1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
   2/*
   3 * Copyright(c) 2015-2018 Intel Corporation.
   4 */
   5
   6#include <linux/delay.h>
   7#include "hfi.h"
   8#include "qp.h"
   9#include "trace.h"
  10
  11#define SC(name) SEND_CTXT_##name
  12/*
  13 * Send Context functions
  14 */
  15static void sc_wait_for_packet_egress(struct send_context *sc, int pause);
  16
  17/*
  18 * Set the CM reset bit and wait for it to clear.  Use the provided
  19 * sendctrl register.  This routine has no locking.
  20 */
  21void __cm_reset(struct hfi1_devdata *dd, u64 sendctrl)
  22{
  23	write_csr(dd, SEND_CTRL, sendctrl | SEND_CTRL_CM_RESET_SMASK);
  24	while (1) {
  25		udelay(1);
  26		sendctrl = read_csr(dd, SEND_CTRL);
  27		if ((sendctrl & SEND_CTRL_CM_RESET_SMASK) == 0)
  28			break;
  29	}
  30}
  31
  32/* global control of PIO send */
  33void pio_send_control(struct hfi1_devdata *dd, int op)
  34{
  35	u64 reg, mask;
  36	unsigned long flags;
  37	int write = 1;	/* write sendctrl back */
  38	int flush = 0;	/* re-read sendctrl to make sure it is flushed */
  39	int i;
  40
  41	spin_lock_irqsave(&dd->sendctrl_lock, flags);
  42
  43	reg = read_csr(dd, SEND_CTRL);
  44	switch (op) {
  45	case PSC_GLOBAL_ENABLE:
  46		reg |= SEND_CTRL_SEND_ENABLE_SMASK;
  47		fallthrough;
  48	case PSC_DATA_VL_ENABLE:
  49		mask = 0;
  50		for (i = 0; i < ARRAY_SIZE(dd->vld); i++)
  51			if (!dd->vld[i].mtu)
  52				mask |= BIT_ULL(i);
  53		/* Disallow sending on VLs not enabled */
  54		mask = (mask & SEND_CTRL_UNSUPPORTED_VL_MASK) <<
  55			SEND_CTRL_UNSUPPORTED_VL_SHIFT;
  56		reg = (reg & ~SEND_CTRL_UNSUPPORTED_VL_SMASK) | mask;
  57		break;
  58	case PSC_GLOBAL_DISABLE:
  59		reg &= ~SEND_CTRL_SEND_ENABLE_SMASK;
  60		break;
  61	case PSC_GLOBAL_VLARB_ENABLE:
  62		reg |= SEND_CTRL_VL_ARBITER_ENABLE_SMASK;
  63		break;
  64	case PSC_GLOBAL_VLARB_DISABLE:
  65		reg &= ~SEND_CTRL_VL_ARBITER_ENABLE_SMASK;
  66		break;
  67	case PSC_CM_RESET:
  68		__cm_reset(dd, reg);
  69		write = 0; /* CSR already written (and flushed) */
  70		break;
  71	case PSC_DATA_VL_DISABLE:
  72		reg |= SEND_CTRL_UNSUPPORTED_VL_SMASK;
  73		flush = 1;
  74		break;
  75	default:
  76		dd_dev_err(dd, "%s: invalid control %d\n", __func__, op);
  77		break;
  78	}
  79
  80	if (write) {
  81		write_csr(dd, SEND_CTRL, reg);
  82		if (flush)
  83			(void)read_csr(dd, SEND_CTRL); /* flush write */
  84	}
  85
  86	spin_unlock_irqrestore(&dd->sendctrl_lock, flags);
  87}
  88
  89/* number of send context memory pools */
  90#define NUM_SC_POOLS 2
  91
  92/* Send Context Size (SCS) wildcards */
  93#define SCS_POOL_0 -1
  94#define SCS_POOL_1 -2
  95
  96/* Send Context Count (SCC) wildcards */
  97#define SCC_PER_VL -1
  98#define SCC_PER_CPU  -2
  99#define SCC_PER_KRCVQ  -3
 100
 101/* Send Context Size (SCS) constants */
 102#define SCS_ACK_CREDITS  32
 103#define SCS_VL15_CREDITS 102	/* 3 pkts of 2048B data + 128B header */
 104
 105#define PIO_THRESHOLD_CEILING 4096
 106
 107#define PIO_WAIT_BATCH_SIZE 5
 108
 109/* default send context sizes */
 110static struct sc_config_sizes sc_config_sizes[SC_MAX] = {
 111	[SC_KERNEL] = { .size  = SCS_POOL_0,	/* even divide, pool 0 */
 112			.count = SCC_PER_VL },	/* one per NUMA */
 113	[SC_ACK]    = { .size  = SCS_ACK_CREDITS,
 114			.count = SCC_PER_KRCVQ },
 115	[SC_USER]   = { .size  = SCS_POOL_0,	/* even divide, pool 0 */
 116			.count = SCC_PER_CPU },	/* one per CPU */
 117	[SC_VL15]   = { .size  = SCS_VL15_CREDITS,
 118			.count = 1 },
 119
 120};
 121
 122/* send context memory pool configuration */
 123struct mem_pool_config {
 124	int centipercent;	/* % of memory, in 100ths of 1% */
 125	int absolute_blocks;	/* absolute block count */
 126};
 127
 128/* default memory pool configuration: 100% in pool 0 */
 129static struct mem_pool_config sc_mem_pool_config[NUM_SC_POOLS] = {
 130	/* centi%, abs blocks */
 131	{  10000,     -1 },		/* pool 0 */
 132	{      0,     -1 },		/* pool 1 */
 133};
 134
 135/* memory pool information, used when calculating final sizes */
 136struct mem_pool_info {
 137	int centipercent;	/*
 138				 * 100th of 1% of memory to use, -1 if blocks
 139				 * already set
 140				 */
 141	int count;		/* count of contexts in the pool */
 142	int blocks;		/* block size of the pool */
 143	int size;		/* context size, in blocks */
 144};
 145
 146/*
 147 * Convert a pool wildcard to a valid pool index.  The wildcards
 148 * start at -1 and increase negatively.  Map them as:
 149 *	-1 => 0
 150 *	-2 => 1
 151 *	etc.
 152 *
 153 * Return -1 on non-wildcard input, otherwise convert to a pool number.
