xref: /linux/drivers/infiniband/hw/hfi1/pio.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
1 /*
2  * Copyright(c) 2015, 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/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 	dma_addr_t dma;
707 	unsigned long flags;
708 	u64 reg;
709 	u32 thresh;
710 	u32 sw_index;
711 	u32 hw_context;
712 	int ret;
713 	u8 opval, opmask;
714 
715 	/* do not allocate while frozen */
716 	if (dd->flags & HFI1_FROZEN)
717 		return NULL;
718 
719 	sc = kzalloc_node(sizeof(*sc), GFP_KERNEL, numa);
720 	if (!sc)
721 		return NULL;
722 
723 	sc->buffers_allocated = alloc_percpu(u32);
724 	if (!sc->buffers_allocated) {
725 		kfree(sc);
726 		dd_dev_err(dd,
727 			   "Cannot allocate buffers_allocated per cpu counters\n"
728 			  );
729 		return NULL;
730 	}
731 
732 	spin_lock_irqsave(&dd->sc_lock, flags);
733 	ret = sc_hw_alloc(dd, type, &sw_index, &hw_context);
734 	if (ret) {
735 		spin_unlock_irqrestore(&dd->sc_lock, flags);
736 		free_percpu(sc->buffers_allocated);
737 		kfree(sc);
738 		return NULL;
739 	}
740 
741 	sci = &dd->send_contexts[sw_index];
742 	sci->sc = sc;
743 
744 	sc->dd = dd;
745 	sc->node = numa;
746 	sc->type = type;
747 	spin_lock_init(&sc->alloc_lock);
748 	spin_lock_init(&sc->release_lock);
749 	spin_lock_init(&sc->credit_ctrl_lock);
750 	INIT_LIST_HEAD(&sc->piowait);
751 	INIT_WORK(&sc->halt_work, sc_halted);
752 	init_waitqueue_head(&sc->halt_wait);
753 
754 	/* grouping is always single context for now */
755 	sc->group = 0;
756 
757 	sc->sw_index = sw_index;
758 	sc->hw_context = hw_context;
759 	cr_group_addresses(sc, &dma);
760 	sc->credits = sci->credits;
761 	sc->size = sc->credits * PIO_BLOCK_SIZE;
762 
763 /* PIO Send Memory Address details */
764 #define PIO_ADDR_CONTEXT_MASK 0xfful
765 #define PIO_ADDR_CONTEXT_SHIFT 16
766 	sc->base_addr = dd->piobase + ((hw_context & PIO_ADDR_CONTEXT_MASK)
767 					<< PIO_ADDR_CONTEXT_SHIFT);
768 
769 	/* set base and credits */
770 	reg = ((sci->credits & SC(CTRL_CTXT_DEPTH_MASK))
771 					<< SC(CTRL_CTXT_DEPTH_SHIFT))
772 		| ((sci->base & SC(CTRL_CTXT_BASE_MASK))
773 					<< SC(CTRL_CTXT_BASE_SHIFT));
774 	write_kctxt_csr(dd, hw_context, SC(CTRL), reg);
775 
776 	set_pio_integrity(sc);
777 
778 	/* unmask all errors */
779 	write_kctxt_csr(dd, hw_context, SC(ERR_MASK), (u64)-1);
780 
781 	/* set the default partition key */
782 	write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY),
783 			(SC(CHECK_PARTITION_KEY_VALUE_MASK) &
784 			 DEFAULT_PKEY) <<
785 			SC(CHECK_PARTITION_KEY_VALUE_SHIFT));
786 
787 	/* per context type checks */
788 	if (type == SC_USER) {
789 		opval = USER_OPCODE_CHECK_VAL;
790 		opmask = USER_OPCODE_CHECK_MASK;
791 	} else {
792 		opval = OPCODE_CHECK_VAL_DISABLED;
793 		opmask = OPCODE_CHECK_MASK_DISABLED;
794 	}
795 
796 	/* set the send context check opcode mask and value */
797 	write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE),
798 			((u64)opmask << SC(CHECK_OPCODE_MASK_SHIFT)) |
799 			((u64)opval << SC(CHECK_OPCODE_VALUE_SHIFT)));
800 
801 	/* set up credit return */
802 	reg = dma & SC(CREDIT_RETURN_ADDR_ADDRESS_SMASK);
803 	write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), reg);
804 
805 	/*
806 	 * Calculate the initial credit return threshold.
807 	 *
808 	 * For Ack contexts, set a threshold for half the credits.
809 	 * For User contexts use the given percentage.  This has been
810 	 * sanitized on driver start-up.
811 	 * For Kernel contexts, use the default MTU plus a header
812 	 * or half the credits, whichever is smaller. This should
813 	 * work for both the 3-deep buffering allocation and the
814 	 * pooling allocation.
815 	 */
816 	if (type == SC_ACK) {
817 		thresh = sc_percent_to_threshold(sc, 50);
818 	} else if (type == SC_USER) {
819 		thresh = sc_percent_to_threshold(sc,
820 						 user_credit_return_threshold);
821 	} else { /* kernel */
822 		thresh = min(sc_percent_to_threshold(sc, 50),
823 			     sc_mtu_to_threshold(sc, hfi1_max_mtu,
824 						 hdrqentsize));
825 	}
826 	reg = thresh << SC(CREDIT_CTRL_THRESHOLD_SHIFT);
827 	/* add in early return */
828 	if (type == SC_USER && HFI1_CAP_IS_USET(EARLY_CREDIT_RETURN))
829 		reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK);
830 	else if (HFI1_CAP_IS_KSET(EARLY_CREDIT_RETURN)) /* kernel, ack */
831 		reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK);
832 
833 	/* set up write-through credit_ctrl */
834 	sc->credit_ctrl = reg;
835 	write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), reg);
836 
837 	/* User send contexts should not allow sending on VL15 */
838 	if (type == SC_USER) {
839 		reg = 1ULL << 15;
840 		write_kctxt_csr(dd, hw_context, SC(CHECK_VL), reg);
841 	}
842 
843 	spin_unlock_irqrestore(&dd->sc_lock, flags);
844 
845 	/*
846 	 * Allocate shadow ring to track outstanding PIO buffers _after_
847 	 * unlocking.  We don't know the size until the lock is held and
848 	 * we can't allocate while the lock is held.  No one is using
849 	 * the context yet, so allocate it now.
850 	 *
851 	 * User contexts do not get a shadow ring.
852 	 */
853 	if (type != SC_USER) {
854 		/*
855 		 * Size the shadow ring 1 larger than the number of credits
856 		 * so head == tail can mean empty.
