contexts)) {
1721 			dd_dev_err(dd,
1722 				   "%s: invalid hw (%u) to sw (%u) mapping\n",
1723 				   __func__, hw_context, sw_index);
1724 			continue;
1725 		}
1726 		sc_release_update(dd->send_contexts[sw_index].sc);
1727 	}
1728 done:
1729 	spin_unlock(&dd->sc_lock);
1730 }
1731 
1732 /*
1733  * pio_select_send_context_vl() - select send context
1734  * @dd: devdata
1735  * @selector: a spreading factor
1736  * @vl: this vl
1737  *
1738  * This function returns a send context based on the selector and a vl.
1739  * The mapping fields are protected by RCU
1740  */
1741 struct send_context *pio_select_send_context_vl(struct hfi1_devdata *dd,
1742 						u32 selector, u8 vl)
1743 {
1744 	struct pio_vl_map *m;
1745 	struct pio_map_elem *e;
1746 	struct send_context *rval;
1747 
1748 	/*
1749 	 * NOTE This should only happen if SC->VL changed after the initial
1750 	 * checks on the QP/AH
1751 	 * Default will return VL0's send context below
1752 	 */
1753 	if (unlikely(vl >= num_vls)) {
1754 		rval = NULL;
1755 		goto done;
1756 	}
1757 
1758 	rcu_read_lock();
1759 	m = rcu_dereference(dd->pio_map);
1760 	if (unlikely(!m)) {
1761 		rcu_read_unlock();
1762 		return dd->vld[0].sc;
1763 	}
1764 	e = m->map[vl & m->mask];
1765 	rval = e->ksc[selector & e->mask];
1766 	rcu_read_unlock();
1767 
1768 done:
1769 	rval = !rval ? dd->vld[0].sc : rval;
1770 	return rval;
1771 }
1772 
1773 /*
1774  * pio_select_send_context_sc() - select send context
1775  * @dd: devdata
1776  * @selector: a spreading factor
1777  * @sc5: the 5 bit sc
1778  *
1779  * This function returns an send context based on the selector and an sc
1780  */
1781 struct send_context *pio_select_send_context_sc(struct hfi1_devdata *dd,
1782 						u32 selector, u8 sc5)
1783 {
1784 	u8 vl = sc_to_vlt(dd, sc5);
1785 
1786 	return pio_select_send_context_vl(dd, selector, vl);
1787 }
1788 
1789 /*
1790  * Free the indicated map struct
1791  */
1792 static void pio_map_free(struct pio_vl_map *m)
1793 {
1794 	int i;
1795 
1796 	for (i = 0; m && i < m->actual_vls; i++)
1797 		kfree(m->map[i]);
1798 	kfree(m);
1799 }
1800 
1801 /*
1802  * Handle RCU callback
1803  */
1804 static void pio_map_rcu_callback(struct rcu_head *list)
1805 {
1806 	struct pio_vl_map *m = container_of(list, struct pio_vl_map, list);
1807 
1808 	pio_map_free(m);
1809 }
1810 
1811 /*
1812  * Set credit return threshold for the kernel send context
1813  */
1814 static void set_threshold(struct hfi1_devdata *dd, int scontext, int i)
1815 {
1816 	u32 thres;
1817 
1818 	thres = min(sc_percent_to_threshold(dd->kernel_send_context[scontext],
1819 					    50),
1820 		    sc_mtu_to_threshold(dd->kernel_send_context[scontext],
1821 					dd->vld[i].mtu,
1822 					dd->rcd[0]->rcvhdrqentsize));
1823 	sc_set_cr_threshold(dd->kernel_send_context[scontext], thres);
1824 }
1825 
1826 /*
1827  * pio_map_init - called when #vls change
1828  * @dd: hfi1_devdata
1829  * @port: port number
1830  * @num_vls: number of vls
1831  * @vl_scontexts: per vl send context mapping (optional)
1832  *
1833  * This routine changes the mapping based on the number of vls.
1834  *
1835  * vl_scontexts is used to specify a non-uniform vl/send context
1836  * loading. NULL implies auto computing the loading and giving each
1837  * VL an uniform distribution of send contexts per VL.
1838  *
1839  * The auto algorithm computers the sc_per_vl and the number of extra
1840  * send contexts. Any extra send contexts are added from the last VL
1841  * on down
1842  *
1843  * rcu locking is used here to control access to the mapping fields.
1844  *
1845  * If either the num_vls or num_send_contexts are non-power of 2, the
1846  * array sizes in the struct pio_vl_map and the struct pio_map_elem are
1847  * rounded up to the next highest power of 2 and the first entry is
1848  * reused in a round robin fashion.
1849  *
1850  * If an error occurs the map change is not done and the mapping is not
1851  * chaged.