 154 */
 155static int wildcard_to_pool(int wc)
 156{
 157	if (wc >= 0)
 158		return -1;	/* non-wildcard */
 159	return -wc - 1;
 160}
 161
 162static const char *sc_type_names[SC_MAX] = {
 163	"kernel",
 164	"ack",
 165	"user",
 166	"vl15"
 167};
 168
 169static const char *sc_type_name(int index)
 170{
 171	if (index < 0 || index >= SC_MAX)
 172		return "unknown";
 173	return sc_type_names[index];
 174}
 175
 176/*
 177 * Read the send context memory pool configuration and send context
 178 * size configuration.  Replace any wildcards and come up with final
 179 * counts and sizes for the send context types.
 180 */
 181int init_sc_pools_and_sizes(struct hfi1_devdata *dd)
 182{
 183	struct mem_pool_info mem_pool_info[NUM_SC_POOLS] = { { 0 } };
 184	int total_blocks = (chip_pio_mem_size(dd) / PIO_BLOCK_SIZE) - 1;
 185	int total_contexts = 0;
 186	int fixed_blocks;
 187	int pool_blocks;
 188	int used_blocks;
 189	int cp_total;		/* centipercent total */
 190	int ab_total;		/* absolute block total */
 191	int extra;
 192	int i;
 193
 194	/*
 195	 * When SDMA is enabled, kernel context pio packet size is capped by
 196	 * "piothreshold". Reduce pio buffer allocation for kernel context by
 197	 * setting it to a fixed size. The allocation allows 3-deep buffering
 198	 * of the largest pio packets plus up to 128 bytes header, sufficient
 199	 * to maintain verbs performance.
 200	 *
 201	 * When SDMA is disabled, keep the default pooling allocation.
 202	 */
 203	if (HFI1_CAP_IS_KSET(SDMA)) {
 204		u16 max_pkt_size = (piothreshold < PIO_THRESHOLD_CEILING) ?
 205					 piothreshold : PIO_THRESHOLD_CEILING;
 206		sc_config_sizes[SC_KERNEL].size =
 207			3 * (max_pkt_size + 128) / PIO_BLOCK_SIZE;
 208	}
 209
 210	/*
 211	 * Step 0:
 212	 *	- copy the centipercents/absolute sizes from the pool config
 213	 *	- sanity check these values
 214	 *	- add up centipercents, then later check for full value
 215	 *	- add up absolute blocks, then later check for over-commit
 216	 */
 217	cp_total = 0;
 218	ab_total = 0;
 219	for (i = 0; i < NUM_SC_POOLS; i++) {
 220		int cp = sc_mem_pool_config[i].centipercent;
 221		int ab = sc_mem_pool_config[i].absolute_blocks;
 222
 223		/*
 224		 * A negative value is "unused" or "invalid".  Both *can*
 225		 * be valid, but centipercent wins, so check that first
 226		 */
 227		if (cp >= 0) {			/* centipercent valid */
 228			cp_total += cp;
 229		} else if (ab >= 0) {		/* absolute blocks valid */
 230			ab_total += ab;
 231		} else {			/* neither valid */
 232			dd_dev_err(
 233				dd,
 234				"Send context memory pool %d: both the block count and centipercent are invalid\n",
 235				i);
 236			return -EINVAL;
 237		}
 238
 239		mem_pool_info[i].centipercent = cp;
 240		mem_pool_info[i].blocks = ab;
 241	}
 242
 243	/* do not use both % and absolute blocks for different pools */
 244	if (cp_total != 0 && ab_total != 0) {
 245		dd_dev_err(
 246			dd,
 247			"All send context memory pools must be described as either centipercent or blocks, no mixing between pools\n");
 248		return -EINVAL;
 249	}
 250
 251	/* if any percentages are present, they must add up to 100% x 100 */
 252	if (cp_total != 0 && cp_total != 10000) {
 253		dd_dev_err(
 254			dd,
 255			"Send context memory pool centipercent is %d, expecting 10000\n",
 256			cp_total);
 257		return -EINVAL;
 258	}
 259
 260	/* the absolute pool total cannot be more than the mem total */
 261	if (ab_total > total_blocks) {
 262		dd_dev_err(
 263			dd,
 264			"Send context memory pool absolute block count %d is larger than the memory size %d\n",
 265			ab_total, total_blocks);
 266		return -EINVAL;
 267	}
 268
 269	/*
 270	 * Step 2:
 271	 *	- copy from the context size config
 272	 *	- replace context type wildcard counts with real values
 273	 *	- add up non-memory pool block sizes
 274	 *	- add up memory pool user counts
 275	 */
 276	fixed_blocks = 0;
 277	for (i = 0; i < SC_MAX; i++) {
 278		int count = sc_config_sizes[i].count;
 279		int size = sc_config_sizes[i].size;
 280		int pool;
 281
 282		/*
 283		 * Sanity check count: Either a positive value or
 284		 * one of the expected wildcards is valid.  The positive
 285		 * value is checked later when we compare against total
 286		 * memory available.
 287		 */
 288		if (i == SC_ACK) {
 289			count = dd->n_krcv_queues;
 290		} else if (i == SC_KERNEL) {
 291			count = INIT_SC_PER_VL * num_vls;
 292		} else if (count == SCC_PER_CPU) {
 293			count = dd->num_rcv_contexts - dd->n_krcv_queues;
 294		} else if (count < 0) {
 295			dd_dev_err(
 296				dd,
 297				"%s send context invalid count wildcard %d\n",
 298				sc_type_name(i), count);
 299			return -EINVAL;
 300		}
 301		if (total_contexts + count > chip_send_contexts(dd))
 302			count = chip_send_contexts(dd) - total_contexts;
 303
 304		total_contexts += count;
 305
 306		/*
 307		 * Sanity check pool: The conversion will return a pool
 308		 * number or -1 if a fixed (non-negative) value.  The fixed
 309		 * value is checked later when we compare against
 310		 * total memory available.