857 		 */
858 		sc->sr_size = sci->credits + 1;
859 		sc->sr = kzalloc_node(sizeof(union pio_shadow_ring) *
860 				sc->sr_size, GFP_KERNEL, numa);
861 		if (!sc->sr) {
862 			sc_free(sc);
863 			return NULL;
864 		}
865 	}
866 
867 	hfi1_cdbg(PIO,
868 		  "Send context %u(%u) %s group %u credits %u credit_ctrl 0x%llx threshold %u\n",
869 		  sw_index,
870 		  hw_context,
871 		  sc_type_name(type),
872 		  sc->group,
873 		  sc->credits,
874 		  sc->credit_ctrl,
875 		  thresh);
876 
877 	return sc;
878 }
879 
880 /* free a per-NUMA send context structure */
881 void sc_free(struct send_context *sc)
882 {
883 	struct hfi1_devdata *dd;
884 	unsigned long flags;
885 	u32 sw_index;
886 	u32 hw_context;
887 
888 	if (!sc)
889 		return;
890 
891 	sc->flags |= SCF_IN_FREE;	/* ensure no restarts */
892 	dd = sc->dd;
893 	if (!list_empty(&sc->piowait))
894 		dd_dev_err(dd, "piowait list not empty!\n");
895 	sw_index = sc->sw_index;
896 	hw_context = sc->hw_context;
897 	sc_disable(sc);	/* make sure the HW is disabled */
898 	flush_work(&sc->halt_work);
899 
900 	spin_lock_irqsave(&dd->sc_lock, flags);
901 	dd->send_contexts[sw_index].sc = NULL;
902 
903 	/* clear/disable all registers set in sc_alloc */
904 	write_kctxt_csr(dd, hw_context, SC(CTRL), 0);
905 	write_kctxt_csr(dd, hw_context, SC(CHECK_ENABLE), 0);
906 	write_kctxt_csr(dd, hw_context, SC(ERR_MASK), 0);
907 	write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY), 0);
908 	write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE), 0);
909 	write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), 0);
910 	write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), 0);
911 
912 	/* release the index and context for re-use */
913 	sc_hw_free(dd, sw_index, hw_context);
914 	spin_unlock_irqrestore(&dd->sc_lock, flags);
915 
916 	kfree(sc->sr);
917 	free_percpu(sc->buffers_allocated);
918 	kfree(sc);
919 }
920 
921 /* disable the context */
922 void sc_disable(struct send_context *sc)
923 {
924 	u64 reg;
925 	unsigned long flags;
926 	struct pio_buf *pbuf;
927 
928 	if (!sc)
929 		return;
930 
931 	/* do all steps, even if already disabled */
932 	spin_lock_irqsave(&sc->alloc_lock, flags);
933 	reg = read_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL));
934 	reg &= ~SC(CTRL_CTXT_ENABLE_SMASK);
935 	sc->flags &= ~SCF_ENABLED;
936 	sc_wait_for_packet_egress(sc, 1);
937 	write_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL), reg);
938 	spin_unlock_irqrestore(&sc->alloc_lock, flags);
939 
940 	/*
941 	 * Flush any waiters.  Once the context is disabled,
942 	 * credit return interrupts are stopped (although there
943 	 * could be one in-process when the context is disabled).
944 	 * Wait one microsecond for any lingering interrupts, then
945 	 * proceed with the flush.
946 	 */
947 	udelay(1);
948 	spin_lock_irqsave(&sc->release_lock, flags);
949 	if (sc->sr) {	/* this context has a shadow ring */
950 		while (sc->sr_tail != sc->sr_head) {
951 			pbuf = &sc->sr[sc->sr_tail].pbuf;
952 			if (pbuf->cb)
953 				(*pbuf->cb)(pbuf->arg, PRC_SC_DISABLE);
954 			sc->sr_tail++;
955 			if (sc->sr_tail >= sc->sr_size)
956 				sc->sr_tail = 0;
957 		}
958 	}
959 	spin_unlock_irqrestore(&sc->release_lock, flags);
960 }
961 
962 /* return SendEgressCtxtStatus.PacketOccupancy */
963 #define packet_occupancy(r) \
964 	(((r) & SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SMASK)\
965 	>> SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SHIFT)
966 
967 /* is egress halted on the context? */
968 #define egress_halted(r) \
969 	((r) & SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_HALT_STATUS_SMASK)
970 
971 /* wait for packet egress, optionally pause for credit return  */
972 static void sc_wait_for_packet_egress(struct send_context *sc, int pause)
973 {
974 	struct hfi1_devdata *dd = sc->dd;
975 	u64 reg = 0;
976 	u64 reg_prev;
977 	u32 loop = 0;
978 
979 	while (1) {
980 		reg_prev = reg;
981 		reg = read_csr(dd, sc->hw_context * 8 +
982 			       SEND_EGRESS_CTXT_STATUS);
983 		/* done if egress is stopped */
984 		if (egress_halted(reg))
985 			break;
986 		reg = packet_occupancy(reg);
987 		if (reg == 0)
988 			break;
989 		/* counter is reset if occupancy count changes */
990 		if (reg != reg_prev)
991 			loop = 0;
992 		if (loop > 50000) {
993 			/* timed out - bounce the link */
994 			dd_dev_err(dd,
995 				   "%s: context %u(%u) timeout waiting for packets to egress, remaining count %u, bouncing link\n",
996 				   __func__, sc->sw_index,
997 				   sc->hw_context, (u32)reg);
998 			queue_work(dd->pport->hfi1_wq,
999 				   &dd->pport->link_bounce_work);
1000 			break;
1001 		}
1002 		loop++;
1003 		udelay(1);
1004 	}
1005 
1006 	if (pause)
1007 		/* Add additional delay to ensure chip returns all credits */
1008 		pause_for_credit_return(dd);
1009 }
1010 
1011 void sc_wait(struct hfi1_devdata *dd)
1012 {
1013 	int i;
1014 
1015 	for (i = 0; i < dd->num_send_contexts; i++) {
1016 		struct send_context *sc = dd->send_contexts[i].sc;
1017 
1018 		if (!sc)
1019 			continue;
1020 		sc_wait_for_packet_egress(sc, 0);
1021 	}
1022 }
1023 
1024 /*
1025  * Restart a context after it has been halted due to error.
1026  *
1027  * If the first step fails - wait for the halt to be asserted, return early.
1028  * Otherwise complain about timeouts but keep going.
1029  *
1030  * It is expected that allocations (enabled flag bit) have been shut off
1031  * already (only applies to kernel contexts).
1032  */
1033 int sc_restart(struct send_context *sc)
1034 {
1035 	struct hfi1_devdata *dd = sc->dd;
1036 	u64 reg;
1037 	u32 loop;
1038 	int count;
1039 
1040 	/* bounce off if not halted, or being free'd */
1041 	if (!(sc->flags & SCF_HALTED) || (sc->flags & SCF_IN_FREE))
1042 		return -EINVAL;
1043 
1044 	dd_dev_info(dd, "restarting send context %u(%u)\n", sc->sw_index,
1045 		    sc->hw_context);
1046 
1047 	/*
1048 	 * Step 1: Wait for the context to actually halt.