1852  *
1853  */
1854 int pio_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_scontexts)
1855 {
1856 	int i, j;
1857 	int extra, sc_per_vl;
1858 	int scontext = 1;
1859 	int num_kernel_send_contexts = 0;
1860 	u8 lvl_scontexts[OPA_MAX_VLS];
1861 	struct pio_vl_map *oldmap, *newmap;
1862 
1863 	if (!vl_scontexts) {
1864 		for (i = 0; i < dd->num_send_contexts; i++)
1865 			if (dd->send_contexts[i].type == SC_KERNEL)
1866 				num_kernel_send_contexts++;
1867 		/* truncate divide */
1868 		sc_per_vl = num_kernel_send_contexts / num_vls;
1869 		/* extras */
1870 		extra = num_kernel_send_contexts % num_vls;
1871 		vl_scontexts = lvl_scontexts;
1872 		/* add extras from last vl down */
1873 		for (i = num_vls - 1; i >= 0; i--, extra--)
1874 			vl_scontexts[i] = sc_per_vl + (extra > 0 ? 1 : 0);
1875 	}
1876 	/* build new map */
1877 	newmap = kzalloc(sizeof(*newmap) +
1878 			 roundup_pow_of_two(num_vls) *
1879 			 sizeof(struct pio_map_elem *),
1880 			 GFP_KERNEL);
1881 	if (!newmap)
1882 		goto bail;
1883 	newmap->actual_vls = num_vls;
1884 	newmap->vls = roundup_pow_of_two(num_vls);
1885 	newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1886 	for (i = 0; i < newmap->vls; i++) {
1887 		/* save for wrap around */
1888 		int first_scontext = scontext;
1889 
1890 		if (i < newmap->actual_vls) {
1891 			int sz = roundup_pow_of_two(vl_scontexts[i]);
1892 
1893 			/* only allocate once */
1894 			newmap->map[i] = kzalloc(sizeof(*newmap->map[i]) +
1895 						 sz * sizeof(struct
1896 							     send_context *),
1897 						 GFP_KERNEL);
1898 			if (!newmap->map[i])
1899 				goto bail;
1900 			newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1901 			/*
1902 			 * assign send contexts and
1903 			 * adjust credit return threshold
1904 			 */
1905 			for (j = 0; j < sz; j++) {
1906 				if (dd->kernel_send_context[scontext]) {
1907 					newmap->map[i]->ksc[j] =
1908 					dd->kernel_send_context[scontext];
1909 					set_threshold(dd, scontext, i);
1910 				}
1911 				if (++scontext >= first_scontext +
1912 						  vl_scontexts[i])
1913 					/* wrap back to first send context */
1914 					scontext = first_scontext;
1915 			}
1916 		} else {
1917 			/* just re-use entry without allocating */
1918 			newmap->map[i] = newmap->map[i % num_vls];
1919 		}
1920 		scontext = first_scontext + vl_scontexts[i];
1921 	}
1922 	/* newmap in hand, save old map */
1923 	spin_lock_irq(&dd->pio_map_lock);
1924 	oldmap = rcu_dereference_protected(dd->pio_map,
1925 					   lockdep_is_held(&dd->pio_map_lock));
1926 
1927 	/* publish newmap */
1928 	rcu_assign_pointer(dd->pio_map, newmap);
1929 
1930 	spin_unlock_irq(&dd->pio_map_lock);
1931 	/* success, free any old map after grace period */
1932 	if (oldmap)
1933 		call_rcu(&oldmap->list, pio_map_rcu_callback);
1934 	return 0;
1935 bail:
1936 	/* free any partial allocation */
1937 	pio_map_free(newmap);
1938 	return -ENOMEM;
1939 }
1940 
1941 void free_pio_map(struct hfi1_devdata *dd)
1942 {
1943 	/* Free PIO map if allocated */
1944 	if (rcu_access_pointer(dd->pio_map)) {
1945 		spin_lock_irq(&dd->pio_map_lock);
1946 		pio_map_free(rcu_access_pointer(dd->pio_map));
1947 		RCU_INIT_POINTER(dd->pio_map, NULL);
1948 		spin_unlock_irq(&dd->pio_map_lock);
1949 		synchronize_rcu();
1950 	}
1951 	kfree(dd->kernel_send_context);
1952 	dd->kernel_send_context = NULL;
1953 }
1954 
1955 int init_pervl_scs(struct hfi1_devdata *dd)
1956 {
1957 	int i;
1958 	u64 mask, all_vl_mask = (u64)0x80ff; /* VLs 0-7, 15 */
1959 	u64 data_vls_mask = (u64)0x00ff; /* VLs 0-7 */
1960 	u32 ctxt;
1961 	struct hfi1_pportdata *ppd = dd->pport;
1962 
1963 	dd->vld[15].sc = sc_alloc(dd, SC_VL15,
1964 				  dd->rcd[0]->rcvhdrqentsize, dd->node);
1965 	if (!dd->vld[15].sc)
1966 		return -ENOMEM;
1967 
1968 	hfi1_init_ctxt(dd->vld[15].sc);
1969 	dd->vld[15].mtu = enum_to_mtu(OPA_MTU_2048);
1970 
1971 	dd->kernel_send_context = kzalloc_node(dd->num_send_contexts *
1972 					sizeof(struct send_context *),
1973 					GFP_KERNEL, dd->node);
1974 	if (!dd->kernel_send_context)
1975 		goto freesc15;
1976 
1977 	dd->kernel_send_context[0] = dd->vld[15].sc;
1978 
1979 	for (i = 0; i < num_vls; i++) {
1980 		/*
1981 		 * Since this function does not deal with a specific
1982 		 * receive context but we need the RcvHdrQ entry size,
1983 		 * use the size from rcd[0]. It is guaranteed to be
1984 		 * valid at this point and will remain the same for all
1985 		 * receive contexts.