 311		 */
 312		pool = wildcard_to_pool(size);
 313		if (pool == -1) {			/* non-wildcard */
 314			fixed_blocks += size * count;
 315		} else if (pool < NUM_SC_POOLS) {	/* valid wildcard */
 316			mem_pool_info[pool].count += count;
 317		} else {				/* invalid wildcard */
 318			dd_dev_err(
 319				dd,
 320				"%s send context invalid pool wildcard %d\n",
 321				sc_type_name(i), size);
 322			return -EINVAL;
 323		}
 324
 325		dd->sc_sizes[i].count = count;
 326		dd->sc_sizes[i].size = size;
 327	}
 328	if (fixed_blocks > total_blocks) {
 329		dd_dev_err(
 330			dd,
 331			"Send context fixed block count, %u, larger than total block count %u\n",
 332			fixed_blocks, total_blocks);
 333		return -EINVAL;
 334	}
 335
 336	/* step 3: calculate the blocks in the pools, and pool context sizes */
 337	pool_blocks = total_blocks - fixed_blocks;
 338	if (ab_total > pool_blocks) {
 339		dd_dev_err(
 340			dd,
 341			"Send context fixed pool sizes, %u, larger than pool block count %u\n",
 342			ab_total, pool_blocks);
 343		return -EINVAL;
 344	}
 345	/* subtract off the fixed pool blocks */
 346	pool_blocks -= ab_total;
 347
 348	for (i = 0; i < NUM_SC_POOLS; i++) {
 349		struct mem_pool_info *pi = &mem_pool_info[i];
 350
 351		/* % beats absolute blocks */
 352		if (pi->centipercent >= 0)
 353			pi->blocks = (pool_blocks * pi->centipercent) / 10000;
 354
 355		if (pi->blocks == 0 && pi->count != 0) {
 356			dd_dev_err(
 357				dd,
 358				"Send context memory pool %d has %u contexts, but no blocks\n",
 359				i, pi->count);
 360			return -EINVAL;
 361		}
 362		if (pi->count == 0) {
 363			/* warn about wasted blocks */
 364			if (pi->blocks != 0)
 365				dd_dev_err(
 366					dd,
 367					"Send context memory pool %d has %u blocks, but zero contexts\n",
 368					i, pi->blocks);
 369			pi->size = 0;
 370		} else {
 371			pi->size = pi->blocks / pi->count;
 372		}
 373	}
 374
 375	/* step 4: fill in the context type sizes from the pool sizes */
 376	used_blocks = 0;
 377	for (i = 0; i < SC_MAX; i++) {
 378		if (dd->sc_sizes[i].size < 0) {
 379			unsigned pool = wildcard_to_pool(dd->sc_sizes[i].size);
 380
 381			WARN_ON_ONCE(pool >= NUM_SC_POOLS);
 382			dd->sc_sizes[i].size = mem_pool_info[pool].size;
 383		}
 384		/* make sure we are not larger than what is allowed by the HW */
 385#define PIO_MAX_BLOCKS 1024
 386		if (dd->sc_sizes[i].size > PIO_MAX_BLOCKS)
 387			dd->sc_sizes[i].size = PIO_MAX_BLOCKS;
 388
 389		/* calculate our total usage */
 390		used_blocks += dd->sc_sizes[i].size * dd->sc_sizes[i].count;
 391	}
 392	extra = total_blocks - used_blocks;
 393	if (extra != 0)
 394		dd_dev_info(dd, "unused send context blocks: %d\n", extra);
 395
 396	return total_contexts;
 397}
 398
 399int init_send_contexts(struct hfi1_devdata *dd)
 400{
 401	u16 base;
 402	int ret, i, j, context;
 403
 404	ret = init_credit_return(dd);
 405	if (ret)
 406		return ret;
 407
 408	dd->hw_to_sw = kmalloc_array(TXE_NUM_CONTEXTS, sizeof(u8),
 409					GFP_KERNEL);
 410	dd->send_contexts = kcalloc(dd->num_send_contexts,
 411				    sizeof(struct send_context_info),
 412				    GFP_KERNEL);
 413	if (!dd->send_contexts || !dd->hw_to_sw) {
 414		kfree(dd->hw_to_sw);
 415		kfree(dd->send_contexts);
 416		free_credit_return(dd);
 417		return -ENOMEM;
 418	}
 419
 420	/* hardware context map starts with invalid send context indices */
 421	for (i = 0; i < TXE_NUM_CONTEXTS; i++)
 422		dd->hw_to_sw[i] = INVALID_SCI;
 423
 424	/*
 425	 * All send contexts have their credit sizes.  Allocate credits
 426	 * for each context one after another from the global space.
 427	 */
 428	context = 0;
 429	base = 1;
 430	for (i = 0; i < SC_MAX; i++) {
 431		struct sc_config_sizes *scs = &dd->sc_sizes[i];
 432
 433		for (j = 0; j < scs->count; j++) {
 434			struct send_context_info *sci =
 435						&dd->send_contexts[context];
 436			sci->type = i;
 437			sci->base = base;
 438			sci->credits = scs->size;
 439
 440			context++;
 441			base += scs->size;
 442		}
 443	}
 444
 445	return 0;
 446}
 447
 448/*
 449 * Allocate a software index and hardware context of the given type.
 450 *
 451 * Must be called with dd->sc_lock held.
 452 */
 453static int sc_hw_alloc(struct hfi1_devdata *dd, int type, u32 *sw_index,
 454		       u32 *hw_context)
 455{
 456	struct send_context_info *sci;
 457	u32 index;
 458	u32 context;
 459
 460	for (index = 0, sci = &dd->send_contexts[0];
 461			index < dd->num_send_contexts; index++, sci++) {
 462		if (sci->type == type && sci->allocated == 0) {
 463			sci->allocated = 1;
 464			/* use a 1:1 mapping, but make them non-equal */
 465			context = chip_send_contexts(dd) - index - 1;
 466			dd->hw_to_sw[context] = index;
 467			*sw_index = index;
 468			*hw_context = context;
 469			return 0; /* success */
 470		}
 471	}
 472	dd_dev_err(dd, "Unable to locate a free type %d send context\n", type);
 473	return -ENOSPC;
 474}
 475
 476/*
 477 * Free the send context given by its software index.
 478 *
 479 * Must be called with dd->sc_lock held.
 480 */
 481static void sc_hw_free(struct hfi1_devdata *dd, u32 sw_index, u32 hw_context)
 482{
 483	struct send_context_info *sci;
 484
 485	sci = &dd->send_contexts[sw_index];
 486	if (!sci->allocated) {
 487		dd_dev_err(dd, "%s: sw_index %u not allocated? hw_context %u\n",
 488			   __func__, sw_index, hw_context);
 489	}
 490	sci->allocated = 0;
 491	dd->hw_to_sw[hw_context] = INVALID_SCI;
 492}
 493
 494/* return the base context of a context in a group */
 495static inline u32 group_context(u32 context, u32 group)
 496{
 497	return (context >> group) << group;
 498}
 499
 500/* return the size of a group */
 501static inline u32 group_size(u32 group)
 502{
 503	return 1 << group;
 504}
 505
 506/*
 507 * Obtain the credit return addresses, kernel virtual and bus, for the
 508 * given sc.