1049 	 *
1050 	 * The error interrupt is asynchronous to actually setting halt
1051 	 * on the context.
1052 	 */
1053 	loop = 0;
1054 	while (1) {
1055 		reg = read_kctxt_csr(dd, sc->hw_context, SC(STATUS));
1056 		if (reg & SC(STATUS_CTXT_HALTED_SMASK))
1057 			break;
1058 		if (loop > 100) {
1059 			dd_dev_err(dd, "%s: context %u(%u) not halting, skipping\n",
1060 				   __func__, sc->sw_index, sc->hw_context);
1061 			return -ETIME;
1062 		}
1063 		loop++;
1064 		udelay(1);
1065 	}
1066 
1067 	/*
1068 	 * Step 2: Ensure no users are still trying to write to PIO.
1069 	 *
1070 	 * For kernel contexts, we have already turned off buffer allocation.
1071 	 * Now wait for the buffer count to go to zero.
1072 	 *
1073 	 * For user contexts, the user handling code has cut off write access
1074 	 * to the context's PIO pages before calling this routine and will
1075 	 * restore write access after this routine returns.
1076 	 */
1077 	if (sc->type != SC_USER) {
1078 		/* kernel context */
1079 		loop = 0;
1080 		while (1) {
1081 			count = get_buffers_allocated(sc);
1082 			if (count == 0)
1083 				break;
1084 			if (loop > 100) {
1085 				dd_dev_err(dd,
1086 					   "%s: context %u(%u) timeout waiting for PIO buffers to zero, remaining %d\n",
1087 					   __func__, sc->sw_index,
1088 					   sc->hw_context, count);
1089 			}
1090 			loop++;
1091 			udelay(1);
1092 		}
1093 	}
1094 
1095 	/*
1096 	 * Step 3: Wait for all packets to egress.
1097 	 * This is done while disabling the send context
1098 	 *
1099 	 * Step 4: Disable the context
1100 	 *
1101 	 * This is a superset of the halt.  After the disable, the
1102 	 * errors can be cleared.
1103 	 */
1104 	sc_disable(sc);
1105 
1106 	/*
1107 	 * Step 5: Enable the context
1108 	 *
1109 	 * This enable will clear the halted flag and per-send context
1110 	 * error flags.
1111 	 */
1112 	return sc_enable(sc);
1113 }
1114 
1115 /*
1116  * PIO freeze processing.  To be called after the TXE block is fully frozen.
1117  * Go through all frozen send contexts and disable them.  The contexts are
1118  * already stopped by the freeze.
1119  */
1120 void pio_freeze(struct hfi1_devdata *dd)
1121 {
1122 	struct send_context *sc;
1123 	int i;
1124 
1125 	for (i = 0; i < dd->num_send_contexts; i++) {
1126 		sc = dd->send_contexts[i].sc;
1127 		/*
1128 		 * Don't disable unallocated, unfrozen, or user send contexts.
1129 		 * User send contexts will be disabled when the process
1130 		 * calls into the driver to reset its context.
1131 		 */
1132 		if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER)
1133 			continue;
1134 
1135 		/* only need to disable, the context is already stopped */
1136 		sc_disable(sc);
1137 	}
1138 }
1139 
1140 /*
1141  * Unfreeze PIO for kernel send contexts.  The precondition for calling this
1142  * is that all PIO send contexts have been disabled and the SPC freeze has
1143  * been cleared.  Now perform the last step and re-enable each kernel context.
1144  * User (PSM) processing will occur when PSM calls into the kernel to
1145  * acknowledge the freeze.
1146  */
1147 void pio_kernel_unfreeze(struct hfi1_devdata *dd)
1148 {
1149 	struct send_context *sc;
1150 	int i;
1151 
1152 	for (i = 0; i < dd->num_send_contexts; i++) {
1153 		sc = dd->send_contexts[i].sc;
1154 		if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER)
1155 			continue;
1156 
1157 		sc_enable(sc);	/* will clear the sc frozen flag */
1158 	}
1159 }
1160 
1161 /*
1162  * Wait for the SendPioInitCtxt.PioInitInProgress bit to clear.
1163  * Returns:
1164  *	-ETIMEDOUT - if we wait too long
1165  *	-EIO	   - if there was an error
1166  */
1167 static int pio_init_wait_progress(struct hfi1_devdata *dd)
1168 {
1169 	u64 reg;
1170 	int max, count = 0;
1171 
1172 	/* max is the longest possible HW init time / delay */
1173 	max = (dd->icode == ICODE_FPGA_EMULATION) ? 120 : 5;
1174 	while (1) {
1175 		reg = read_csr(dd, SEND_PIO_INIT_CTXT);
1176 		if (!(reg & SEND_PIO_INIT_CTXT_PIO_INIT_IN_PROGRESS_SMASK))
1177 			break;
1178 		if (count >= max)
1179 			return -ETIMEDOUT;
1180 		udelay(5);
1181 		count++;
1182 	}
1183 
1184 	return reg & SEND_PIO_INIT_CTXT_PIO_INIT_ERR_SMASK ? -EIO : 0;
1185 }
1186 
1187 /*
1188  * Reset all of the send contexts to their power-on state.  Used
1189  * only during manual init - no lock against sc_enable needed.
1190  */
1191 void pio_reset_all(struct hfi1_devdata *dd)
1192 {
1193 	int ret;
1194 
1195 	/* make sure the init engine is not busy */
1196 	ret = pio_init_wait_progress(dd);
1197 	/* ignore any timeout */
1198 	if (ret == -EIO) {
1199 		/* clear the error */
1200 		write_csr(dd, SEND_PIO_ERR_CLEAR,
1201 			  SEND_PIO_ERR_CLEAR_PIO_INIT_SM_IN_ERR_SMASK);
1202 	}
1203 
1204 	/* reset init all */
1205 	write_csr(dd, SEND_PIO_INIT_CTXT,
1206 		  SEND_PIO_INIT_CTXT_PIO_ALL_CTXT_INIT_SMASK);
1207 	udelay(2);
1208 	ret = pio_init_wait_progress(dd);
1209 	if (ret < 0) {
1210 		dd_dev_err(dd,
1211 			   "PIO send context init %s while initializing all PIO blocks\n",
1212 			   ret == -ETIMEDOUT ? "is stuck" : "had an error");
1213 	}
1214 }
1215 
1216 /* enable the context */
1217 int sc_enable(struct send_context *sc)
1218 {
1219 	u64 sc_ctrl, reg, pio;
1220 	struct hfi1_devdata *dd;
1221 	unsigned long flags;
1222 	int ret = 0;
1223 
1224 	if (!sc)
1225 		return -EINVAL;
1226 	dd = sc->dd;
1227 
1228 	/*
1229 	 * Obtain the allocator lock to guard against any allocation
1230 	 * attempts (which should not happen prior to context being
1231 	 * enabled). On the release/disable side we don't need to
1232 	 * worry about locking since the releaser will not do anything
1233 	 * if the context accounting values have not changed.