1986 		 */
1987 		dd->vld[i].sc = sc_alloc(dd, SC_KERNEL,
1988 					 dd->rcd[0]->rcvhdrqentsize, dd->node);
1989 		if (!dd->vld[i].sc)
1990 			goto nomem;
1991 		dd->kernel_send_context[i + 1] = dd->vld[i].sc;
1992 		hfi1_init_ctxt(dd->vld[i].sc);
1993 		/* non VL15 start with the max MTU */
1994 		dd->vld[i].mtu = hfi1_max_mtu;
1995 	}
1996 	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++) {
1997 		dd->kernel_send_context[i + 1] =
1998 		sc_alloc(dd, SC_KERNEL, dd->rcd[0]->rcvhdrqentsize, dd->node);
1999 		if (!dd->kernel_send_context[i + 1])
2000 			goto nomem;
2001 		hfi1_init_ctxt(dd->kernel_send_context[i + 1]);
2002 	}
2003 
2004 	sc_enable(dd->vld[15].sc);
2005 	ctxt = dd->vld[15].sc->hw_context;
2006 	mask = all_vl_mask & ~(1LL << 15);
2007 	write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2008 	dd_dev_info(dd,
2009 		    "Using send context %u(%u) for VL15\n",
2010 		    dd->vld[15].sc->sw_index, ctxt);
2011 
2012 	for (i = 0; i < num_vls; i++) {
2013 		sc_enable(dd->vld[i].sc);
2014 		ctxt = dd->vld[i].sc->hw_context;
2015 		mask = all_vl_mask & ~(data_vls_mask);
2016 		write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2017 	}
2018 	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++) {
2019 		sc_enable(dd->kernel_send_context[i + 1]);
2020 		ctxt = dd->kernel_send_context[i + 1]->hw_context;
2021 		mask = all_vl_mask & ~(data_vls_mask);
2022 		write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2023 	}
2024 
2025 	if (pio_map_init(dd, ppd->port - 1, num_vls, NULL))
2026 		goto nomem;
2027 	return 0;
2028 
2029 nomem:
2030 	for (i = 0; i < num_vls; i++) {
2031 		sc_free(dd->vld[i].sc);
2032 		dd->vld[i].sc = NULL;
2033 	}
2034 
2035 	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++)
2036 		sc_free(dd->kernel_send_context[i + 1]);
2037 
2038 	kfree(dd->kernel_send_context);
2039 	dd->kernel_send_context = NULL;
2040 
2041 freesc15:
2042 	sc_free(dd->vld[15].sc);
2043 	return -ENOMEM;
2044 }
2045 
2046 int init_credit_return(struct hfi1_devdata *dd)
2047 {
2048 	int ret;
2049 	int i;
2050 
2051 	dd->cr_base = kcalloc(
2052 		node_affinity.num_possible_nodes,
2053 		sizeof(struct credit_return_base),
2054 		GFP_KERNEL);
2055 	if (!dd->cr_base) {
2056 		ret = -ENOMEM;
2057 		goto done;
2058 	}
2059 	for_each_node_with_cpus(i) {
2060 		int bytes = TXE_NUM_CONTEXTS * sizeof(struct credit_return);
2061 
2062 		set_dev_node(&dd->pcidev->dev, i);
2063 		dd->cr_base[i].va = dma_zalloc_coherent(
2064 					&dd->pcidev->dev,
2065 					bytes,
2066 					&dd->cr_base[i].dma,
2067 					GFP_KERNEL);
2068 		if (!dd->cr_base[i].va) {
2069 			set_dev_node(&dd->pcidev->dev, dd->node);
2070 			dd_dev_err(dd,
2071 				   "Unable to allocate credit return DMA range for NUMA %d\n",
2072 				   i);
2073 			ret = -ENOMEM;
2074 			goto done;
2075 		}
2076 	}
2077 	set_dev_node(&dd->pcidev->dev, dd->node);
2078 
2079 	ret = 0;
2080 done:
2081 	return ret;
2082 }
2083 
2084 void free_credit_return(struct hfi1_devdata *dd)
2085 {
2086 	int i;
2087 
2088 	if (!dd->cr_base)
2089 		return;
2090 	for (i = 0; i < node_affinity.num_possible_nodes; i++) {
2091 		if (dd->cr_base[i].va) {
2092 			dma_free_coherent(&dd->pcidev->dev,
2093 					  TXE_NUM_CONTEXTS *
2094 					  sizeof(struct credit_return),
2095 					  dd->cr_base[i].va,
2096 					  dd->cr_base[i].dma);
2097 		}
2098 	}
2099 	kfree(dd->cr_base);
2100 	dd->cr_base = NULL;
2101 }
2102