 509 *
 510 * To understand this routine:
 511 * o va and dma are arrays of struct credit_return.  One for each physical
 512 *   send context, per NUMA.
 513 * o Each send context always looks in its relative location in a struct
 514 *   credit_return for its credit return.
 515 * o Each send context in a group must have its return address CSR programmed
 516 *   with the same value.  Use the address of the first send context in the
 517 *   group.
 518 */
 519static void cr_group_addresses(struct send_context *sc, dma_addr_t *dma)
 520{
 521	u32 gc = group_context(sc->hw_context, sc->group);
 522	u32 index = sc->hw_context & 0x7;
 523
 524	sc->hw_free = &sc->dd->cr_base[sc->node].va[gc].cr[index];
 525	*dma = (unsigned long)
 526	       &((struct credit_return *)sc->dd->cr_base[sc->node].dma)[gc];
 527}
 528
 529/*
 530 * Work queue function triggered in error interrupt routine for
 531 * kernel contexts.
 532 */
 533static void sc_halted(struct work_struct *work)
 534{
 535	struct send_context *sc;
 536
 537	sc = container_of(work, struct send_context, halt_work);
 538	sc_restart(sc);
 539}
 540
 541/*
 542 * Calculate PIO block threshold for this send context using the given MTU.
 543 * Trigger a return when one MTU plus optional header of credits remain.
 544 *
 545 * Parameter mtu is in bytes.
 546 * Parameter hdrqentsize is in DWORDs.
 547 *
 548 * Return value is what to write into the CSR: trigger return when
 549 * unreturned credits pass this count.
 550 */
 551u32 sc_mtu_to_threshold(struct send_context *sc, u32 mtu, u32 hdrqentsize)
 552{
 553	u32 release_credits;
 554	u32 threshold;
 555
 556	/* add in the header size, then divide by the PIO block size */
 557	mtu += hdrqentsize << 2;
 558	release_credits = DIV_ROUND_UP(mtu, PIO_BLOCK_SIZE);
 559
 560	/* check against this context's credits */
 561	if (sc->credits <= release_credits)
 562		threshold = 1;
 563	else
 564		threshold = sc->credits - release_credits;
 565
 566	return threshold;
 567}
 568
 569/*
 570 * Calculate credit threshold in terms of percent of the allocated credits.
 571 * Trigger when unreturned credits equal or exceed the percentage of the whole.
 572 *
 573 * Return value is what to write into the CSR: trigger return when
 574 * unreturned credits pass this count.
 575 */
 576u32 sc_percent_to_threshold(struct send_context *sc, u32 percent)
 577{
 578	return (sc->credits * percent) / 100;
 579}
 580
 581/*
 582 * Set the credit return threshold.
 583 */
 584void sc_set_cr_threshold(struct send_context *sc, u32 new_threshold)
 585{
 586	unsigned long flags;
 587	u32 old_threshold;
 588	int force_return = 0;
 589
 590	spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
 591
 592	old_threshold = (sc->credit_ctrl >>
 593				SC(CREDIT_CTRL_THRESHOLD_SHIFT))
 594			 & SC(CREDIT_CTRL_THRESHOLD_MASK);
 595
 596	if (new_threshold != old_threshold) {
 597		sc->credit_ctrl =
 598			(sc->credit_ctrl
 599				& ~SC(CREDIT_CTRL_THRESHOLD_SMASK))
 600			| ((new_threshold
 601				& SC(CREDIT_CTRL_THRESHOLD_MASK))
 602			   << SC(CREDIT_CTRL_THRESHOLD_SHIFT));
 603		write_kctxt_csr(sc->dd, sc->hw_context,
 604				SC(CREDIT_CTRL), sc->credit_ctrl);
 605
 606		/* force a credit return on change to avoid a possible stall */
 607		force_return = 1;
 608	}
 609
 610	spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
 611
 612	if (force_return)
 613		sc_return_credits(sc);
 614}
 615
 616/*
 617 * set_pio_integrity
 618 *
 619 * Set the CHECK_ENABLE register for the send context 'sc'.
 620 */
 621void set_pio_integrity(struct send_context *sc)
 622{
 623	struct hfi1_devdata *dd = sc->dd;
 624	u32 hw_context = sc->hw_context;
 625	int type = sc->type;
 626
 627	write_kctxt_csr(dd, hw_context,
 628			SC(CHECK_ENABLE),
 629			hfi1_pkt_default_send_ctxt_mask(dd, type));
 630}
 631
 632static u32 get_buffers_allocated(struct send_context *sc)
 633{
 634	int cpu;
 635	u32 ret = 0;
 636
 637	for_each_possible_cpu(cpu)
 638		ret += *per_cpu_ptr(sc->buffers_allocated, cpu);
 639	return ret;
 640}
 641
 642static void reset_buffers_allocated(struct send_context *sc)
 643{
 644	int cpu;
 645
 646	for_each_possible_cpu(cpu)
 647		(*per_cpu_ptr(sc->buffers_allocated, cpu)) = 0;
 648}
 649
 650/*
 651 * Allocate a NUMA relative send context structure of the given type along
 652 * with a HW context.