1234 	 */
1235 	spin_lock_irqsave(&sc->alloc_lock, flags);
1236 	sc_ctrl = read_kctxt_csr(dd, sc->hw_context, SC(CTRL));
1237 	if ((sc_ctrl & SC(CTRL_CTXT_ENABLE_SMASK)))
1238 		goto unlock; /* already enabled */
1239 
1240 	/* IMPORTANT: only clear free and fill if transitioning 0 -> 1 */
1241 
1242 	*sc->hw_free = 0;
1243 	sc->free = 0;
1244 	sc->alloc_free = 0;
1245 	sc->fill = 0;
1246 	sc->fill_wrap = 0;
1247 	sc->sr_head = 0;
1248 	sc->sr_tail = 0;
1249 	sc->flags = 0;
1250 	/* the alloc lock insures no fast path allocation */
1251 	reset_buffers_allocated(sc);
1252 
1253 	/*
1254 	 * Clear all per-context errors.  Some of these will be set when
1255 	 * we are re-enabling after a context halt.  Now that the context
1256 	 * is disabled, the halt will not clear until after the PIO init
1257 	 * engine runs below.
1258 	 */
1259 	reg = read_kctxt_csr(dd, sc->hw_context, SC(ERR_STATUS));
1260 	if (reg)
1261 		write_kctxt_csr(dd, sc->hw_context, SC(ERR_CLEAR), reg);
1262 
1263 	/*
1264 	 * The HW PIO initialization engine can handle only one init
1265 	 * request at a time. Serialize access to each device's engine.
1266 	 */
1267 	spin_lock(&dd->sc_init_lock);
1268 	/*
1269 	 * Since access to this code block is serialized and
1270 	 * each access waits for the initialization to complete
1271 	 * before releasing the lock, the PIO initialization engine
1272 	 * should not be in use, so we don't have to wait for the
1273 	 * InProgress bit to go down.
1274 	 */
1275 	pio = ((sc->hw_context & SEND_PIO_INIT_CTXT_PIO_CTXT_NUM_MASK) <<
1276 	       SEND_PIO_INIT_CTXT_PIO_CTXT_NUM_SHIFT) |
1277 		SEND_PIO_INIT_CTXT_PIO_SINGLE_CTXT_INIT_SMASK;
1278 	write_csr(dd, SEND_PIO_INIT_CTXT, pio);
1279 	/*
1280 	 * Wait until the engine is done.  Give the chip the required time
1281 	 * so, hopefully, we read the register just once.
1282 	 */
1283 	udelay(2);
1284 	ret = pio_init_wait_progress(dd);
1285 	spin_unlock(&dd->sc_init_lock);
1286 	if (ret) {
1287 		dd_dev_err(dd,
1288 			   "sctxt%u(%u): Context not enabled due to init failure %d\n",
1289 			   sc->sw_index, sc->hw_context, ret);
1290 		goto unlock;
1291 	}
1292 
1293 	/*
1294 	 * All is well. Enable the context.
1295 	 */
1296 	sc_ctrl |= SC(CTRL_CTXT_ENABLE_SMASK);
1297 	write_kctxt_csr(dd, sc->hw_context, SC(CTRL), sc_ctrl);
1298 	/*
1299 	 * Read SendCtxtCtrl to force the write out and prevent a timing
1300 	 * hazard where a PIO write may reach the context before the enable.
1301 	 */
1302 	read_kctxt_csr(dd, sc->hw_context, SC(CTRL));
1303 	sc->flags |= SCF_ENABLED;
1304 
1305 unlock:
1306 	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1307 
1308 	return ret;
1309 }
1310 
1311 /* force a credit return on the context */
1312 void sc_return_credits(struct send_context *sc)
1313 {
1314 	if (!sc)
1315 		return;
1316 
1317 	/* a 0->1 transition schedules a credit return */
1318 	write_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE),
1319 			SC(CREDIT_FORCE_FORCE_RETURN_SMASK));
1320 	/*
1321 	 * Ensure that the write is flushed and the credit return is
1322 	 * scheduled. We care more about the 0 -> 1 transition.
1323 	 */
1324 	read_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE));
1325 	/* set back to 0 for next time */
1326 	write_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE), 0);
1327 }
1328 
1329 /* allow all in-flight packets to drain on the context */
1330 void sc_flush(struct send_context *sc)
1331 {
1332 	if (!sc)
1333 		return;
1334 
1335 	sc_wait_for_packet_egress(sc, 1);
1336 }
1337 
1338 /* drop all packets on the context, no waiting until they are sent */
1339 void sc_drop(struct send_context *sc)
1340 {
1341 	if (!sc)
1342 		return;
1343 
1344 	dd_dev_info(sc->dd, "%s: context %u(%u) - not implemented\n",
1345 		    __func__, sc->sw_index, sc->hw_context);
1346 }
1347 
1348 /*
1349  * Start the software reaction to a context halt or SPC freeze:
1350  *	- mark the context as halted or frozen
1351  *	- stop buffer allocations
1352  *
1353  * Called from the error interrupt.  Other work is deferred until
1354  * out of the interrupt.
1355  */
1356 void sc_stop(struct send_context *sc, int flag)
1357 {
1358 	unsigned long flags;
1359 
1360 	/* mark the context */
1361 	sc->flags |= flag;
1362 
1363 	/* stop buffer allocations */
1364 	spin_lock_irqsave(&sc->alloc_lock, flags);
1365 	sc->flags &= ~SCF_ENABLED;
1366 	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1367 	wake_up(&sc->halt_wait);
1368 }
1369 
1370 #define BLOCK_DWORDS (PIO_BLOCK_SIZE / sizeof(u32))
1371 #define dwords_to_blocks(x) DIV_ROUND_UP(x, BLOCK_DWORDS)
1372 
1373 /*
1374  * The send context buffer "allocator".
1375  *
1376  * @sc: the PIO send context we are allocating from
1377  * @len: length of whole packet - including PBC - in dwords
1378  * @cb: optional callback to call when the buffer is finished sending
1379  * @arg: argument for cb
1380  *
1381  * Return a pointer to a PIO buffer if successful, NULL if not enough room.