 653 */
 654struct send_context *sc_alloc(struct hfi1_devdata *dd, int type,
 655			      uint hdrqentsize, int numa)
 656{
 657	struct send_context_info *sci;
 658	struct send_context *sc = NULL;
 659	dma_addr_t dma;
 660	unsigned long flags;
 661	u64 reg;
 662	u32 thresh;
 663	u32 sw_index;
 664	u32 hw_context;
 665	int ret;
 666	u8 opval, opmask;
 667
 668	/* do not allocate while frozen */
 669	if (dd->flags & HFI1_FROZEN)
 670		return NULL;
 671
 672	sc = kzalloc_node(sizeof(*sc), GFP_KERNEL, numa);
 673	if (!sc)
 674		return NULL;
 675
 676	sc->buffers_allocated = alloc_percpu(u32);
 677	if (!sc->buffers_allocated) {
 678		kfree(sc);
 679		dd_dev_err(dd,
 680			   "Cannot allocate buffers_allocated per cpu counters\n"
 681			  );
 682		return NULL;
 683	}
 684
 685	spin_lock_irqsave(&dd->sc_lock, flags);
 686	ret = sc_hw_alloc(dd, type, &sw_index, &hw_context);
 687	if (ret) {
 688		spin_unlock_irqrestore(&dd->sc_lock, flags);
 689		free_percpu(sc->buffers_allocated);
 690		kfree(sc);
 691		return NULL;
 692	}
 693
 694	sci = &dd->send_contexts[sw_index];
 695	sci->sc = sc;
 696
 697	sc->dd = dd;
 698	sc->node = numa;
 699	sc->type = type;
 700	spin_lock_init(&sc->alloc_lock);
 701	spin_lock_init(&sc->release_lock);
 702	spin_lock_init(&sc->credit_ctrl_lock);
 703	seqlock_init(&sc->waitlock);
 704	INIT_LIST_HEAD(&sc->piowait);
 705	INIT_WORK(&sc->halt_work, sc_halted);
 706	init_waitqueue_head(&sc->halt_wait);
 707
 708	/* grouping is always single context for now */
 709	sc->group = 0;
 710
 711	sc->sw_index = sw_index;
 712	sc->hw_context = hw_context;
 713	cr_group_addresses(sc, &dma);
 714	sc->credits = sci->credits;
 715	sc->size = sc->credits * PIO_BLOCK_SIZE;
 716
 717/* PIO Send Memory Address details */
 718#define PIO_ADDR_CONTEXT_MASK 0xfful
 719#define PIO_ADDR_CONTEXT_SHIFT 16
 720	sc->base_addr = dd->piobase + ((hw_context & PIO_ADDR_CONTEXT_MASK)
 721					<< PIO_ADDR_CONTEXT_SHIFT);
 722
 723	/* set base and credits */
 724	reg = ((sci->credits & SC(CTRL_CTXT_DEPTH_MASK))
 725					<< SC(CTRL_CTXT_DEPTH_SHIFT))
 726		| ((sci->base & SC(CTRL_CTXT_BASE_MASK))
 727					<< SC(CTRL_CTXT_BASE_SHIFT));
 728	write_kctxt_csr(dd, hw_context, SC(CTRL), reg);
 729
 730	set_pio_integrity(sc);
 731
 732	/* unmask all errors */
 733	write_kctxt_csr(dd, hw_context, SC(ERR_MASK), (u64)-1);
 734
 735	/* set the default partition key */
 736	write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY),
 737			(SC(CHECK_PARTITION_KEY_VALUE_MASK) &
 738			 DEFAULT_PKEY) <<
 739			SC(CHECK_PARTITION_KEY_VALUE_SHIFT));
 740
 741	/* per context type checks */
 742	if (type == SC_USER) {
 743		opval = USER_OPCODE_CHECK_VAL;
 744		opmask = USER_OPCODE_CHECK_MASK;
 745	} else {
 746		opval = OPCODE_CHECK_VAL_DISABLED;
 747		opmask = OPCODE_CHECK_MASK_DISABLED;
 748	}
 749
 750	/* set the send context check opcode mask and value */
 751	write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE),
 752			((u64)opmask << SC(CHECK_OPCODE_MASK_SHIFT)) |
 753			((u64)opval << SC(CHECK_OPCODE_VALUE_SHIFT)));
 754
 755	/* set up credit return */
 756	reg = dma & SC(CREDIT_RETURN_ADDR_ADDRESS_SMASK);
 757	write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), reg);
 758
 759	/*
 760	 * Calculate the initial credit return threshold.
 761	 *
 762	 * For Ack contexts, set a threshold for half the credits.
 763	 * For User contexts use the given percentage.  This has been
 764	 * sanitized on driver start-up.
 765	 * For Kernel contexts, use the default MTU plus a header
 766	 * or half the credits, whichever is smaller. This should
 767	 * work for both the 3-deep buffering allocation and the
 768	 * pooling allocation.
 769	 */
 770	if (type == SC_ACK) {
 771		thresh = sc_percent_to_threshold(sc, 50);
 772	} else if (type == SC_USER) {
 773		thresh = sc_percent_to_threshold(sc,
 774						 user_credit_return_threshold);
 775	} else { /* kernel */
 776		thresh = min(sc_percent_to_threshold(sc, 50),
 777			     sc_mtu_to_threshold(sc, hfi1_max_mtu,
 778						 hdrqentsize));
 779	}
 780	reg = thresh << SC(CREDIT_CTRL_THRESHOLD_SHIFT);
 781	/* add in early return */
 782	if (type == SC_USER && HFI1_CAP_IS_USET(EARLY_CREDIT_RETURN))
 783		reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK);
 784	else if (HFI1_CAP_IS_KSET(EARLY_CREDIT_RETURN)) /* kernel, ack */
 785		reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK);
 786
 787	/* set up write-through credit_ctrl */
 788	sc->credit_ctrl = reg;
 789	write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), reg);
 790
 791	/* User send contexts should not allow sending on VL15 */
 792	if (type == SC_USER) {
 793		reg = 1ULL << 15;
 794		write_kctxt_csr(dd, hw_context, SC(CHECK_VL), reg);
 795	}
 796
 797	spin_unlock_irqrestore(&dd->sc_lock, flags);
 798
 799	/*
 800	 * Allocate shadow ring to track outstanding PIO buffers _after_
 801	 * unlocking.  We don't know the size until the lock is held and
 802	 * we can't allocate while the lock is held.  No one is using
 803	 * the context yet, so allocate it now.
 804	 *
 805	 * User contexts do not get a shadow ring.
 806	 */
 807	if (type != SC_USER) {
 808		/*
 809		 * Size the shadow ring 1 larger than the number of credits
 810		 * so head == tail can mean empty.
 811		 */
 812		sc->sr_size = sci->credits + 1;
 813		sc->sr = kcalloc_node(sc->sr_size,
 814				      sizeof(union pio_shadow_ring),
 815				      GFP_KERNEL, numa);
 816		if (!sc->sr) {
 817			sc_free(sc);
 818			return NULL;
 819		}
 820	}
 821
 822	hfi1_cdbg(PIO,
 823		  "Send context %u(%u) %s group %u credits %u credit_ctrl 0x%llx threshold %u",
 824		  sw_index,
 825		  hw_context,
 826		  sc_type_name(type),
 827		  sc->group,
 828		  sc->credits,
 829		  sc->credit_ctrl,
 830		  thresh);
 831
 832	return sc;
 833}
 834
 835/* free a per-NUMA send context structure */
 836void sc_free(struct send_context *sc)
 837{
 838	struct hfi1_devdata *dd;
 839	unsigned long flags;
 840	u32 sw_index;
 841	u32 hw_context;
 842
 843	if (!sc)
 844		return;
 845
 846	sc->flags |= SCF_IN_FREE;	/* ensure no restarts */
 847	dd = sc->dd;
 848	if (!list_empty(&sc->piowait))
 849		dd_dev_err(dd, "piowait list not empty!\n");
 850	sw_index = sc->sw_index;
 851	hw_context = sc->hw_context;
 852	sc_disable(sc);	/* make sure the HW is disabled */
 853	flush_work(&sc->halt_work);
 854
 855	spin_lock_irqsave(&dd->sc_lock, flags);
 856	dd->send_contexts[sw_index].sc = NULL;
 857
 858	/* clear/disable all registers set in sc_alloc */
 859	write_kctxt_csr(dd, hw_context, SC(CTRL), 0);
 860	write_kctxt_csr(dd, hw_context, SC(CHECK_ENABLE), 0);
 861	write_kctxt_csr(dd, hw_context, SC(ERR_MASK), 0);
 862	write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY), 0);
 863	write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE), 0);
 864	write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), 0);
 865	write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), 0);
 866
 867	/* release the index and context for re-use */
 868	sc_hw_free(dd, sw_index, hw_context);
 869	spin_unlock_irqrestore(&dd->sc_lock, flags);
 870
 871	kfree(sc->sr);
 872	free_percpu(sc->buffers_allocated);
 873	kfree(sc);
 874}
 875
 876/* disable the context */
 877void sc_disable(struct send_context *sc)
 878{
 879	u64 reg;
 880	struct pio_buf *pbuf;
 881	LIST_HEAD(wake_list);
 882
 883	if (!sc)
 884		return;
 885
 886	/* do all steps, even if already disabled */
 887	spin_lock_irq(&sc->alloc_lock);
 888	reg = read_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL));
 889	reg &= ~SC(CTRL_CTXT_ENABLE_SMASK);
 890	sc->flags &= ~SCF_ENABLED;
 891	sc_wait_for_packet_egress(sc, 1);
 892	write_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL), reg);
 893
 894	/*
 895	 * Flush any waiters.  Once the context is disabled,
 896	 * credit return interrupts are stopped (although there
 897	 * could be one in-process when the context is disabled).
 898	 * Wait one microsecond for any lingering interrupts, then
 899	 * proceed with the flush.
 900	 */
 901	udelay(1);
 902	spin_lock(&sc->release_lock);
 903	if (sc->sr) {	/* this context has a shadow ring */
 904		while (sc->sr_tail != sc->sr_head) {
 905			pbuf = &sc->sr[sc->sr_tail].pbuf;
 906			if (pbuf->cb)
 907				(*pbuf->cb)(pbuf->arg, PRC_SC_DISABLE);
 908			sc->sr_tail++;
 909			if (sc->sr_tail >= sc->sr_size)
 910				sc->sr_tail = 0;
 911		}
 912	}
 913	spin_unlock(&sc->release_lock);
 914
 915	write_seqlock(&sc->waitlock);
 916	list_splice_init(&sc->piowait, &wake_list);
 917	write_sequnlock(&sc->waitlock);
 918	while (!list_empty(&wake_list)) {
 919		struct iowait *wait;
 920		struct rvt_qp *qp;
 921		struct hfi1_qp_priv *priv;
 922
 923		wait = list_first_entry(&wake_list, struct iowait, list);
 924		qp = iowait_to_qp(wait);
 925		priv = qp->priv;
 926		list_del_init(&priv->s_iowait.list);
 927		priv->s_iowait.lock = NULL;
 928		hfi1_qp_wakeup(qp, RVT_S_WAIT_PIO | HFI1_S_WAIT_PIO_DRAIN);
 929	}
 930
 931	spin_unlock_irq(&sc->alloc_lock);
 932}
 933
 934/* return SendEgressCtxtStatus.PacketOccupancy */
 935static u64 packet_occupancy(u64 reg)
 936{
 937	return (reg &
 938		SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SMASK)
 939		>> SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SHIFT;
 940}
 941
 942/* is egress halted on the context? */
 943static bool egress_halted(u64 reg)
 944{
 945	return !!(reg & SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_HALT_STATUS_SMASK);
 946}
 947
 948/* is the send context halted? */
 949static bool is_sc_halted(struct hfi1_devdata *dd, u32 hw_context)
 950{
 951	return !!(read_kctxt_csr(dd, hw_context, SC(STATUS)) &
 952		  SC(STATUS_CTXT_HALTED_SMASK));
 953}
 954
 955/**
 956 * sc_wait_for_packet_egress - wait for packet
 957 * @sc: valid send context
 958 * @pause: wait for credit return
 959 *
 960 * Wait for packet egress, optionally pause for credit return
 961 *
 962 * Egress halt and Context halt are not necessarily the same thing, so
 963 * check for both.
 964 *
 965 * NOTE: The context halt bit may not be set immediately.  Because of this,
 966 * it is necessary to check the SW SFC_HALTED bit (set in the IRQ) and the HW
 967 * context bit to determine if the context is halted.
 968 */
 969static void sc_wait_for_packet_egress(struct send_context *sc, int pause)
 970{
 971	struct hfi1_devdata *dd = sc->dd;
 972	u64 reg = 0;
 973	u64 reg_prev;
 974	u32 loop = 0;
 975
 976	while (1) {
 977		reg_prev = reg;
 978		reg = read_csr(dd, sc->hw_context * 8 +
 979			       SEND_EGRESS_CTXT_STATUS);
 980		/* done if any halt bits, SW or HW are set */
 981		if (sc->flags & SCF_HALTED ||
 982		    is_sc_halted(dd, sc->hw_context) || egress_halted(reg))
 983			break;
 984		reg = packet_occupancy(reg);
 985		if (reg == 0)
 986			break;
 987		/* counter is reset if occupancy count changes */
 988		if (reg != reg_prev)
 989			loop = 0;
 990		if (loop > 50000) {
 991			/* timed out - bounce the link */
 992			dd_dev_err(dd,
 993				   "%s: context %u(%u) timeout waiting for packets to egress, remaining count %u, bouncing link\n",
 994				   __func__, sc->sw_index,
 995				   sc->hw_context, (u32)reg);
 996			queue_work(dd->pport->link_wq,
 997				   &dd->pport->link_bounce_work);
 998			break;
 999		}
1000		loop++;
1001		udelay(1);
1002	}
1003
1004	if (pause)
1005		/* Add additional delay to ensure chip returns all credits */
1006		pause_for_credit_return(dd);
1007}
1008
1009void sc_wait(struct hfi1_devdata *dd)
1010{
1011	int i;
1012
1013	for (i = 0; i < dd->num_send_contexts; i++) {
1014		struct send_context *sc = dd->send_contexts[i].sc;
1015
1016		if (!sc)
1017			continue;
1018		sc_wait_for_packet_egress(sc, 0);
1019	}
1020}
1021
1022/*
1023 * Restart a context after it has been halted due to error.