1382  */
1383 struct pio_buf *sc_buffer_alloc(struct send_context *sc, u32 dw_len,
1384 				pio_release_cb cb, void *arg)
1385 {
1386 	struct pio_buf *pbuf = NULL;
1387 	unsigned long flags;
1388 	unsigned long avail;
1389 	unsigned long blocks = dwords_to_blocks(dw_len);
1390 	u32 fill_wrap;
1391 	int trycount = 0;
1392 	u32 head, next;
1393 
1394 	spin_lock_irqsave(&sc->alloc_lock, flags);
1395 	if (!(sc->flags & SCF_ENABLED)) {
1396 		spin_unlock_irqrestore(&sc->alloc_lock, flags);
1397 		goto done;
1398 	}
1399 
1400 retry:
1401 	avail = (unsigned long)sc->credits - (sc->fill - sc->alloc_free);
1402 	if (blocks > avail) {
1403 		/* not enough room */
1404 		if (unlikely(trycount))	{ /* already tried to get more room */
1405 			spin_unlock_irqrestore(&sc->alloc_lock, flags);
1406 			goto done;
1407 		}
1408 		/* copy from receiver cache line and recalculate */
1409 		sc->alloc_free = ACCESS_ONCE(sc->free);
1410 		avail =
1411 			(unsigned long)sc->credits -
1412 			(sc->fill - sc->alloc_free);
1413 		if (blocks > avail) {
1414 			/* still no room, actively update */
1415 			sc_release_update(sc);
1416 			sc->alloc_free = ACCESS_ONCE(sc->free);
1417 			trycount++;
1418 			goto retry;
1419 		}
1420 	}
1421 
1422 	/* there is enough room */
1423 
1424 	preempt_disable();
1425 	this_cpu_inc(*sc->buffers_allocated);
1426 
1427 	/* read this once */
1428 	head = sc->sr_head;
1429 
1430 	/* "allocate" the buffer */
1431 	sc->fill += blocks;
1432 	fill_wrap = sc->fill_wrap;
1433 	sc->fill_wrap += blocks;
1434 	if (sc->fill_wrap >= sc->credits)
1435 		sc->fill_wrap = sc->fill_wrap - sc->credits;
1436 
1437 	/*
1438 	 * Fill the parts that the releaser looks at before moving the head.
1439 	 * The only necessary piece is the sent_at field.  The credits
1440 	 * we have just allocated cannot have been returned yet, so the
1441 	 * cb and arg will not be looked at for a "while".  Put them
1442 	 * on this side of the memory barrier anyway.
1443 	 */
1444 	pbuf = &sc->sr[head].pbuf;
1445 	pbuf->sent_at = sc->fill;
1446 	pbuf->cb = cb;
1447 	pbuf->arg = arg;
1448 	pbuf->sc = sc;	/* could be filled in at sc->sr init time */
1449 	/* make sure this is in memory before updating the head */
1450 
1451 	/* calculate next head index, do not store */
1452 	next = head + 1;
1453 	if (next >= sc->sr_size)
1454 		next = 0;
1455 	/*
1456 	 * update the head - must be last! - the releaser can look at fields
1457 	 * in pbuf once we move the head
1458 	 */
1459 	smp_wmb();
1460 	sc->sr_head = next;
1461 	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1462 
1463 	/* finish filling in the buffer outside the lock */
1464 	pbuf->start = sc->base_addr + fill_wrap * PIO_BLOCK_SIZE;
1465 	pbuf->end = sc->base_addr + sc->size;
1466 	pbuf->qw_written = 0;
1467 	pbuf->carry_bytes = 0;
1468 	pbuf->carry.val64 = 0;
1469 done:
1470 	return pbuf;
1471 }
1472 
1473 /*
1474  * There are at least two entities that can turn on credit return
1475  * interrupts and they can overlap.  Avoid problems by implementing
1476  * a count scheme that is enforced by a lock.  The lock is needed because
1477  * the count and CSR write must be paired.
1478  */
1479 
1480 /*
1481  * Start credit return interrupts.  This is managed by a count.  If already
1482  * on, just increment the count.
1483  */
1484 void sc_add_credit_return_intr(struct send_context *sc)
1485 {
1486 	unsigned long flags;
1487 
1488 	/* lock must surround both the count change and the CSR update */
1489 	spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
1490 	if (sc->credit_intr_count == 0) {
1491 		sc->credit_ctrl |= SC(CREDIT_CTRL_CREDIT_INTR_SMASK);
1492 		write_kctxt_csr(sc->dd, sc->hw_context,
1493 				SC(CREDIT_CTRL), sc->credit_ctrl);
1494 	}
1495 	sc->credit_intr_count++;
1496 	spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
1497 }
1498 
1499 /*
1500  * Stop credit return interrupts.  This is managed by a count.  Decrement the
1501  * count, if the last user, then turn the credit interrupts off.
1502  */
1503 void sc_del_credit_return_intr(struct send_context *sc)
1504 {
1505 	unsigned long flags;
1506 
1507 	WARN_ON(sc->credit_intr_count == 0);
1508 
1509 	/* lock must surround both the count change and the CSR update */
1510 	spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
1511 	sc->credit_intr_count--;
1512 	if (sc->credit_intr_count == 0) {
1513 		sc->credit_ctrl &= ~SC(CREDIT_CTRL_CREDIT_INTR_SMASK);
1514 		write_kctxt_csr(sc->dd, sc->hw_context,
1515 				SC(CREDIT_CTRL), sc->credit_ctrl);
1516 	}
1517 	spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
1518 }
1519 
1520 /*
1521  * The caller must be careful when calling this.  All needint calls
1522  * must be paired with !needint.
1523  */
1524 void hfi1_sc_wantpiobuf_intr(struct send_context *sc, u32 needint)
1525 {
1526 	if (needint)
1527 		sc_add_credit_return_intr(sc);
1528 	else
1529 		sc_del_credit_return_intr(sc);
1530 	trace_hfi1_wantpiointr(sc, needint, sc->credit_ctrl);
1531 	if (needint) {
1532 		mmiowb();
1533 		sc_return_credits(sc);
1534 	}
1535 }
1536 
1537 /**
1538  * sc_piobufavail - callback when a PIO buffer is available
1539  * @sc: the send context
1540  *
1541  * This is called from the interrupt handler when a PIO buffer is
1542  * available after hfi1_verbs_send() returned an error that no buffers were
1543  * available. Disable the interrupt if there are no more QPs waiting.
1544  */
1545 static void sc_piobufavail(struct send_context *sc)
1546 {
1547 	struct hfi1_devdata *dd = sc->dd;
1548 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1549 	struct list_head *list;
1550 	struct rvt_qp *qps[PIO_WAIT_BATCH_SIZE];
1551 	struct rvt_qp *qp;
1552 	struct hfi1_qp_priv *priv;
1553 	unsigned long flags;
1554 	unsigned i, n = 0;
1555 
1556 	if (dd->send_contexts[sc->sw_index].type != SC_KERNEL &&
1557 	    dd->send_contexts[sc->sw_index].type != SC_VL15)
1558 		return;
1559 	list = &sc->piowait;
1560 	/*
1561 	 * Note: checking that the piowait list is empty and clearing
1562 	 * the buffer available interrupt needs to be atomic or we
1563 	 * could end up with QPs on the wait list with the interrupt
1564 	 * disabled.