1024 *
1025 * If the first step fails - wait for the halt to be asserted, return early.
1026 * Otherwise complain about timeouts but keep going.
1027 *
1028 * It is expected that allocations (enabled flag bit) have been shut off
1029 * already (only applies to kernel contexts).
1030 */
1031int sc_restart(struct send_context *sc)
1032{
1033	struct hfi1_devdata *dd = sc->dd;
1034	u64 reg;
1035	u32 loop;
1036	int count;
1037
1038	/* bounce off if not halted, or being free'd */
1039	if (!(sc->flags & SCF_HALTED) || (sc->flags & SCF_IN_FREE))
1040		return -EINVAL;
1041
1042	dd_dev_info(dd, "restarting send context %u(%u)\n", sc->sw_index,
1043		    sc->hw_context);
1044
1045	/*
1046	 * Step 1: Wait for the context to actually halt.
1047	 *
1048	 * The error interrupt is asynchronous to actually setting halt
1049	 * on the context.
1050	 */
1051	loop = 0;
1052	while (1) {
1053		reg = read_kctxt_csr(dd, sc->hw_context, SC(STATUS));
1054		if (reg & SC(STATUS_CTXT_HALTED_SMASK))
1055			break;
1056		if (loop > 100) {
1057			dd_dev_err(dd, "%s: context %u(%u) not halting, skipping\n",
1058				   __func__, sc->sw_index, sc->hw_context);
1059			return -ETIME;
1060		}
1061		loop++;
1062		udelay(1);
1063	}
1064
1065	/*
1066	 * Step 2: Ensure no users are still trying to write to PIO.
1067	 *
1068	 * For kernel contexts, we have already turned off buffer allocation.
1069	 * Now wait for the buffer count to go to zero.
1070	 *
1071	 * For user contexts, the user handling code has cut off write access
1072	 * to the context's PIO pages before calling this routine and will
1073	 * restore write access after this routine returns.
1074	 */
1075	if (sc->type != SC_USER) {
1076		/* kernel context */
1077		loop = 0;
1078		while (1) {
1079			count = get_buffers_allocated(sc);
1080			if (count == 0)
1081				break;
1082			if (loop > 100) {
1083				dd_dev_err(dd,
1084					   "%s: context %u(%u) timeout waiting for PIO buffers to zero, remaining %d\n",
1085					   __func__, sc->sw_index,
1086					   sc->hw_context, count);
1087			}
1088			loop++;
1089			udelay(1);
1090		}
1091	}
1092
1093	/*
1094	 * Step 3: Wait for all packets to egress.
1095	 * This is done while disabling the send context
1096	 *
1097	 * Step 4: Disable the context
1098	 *
1099	 * This is a superset of the halt.  After the disable, the
1100	 * errors can be cleared.
1101	 */
1102	sc_disable(sc);
1103
1104	/*
1105	 * Step 5: Enable the context
1106	 *
1107	 * This enable will clear the halted flag and per-send context
1108	 * error flags.
1109	 */
1110	return sc_enable(sc);
1111}
1112
1113/*
1114 * PIO freeze processing.  To be called after the TXE block is fully frozen.
1115 * Go through all frozen send contexts and disable them.  The contexts are
1116 * already stopped by the freeze.
1117 */
1118void pio_freeze(struct hfi1_devdata *dd)
1119{
1120	struct send_context *sc;
1121	int i;
1122
1123	for (i = 0; i < dd->num_send_contexts; i++) {
1124		sc = dd->send_contexts[i].sc;
1125		/*
1126		 * Don't disable unallocated, unfrozen, or user send contexts.
1127		 * User send contexts will be disabled when the process
1128		 * calls into the driver to reset its context.
1129		 */
1130		if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER)
1131			continue;
1132
1133		/* only need to disable, the context is already stopped */
1134		sc_disable(sc);
1135	}
1136}
1137
1138/*
1139 * Unfreeze PIO for kernel send contexts.  The precondition for calling this
1140 * is that all PIO send contexts have been disabled and the SPC freeze has
1141 * been cleared.  Now perform the last step and re-enable each kernel context.
1142 * User (PSM) processing will occur when PSM calls into the kernel to
1143 * acknowledge the freeze.
1144 */
1145void pio_kernel_unfreeze(struct hfi1_devdata *dd)
1146{
1147	struct send_context *sc;
1148	int i;
1149
1150	for (i = 0; i < dd->num_send_contexts; i++) {
1151		sc = dd->send_contexts[i].sc;
1152		if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER)
1153			continue;
1154		if (sc->flags & SCF_LINK_DOWN)
1155			continue;
1156
1157		sc_enable(sc);	/* will clear the sc frozen flag */
1158	}
1159}
1160
1161/**
1162 * pio_kernel_linkup() - Re-enable send contexts after linkup event
1163 * @dd: valid devive data
1164 *
1165 * When the link goes down, the freeze path is taken.  However, a link down
1166 * event is different from a freeze because if the send context is re-enabled
1167 * whowever is sending data will start sending data again, which will hang
1168 * any QP that is sending data.
1169 *
1170 * The freeze path now looks at the type of event that occurs and takes this
1171 * path for link down event.
1172 */
1173void pio_kernel_linkup(struct hfi1_devdata *dd)
1174{
1175	struct send_context *sc;
1176	int i;
1177
1178	for (i = 0; i < dd->num_send_contexts; i++) {
1179		sc = dd->send_contexts[i].sc;
1180		if (!sc || !(sc->flags & SCF_LINK_DOWN) || sc->type == SC_USER)
1181			continue;
1182
1183		sc_enable(sc);	/* will clear the sc link down flag */
1184	}
1185}
1186
1187/*
1188 * Wait for the SendPioInitCtxt.PioInitInProgress bit to clear.