1565 	 */
1566 	write_seqlock_irqsave(&dev->iowait_lock, flags);
1567 	while (!list_empty(list)) {
1568 		struct iowait *wait;
1569 
1570 		if (n == ARRAY_SIZE(qps))
1571 			break;
1572 		wait = list_first_entry(list, struct iowait, list);
1573 		qp = iowait_to_qp(wait);
1574 		priv = qp->priv;
1575 		list_del_init(&priv->s_iowait.list);
1576 		priv->s_iowait.lock = NULL;
1577 		/* refcount held until actual wake up */
1578 		qps[n++] = qp;
1579 	}
1580 	/*
1581 	 * If there had been waiters and there are more
1582 	 * insure that we redo the force to avoid a potential hang.
1583 	 */
1584 	if (n) {
1585 		hfi1_sc_wantpiobuf_intr(sc, 0);
1586 		if (!list_empty(list))
1587 			hfi1_sc_wantpiobuf_intr(sc, 1);
1588 	}
1589 	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
1590 
1591 	for (i = 0; i < n; i++)
1592 		hfi1_qp_wakeup(qps[i],
1593 			       RVT_S_WAIT_PIO | RVT_S_WAIT_PIO_DRAIN);
1594 }
1595 
1596 /* translate a send credit update to a bit code of reasons */
1597 static inline int fill_code(u64 hw_free)
1598 {
1599 	int code = 0;
1600 
1601 	if (hw_free & CR_STATUS_SMASK)
1602 		code |= PRC_STATUS_ERR;
1603 	if (hw_free & CR_CREDIT_RETURN_DUE_TO_PBC_SMASK)
1604 		code |= PRC_PBC;
1605 	if (hw_free & CR_CREDIT_RETURN_DUE_TO_THRESHOLD_SMASK)
1606 		code |= PRC_THRESHOLD;
1607 	if (hw_free & CR_CREDIT_RETURN_DUE_TO_ERR_SMASK)
1608 		code |= PRC_FILL_ERR;
1609 	if (hw_free & CR_CREDIT_RETURN_DUE_TO_FORCE_SMASK)
1610 		code |= PRC_SC_DISABLE;
1611 	return code;
1612 }
1613 
1614 /* use the jiffies compare to get the wrap right */
1615 #define sent_before(a, b) time_before(a, b)	/* a < b */
1616 
1617 /*
1618  * The send context buffer "releaser".
1619  */
1620 void sc_release_update(struct send_context *sc)
1621 {
1622 	struct pio_buf *pbuf;
1623 	u64 hw_free;
1624 	u32 head, tail;
1625 	unsigned long old_free;
1626 	unsigned long free;
1627 	unsigned long extra;
1628 	unsigned long flags;
1629 	int code;
1630 
1631 	if (!sc)
1632 		return;
1633 
1634 	spin_lock_irqsave(&sc->release_lock, flags);
1635 	/* update free */
1636 	hw_free = le64_to_cpu(*sc->hw_free);		/* volatile read */
1637 	old_free = sc->free;
1638 	extra = (((hw_free & CR_COUNTER_SMASK) >> CR_COUNTER_SHIFT)
1639 			- (old_free & CR_COUNTER_MASK))
1640 				& CR_COUNTER_MASK;
1641 	free = old_free + extra;
1642 	trace_hfi1_piofree(sc, extra);
1643 
1644 	/* call sent buffer callbacks */
1645 	code = -1;				/* code not yet set */
1646 	head = ACCESS_ONCE(sc->sr_head);	/* snapshot the head */
1647 	tail = sc->sr_tail;
1648 	while (head != tail) {
1649 		pbuf = &sc->sr[tail].pbuf;
1650 
1651 		if (sent_before(free, pbuf->sent_at)) {
1652 			/* not sent yet */
1653 			break;
1654 		}
1655 		if (pbuf->cb) {
1656 			if (code < 0) /* fill in code on first user */
1657 				code = fill_code(hw_free);
1658 			(*pbuf->cb)(pbuf->arg, code);
1659 		}
1660 
1661 		tail++;
1662 		if (tail >= sc->sr_size)
1663 			tail = 0;
1664 	}
1665 	sc->sr_tail = tail;
1666 	/* make sure tail is updated before free */
1667 	smp_wmb();
1668 	sc->free = free;
1669 	spin_unlock_irqrestore(&sc->release_lock, flags);
1670 	sc_piobufavail(sc);
1671 }
1672 
1673 /*
1674  * Send context group releaser.  Argument is the send context that caused
1675  * the interrupt.  Called from the send context interrupt handler.
1676  *
1677  * Call release on all contexts in the group.
1678  *
1679  * This routine takes the sc_lock without an irqsave because it is only
1680  * called from an interrupt handler.  Adjust if that changes.
1681  */
1682 void sc_group_release_update(struct hfi1_devdata *dd, u32 hw_context)
1683 {
1684 	struct send_context *sc;
1685 	u32 sw_index;
1686 	u32 gc, gc_end;
1687 
1688 	spin_lock(&dd->sc_lock);
1689 	sw_index = dd->hw_to_sw[hw_context];
1690 	if (unlikely(sw_index >= dd->num_send_contexts)) {
1691 		dd_dev_err(dd, "%s: invalid hw (%u) to sw (%u) mapping\n",
1692 			   __func__, hw_context, sw_index);
1693 		goto done;
1694 	}
1695 	sc = dd->send_contexts[sw_index].sc;
1696 	if (unlikely(!sc))
1697 		goto done;
1698 
1699 	gc = group_context(hw_context, sc->group);
1700 	gc_end = gc + group_size(sc->group);
1701 	for (; gc < gc_end; gc++) {
1702 		sw_index = dd->hw_to_sw[gc];
1703 		if (unlikely(sw_index >= dd->num_send_contexts)) {
1704 			dd_dev_err(dd,
1705 				   "%s: invalid hw (%u) to sw (%u) mapping\n",
1706 				   __func__, hw_context, sw_index);
1707 			continue;
1708 		}
1709 		sc_release_update(dd->send_contexts[sw_index].sc);
1710 	}
1711 done:
1712 	spin_unlock(&dd->sc_lock);
1713 }
1714 
1715 /*
1716  * pio_select_send_context_vl() - select send context
1717  * @dd: devdata
1718  * @selector: a spreading factor
1719  * @vl: this vl
1720  *
1721  * This function returns a send context based on the selector and a vl.
1722  * The mapping fields are protected by RCU
1723  */
1724 struct send_context *pio_select_send_context_vl(struct hfi1_devdata *dd,
1725 						u32 selector, u8 vl)
1726 {
1727 	struct pio_vl_map *m;
1728 	struct pio_map_elem *e;
1729 	struct send_context *rval;
1730 
1731 	/*
1732 	 * NOTE This should only happen if SC->VL changed after the initial
1733 	 * checks on the QP/AH
1734 	 * Default will return VL0's send context below
1735 	 */
1736 	if (unlikely(vl >= num_vls)) {
1737 		rval = NULL;
1738 		goto done;
1739 	}
1740 
1741 	rcu_read_lock();
1742 	m = rcu_dereference(dd->pio_map);
1743 	if (unlikely(!m)) {
1744 		rcu_read_unlock();
1745 		return dd->vld[0].sc;
1746 	}
1747 	e = m->map[vl & m->mask];
1748 	rval = e->ksc[selector & e->mask];
1749 	rcu_read_unlock();
1750 
1751 done:
1752 	rval = !rval ? dd->vld[0].sc : rval;
1753 	return rval;
1754 }
1755 
1756 /*
1757  * pio_select_send_context_sc() - select send context
1758  * @dd: devdata
1759  * @selector: a spreading factor
1760  * @sc5: the 5 bit sc
1761  *
1762  * This function returns an send context based on the selector and an sc
1763  */
1764 struct send_context *pio_select_send_context_sc(struct hfi1_devdata *dd,
1765 						u32 selector, u8 sc5)
1766 {
1767 	u8 vl = sc_to_vlt(dd, sc5);
1768 
1769 	return pio_select_send_context_vl(dd, selector, vl);
1770 }
1771 
1772 /*
1773  * Free the indicated map struct
1774  */
1775 static void pio_map_free(struct pio_vl_map *m)
1776 {
1777 	int i;
1778 
1779 	for (i = 0; m && i < m->actual_vls; i++)
1780 		kfree(m->map[i]);
1781 	kfree(m);
1782 }
1783 
1784 /*
1785  * Handle RCU callback
1786  */
1787 static void pio_map_rcu_callback(struct rcu_head *list)
1788 {
1789 	struct pio_vl_map *m = container_of(list, struct pio_vl_map, list);
1790 
1791 	pio_map_free(m);
1792 }
1793 
1794 /*
1795  * Set credit return threshold for the kernel send context
1796  */
1797 static void set_threshold(struct hfi1_devdata *dd, int scontext, int i)
1798 {
1799 	u32 thres;
1800 
1801 	thres = min(sc_percent_to_threshold(dd->kernel_send_context[scontext],
1802 					    50),
1803 		    sc_mtu_to_threshold(dd->kernel_send_context[scontext],
1804 					dd->vld[i].mtu,
1805 					dd->rcd[0]->rcvhdrqentsize));
1806 	sc_set_cr_threshold(dd->kernel_send_context[scontext], thres);
1807 }
1808 
1809 /*
1810  * pio_map_init - called when #vls change
1811  * @dd: hfi1_devdata
1812  * @port: port number
1813  * @num_vls: number of vls
1814  * @vl_scontexts: per vl send context mapping (optional)
1815  *
1816  * This routine changes the mapping based on the number of vls.
1817  *
1818  * vl_scontexts is used to specify a non-uniform vl/send context
1819  * loading. NULL implies auto computing the loading and giving each
1820  * VL an uniform distribution of send contexts per VL.
1821  *
1822  * The auto algorithm computers the sc_per_vl and the number of extra
1823  * send contexts. Any extra send contexts are added from the last VL
1824  * on down
1825  *
1826  * rcu locking is used here to control access to the mapping fields.
1827  *
1828  * If either the num_vls or num_send_contexts are non-power of 2, the
1829  * array sizes in the struct pio_vl_map and the struct pio_map_elem are
1830  * rounded up to the next highest power of 2 and the first entry is
1831  * reused in a round robin fashion.
1832  *
1833  * If an error occurs the map change is not done and the mapping is not
1834  * chaged.
1835  *
1836  */
1837 int pio_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_scontexts)
1838 {
1839 	int i, j;
1840 	int extra, sc_per_vl;
1841 	int scontext = 1;
1842 	int num_kernel_send_contexts = 0;
1843 	u8 lvl_scontexts[OPA_MAX_VLS];
1844 	struct pio_vl_map *oldmap, *newmap;
1845 
1846 	if (!vl_scontexts) {
1847 		for (i = 0; i < dd->num_send_contexts; i++)
1848 			if (dd->send_contexts[i].type == SC_KERNEL)
1849 				num_kernel_send_contexts++;
1850 		/* truncate divide */
1851 		sc_per_vl = num_kernel_send_contexts / num_vls;
1852 		/* extras */
1853 		extra = num_kernel_send_contexts % num_vls;
1854 		vl_scontexts = lvl_scontexts;
1855 		/* add extras from last vl down */
1856 		for (i = num_vls - 1; i >= 0; i--, extra--)
1857 			vl_scontexts[i] = sc_per_vl + (extra > 0 ? 1 : 0);
1858 	}
1859 	/* build new map */
1860 	newmap = kzalloc(sizeof(*newmap) +
1861 			 roundup_pow_of_two(num_vls) *
1862 			 sizeof(struct pio_map_elem *),
1863 			 GFP_KERNEL);
1864 	if (!newmap)
1865 		goto bail;
1866 	newmap->actual_vls = num_vls;
1867 	newmap->vls = roundup_pow_of_two(num_vls);
1868 	newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1869 	for (i = 0; i < newmap->vls; i++) {
1870 		/* save for wrap around */
1871 		int first_scontext = scontext;
1872 
1873 		if (i < newmap->actual_vls) {
1874 			int sz = roundup_pow_of_two(vl_scontexts[i]);
1875 
1876 			/* only allocate once */
1877 			newmap->map[i] = kzalloc(sizeof(*newmap->map[i]) +
1878 						 sz * sizeof(struct
1879 							     send_context *),
1880 						 GFP_KERNEL);
1881 			if (!newmap->map[i])
1882 				goto bail;
1883 			newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1884 			/*
1885 			 * assign send contexts and
1886 			 * adjust credit return threshold
1887 			 */
1888 			for (j = 0; j < sz; j++) {
1889 				if (dd->kernel_send_context[scontext]) {
1890 					newmap->map[i]->ksc[j] =
1891 					dd->kernel_send_context[scontext];
1892 					set_threshold(dd, scontext, i);
1893 				}
1894 				if (++scontext >= first_scontext +
1895 						  vl_scontexts[i])
1896 					/* wrap back to first send context */
1897 					scontext = first_scontext;
1898 			}
1899 		} else {
1900 			/* just re-use entry without allocating */
1901 			newmap->map[i] = newmap->map[i % num_vls];
1902 		}
1903 		scontext = first_scontext + vl_scontexts[i];
1904 	}
1905 	/* newmap in hand, save old map */
1906 	spin_lock_irq(&dd->pio_map_lock);
1907 	oldmap = rcu_dereference_protected(dd->pio_map,
1908 					   lockdep_is_held(&dd->pio_map_lock));
1909 
1910 	/* publish newmap */
1911 	rcu_assign_pointer(dd->pio_map, newmap);
1912 
1913 	spin_unlock_irq(&dd->pio_map_lock);
1914 	/* success, free any old map after grace period */
1915 	if (oldmap)
1916 		call_rcu(&oldmap->list, pio_map_rcu_callback);
1917 	return 0;
1918 bail:
1919 	/* free any partial allocation */
1920 	pio_map_free(newmap);
1921 	return -ENOMEM;
1922 }
1923 
1924 void free_pio_map(struct hfi1_devdata *dd)
1925 {
1926 	/* Free PIO map if allocated */
1927 	if (rcu_access_pointer(dd->pio_map)) {
1928 		spin_lock_irq(&dd->pio_map_lock);
1929 		pio_map_free(rcu_access_pointer(dd->pio_map));
1930 		RCU_INIT_POINTER(dd->pio_map, NULL);
1931 		spin_unlock_irq(&dd->pio_map_lock);
1932 		synchronize_rcu();
1933 	}
1934 	kfree(dd->kernel_send_context);
1935 	dd->kernel_send_context = NULL;
1936 }
1937 
1938 int init_pervl_scs(struct hfi1_devdata *dd)
1939 {
1940 	int i;
1941 	u64 mask, all_vl_mask = (u64)0x80ff; /* VLs 0-7, 15 */
1942 	u64 data_vls_mask = (u64)0x00ff; /* VLs 0-7 */
1943 	u32 ctxt;
1944 	struct hfi1_pportdata *ppd = dd->pport;
1945 
1946 	dd->vld[15].sc = sc_alloc(dd, SC_VL15,
1947 				  dd->rcd[0]->rcvhdrqentsize, dd->node);
1948 	if (!dd->vld[15].sc)
1949 		return -ENOMEM;
1950 
1951 	hfi1_init_ctxt(dd->vld[15].sc);
1952 	dd->vld[15].mtu = enum_to_mtu(OPA_MTU_2048);
1953 
1954 	dd->kernel_send_context = kzalloc_node(dd->num_send_contexts *
1955 					sizeof(struct send_context *),
1956 					GFP_KERNEL, dd->node);
1957 	if (!dd->kernel_send_context)
1958 		goto freesc15;
1959 
1960 	dd->kernel_send_context[0] = dd->vld[15].sc;
1961 
1962 	for (i = 0; i < num_vls; i++) {
1963 		/*
1964 		 * Since this function does not deal with a specific
1965 		 * receive context but we need the RcvHdrQ entry size,
1966 		 * use the size from rcd[0]. It is guaranteed to be
1967 		 * valid at this point and will remain the same for all
1968 		 * receive contexts.
1969 		 */
1970 		dd->vld[i].sc = sc_alloc(dd, SC_KERNEL,
1971 					 dd->rcd[0]->rcvhdrqentsize, dd->node);
1972 		if (!dd->vld[i].sc)
1973 			goto nomem;
1974 		dd->kernel_send_context[i + 1] = dd->vld[i].sc;
1975 		hfi1_init_ctxt(dd->vld[i].sc);
1976 		/* non VL15 start with the max MTU */
1977 		dd->vld[i].mtu = hfi1_max_mtu;
1978 	}
1979 	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++) {
1980 		dd->kernel_send_context[i + 1] =
1981 		sc_alloc(dd, SC_KERNEL, dd->rcd[0]->rcvhdrqentsize, dd->node);
1982 		if (!dd->kernel_send_context[i + 1])
1983 			goto nomem;
1984 		hfi1_init_ctxt(dd->kernel_send_context[i + 1]);
1985 	}
1986 
1987 	sc_enable(dd->vld[15].sc);
1988 	ctxt = dd->vld[15].sc->hw_context;
1989 	mask = all_vl_mask & ~(1LL << 15);
1990 	write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
1991 	dd_dev_info(dd,
1992 		    "Using send context %u(%u) for VL15\n",
1993 		    dd->vld[15].sc->sw_index, ctxt);
1994 
1995 	for (i = 0; i < num_vls; i++) {
1996 		sc_enable(dd->vld[i].sc);
1997 		ctxt = dd->vld[i].sc->hw_context;
1998 		mask = all_vl_mask & ~(data_vls_mask);
1999 		write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2000 	}
2001 	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++) {
2002 		sc_enable(dd->kernel_send_context[i + 1]);
2003 		ctxt = dd->kernel_send_context[i + 1]->hw_context;
2004 		mask = all_vl_mask & ~(data_vls_mask);
2005 		write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2006 	}
2007 
2008 	if (pio_map_init(dd, ppd->port - 1, num_vls, NULL))
2009 		goto nomem;
2010 	return 0;
2011 
2012 nomem:
2013 	for (i = 0; i < num_vls; i++) {
2014 		sc_free(dd->vld[i].sc);
2015 		dd->vld[i].sc = NULL;
2016 	}
2017 
2018 	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++)
2019 		sc_free(dd->kernel_send_context[i + 1]);
2020 
2021 	kfree(dd->kernel_send_context);
2022 	dd->kernel_send_context = NULL;
2023 
2024 freesc15:
2025 	sc_free(dd->vld[15].sc);
2026 	return -ENOMEM;
2027 }
2028 
2029 int init_credit_return(struct hfi1_devdata *dd)
2030 {
2031 	int ret;
2032 	int i;
2033 
2034 	dd->cr_base = kcalloc(
2035 		node_affinity.num_possible_nodes,
2036 		sizeof(struct credit_return_base),
2037 		GFP_KERNEL);
2038 	if (!dd->cr_base) {
2039 		ret = -ENOMEM;
2040 		goto done;
2041 	}
2042 	for_each_node_with_cpus(i) {
2043 		int bytes = TXE_NUM_CONTEXTS * sizeof(struct credit_return);
2044 
2045 		set_dev_node(&dd->pcidev->dev, i);
2046 		dd->cr_base[i].va = dma_zalloc_coherent(
2047 					&dd->pcidev->dev,
2048 					bytes,
2049 					&dd->cr_base[i].dma,
2050 					GFP_KERNEL);
2051 		if (!dd->cr_base[i].va) {
2052 			set_dev_node(&dd->pcidev->dev, dd->node);
2053 			dd_dev_err(dd,
2054 				   "Unable to allocate credit return DMA range for NUMA %d\n",
2055 				   i);
2056 			ret = -ENOMEM;
2057 			goto done;
2058 		}
2059 	}
2060 	set_dev_node(&dd->pcidev->dev, dd->node);
2061 
2062 	ret = 0;
2063 done:
2064 	return ret;
2065 }
2066 
2067 void free_credit_return(struct hfi1_devdata *dd)
2068 {
2069 	int i;
2070 
2071 	if (!dd->cr_base)
2072 		return;
2073 	for (i = 0; i < node_affinity.num_possible_nodes; i++) {
2074 		if (dd->cr_base[i].va) {
2075 			dma_free_coherent(&dd->pcidev->dev,
2076 					  TXE_NUM_CONTEXTS *
2077 					  sizeof(struct credit_return),
2078 					  dd->cr_base[i].va,
2079 					  dd->cr_base[i].dma);
2080 		}
2081 	}
2082 	kfree(dd->cr_base);
2083 	dd->cr_base = NULL;
2084 }
2085