1189 * Returns:
1190 *	-ETIMEDOUT - if we wait too long
1191 *	-EIO	   - if there was an error
1192 */
1193static int pio_init_wait_progress(struct hfi1_devdata *dd)
1194{
1195	u64 reg;
1196	int max, count = 0;
1197
1198	/* max is the longest possible HW init time / delay */
1199	max = (dd->icode == ICODE_FPGA_EMULATION) ? 120 : 5;
1200	while (1) {
1201		reg = read_csr(dd, SEND_PIO_INIT_CTXT);
1202		if (!(reg & SEND_PIO_INIT_CTXT_PIO_INIT_IN_PROGRESS_SMASK))
1203			break;
1204		if (count >= max)
1205			return -ETIMEDOUT;
1206		udelay(5);
1207		count++;
1208	}
1209
1210	return reg & SEND_PIO_INIT_CTXT_PIO_INIT_ERR_SMASK ? -EIO : 0;
1211}
1212
1213/*
1214 * Reset all of the send contexts to their power-on state.  Used
1215 * only during manual init - no lock against sc_enable needed.
1216 */
1217void pio_reset_all(struct hfi1_devdata *dd)
1218{
1219	int ret;
1220
1221	/* make sure the init engine is not busy */
1222	ret = pio_init_wait_progress(dd);
1223	/* ignore any timeout */
1224	if (ret == -EIO) {
1225		/* clear the error */
1226		write_csr(dd, SEND_PIO_ERR_CLEAR,
1227			  SEND_PIO_ERR_CLEAR_PIO_INIT_SM_IN_ERR_SMASK);
1228	}
1229
1230	/* reset init all */
1231	write_csr(dd, SEND_PIO_INIT_CTXT,
1232		  SEND_PIO_INIT_CTXT_PIO_ALL_CTXT_INIT_SMASK);
1233	udelay(2);
1234	ret = pio_init_wait_progress(dd);
1235	if (ret < 0) {
1236		dd_dev_err(dd,
1237			   "PIO send context init %s while initializing all PIO blocks\n",
1238			   ret == -ETIMEDOUT ? "is stuck" : "had an error");
1239	}
1240}
1241
1242/* enable the context */
1243int sc_enable(struct send_context *sc)
1244{
1245	u64 sc_ctrl, reg, pio;
1246	struct hfi1_devdata *dd;
1247	unsigned long flags;
1248	int ret = 0;
1249
1250	if (!sc)
1251		return -EINVAL;
1252	dd = sc->dd;
1253
1254	/*
1255	 * Obtain the allocator lock to guard against any allocation
1256	 * attempts (which should not happen prior to context being
1257	 * enabled). On the release/disable side we don't need to
1258	 * worry about locking since the releaser will not do anything
1259	 * if the context accounting values have not changed.
1260	 */
1261	spin_lock_irqsave(&sc->alloc_lock, flags);
1262	sc_ctrl = read_kctxt_csr(dd, sc->hw_context, SC(CTRL));
1263	if ((sc_ctrl & SC(CTRL_CTXT_ENABLE_SMASK)))
1264		goto unlock; /* already enabled */
1265
1266	/* IMPORTANT: only clear free and fill if transitioning 0 -> 1 */
1267
1268	*sc->hw_free = 0;
1269	sc->free = 0;
1270	sc->alloc_free = 0;
1271	sc->fill = 0;
1272	sc->fill_wrap = 0;
1273	sc->sr_head = 0;
1274	sc->sr_tail = 0;
1275	sc->flags = 0;
1276	/* the alloc lock insures no fast path allocation */
1277	reset_buffers_allocated(sc);
1278
1279	/*
1280	 * Clear all per-context errors.  Some of these will be set when
1281	 * we are re-enabling after a context halt.  Now that the context
1282	 * is disabled, the halt will not clear until after the PIO init
1283	 * engine runs below.
1284	 */
1285	reg = read_kctxt_csr(dd, sc->hw_context, SC(ERR_STATUS));
1286	if (reg)
1287		write_kctxt_csr(dd, sc->hw_context, SC(ERR_CLEAR), reg);
1288
1289	/*
1290	 * The HW PIO initialization engine can handle only one init
1291	 * request at a time. Serialize access to each device's engine.
1292	 */
1293	spin_lock(&dd->sc_init_lock);
1294	/*
1295	 * Since access to this code block is serialized and
1296	 * each access waits for the initialization to complete
1297	 * before releasing the lock, the PIO initialization engine
1298	 * should not be in use, so we don't have to wait for the
1299	 * InProgress bit to go down.
1300	 */
1301	pio = ((sc->hw_context & SEND_PIO_INIT_CTXT_PIO_CTXT_NUM_MASK) <<
1302	       SEND_PIO_INIT_CTXT_PIO_CTXT_NUM_SHIFT) |
1303		SEND_PIO_INIT_CTXT_PIO_SINGLE_CTXT_INIT_SMASK;
1304	write_csr(dd, SEND_PIO_INIT_CTXT, pio);
1305	/*
1306	 * Wait until the engine is done.  Give the chip the required time
1307	 * so, hopefully, we read the register just once.
1308	 */
1309	udelay(2);
1310	ret = pio_init_wait_progress(dd);
1311	spin_unlock(&dd->sc_init_lock);
1312	if (ret) {
1313		dd_dev_err(dd,
1314			   "sctxt%u(%u): Context not enabled due to init failure %d\n",
1315			   sc->sw_index, sc->hw_context, ret);
1316		goto unlock;
1317	}
1318
1319	/*
1320	 * All is well. Enable the context.
1321	 */
1322	sc_ctrl |= SC(CTRL_CTXT_ENABLE_SMASK);
1323	write_kctxt_csr(dd, sc->hw_context, SC(CTRL), sc_ctrl);
1324	/*
1325	 * Read SendCtxtCtrl to force the write out and prevent a timing
1326	 * hazard where a PIO write may reach the context before the enable.
1327	 */
1328	read_kctxt_csr(dd, sc->hw_context, SC(CTRL));
1329	sc->flags |= SCF_ENABLED;
1330
1331unlock:
1332	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1333
1334	return ret;
1335}
1336
1337/* force a credit return on the context */
1338void sc_return_credits(struct send_context *sc)
1339{
1340	if (!sc)
1341		return;
1342
1343	/* a 0->1 transition schedules a credit return */
1344	write_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE),
1345			SC(CREDIT_FORCE_FORCE_RETURN_SMASK));
1346	/*
1347	 * Ensure that the write is flushed and the credit return is
1348	 * scheduled. We care more about the 0 -> 1 transition.
1349	 */
1350	read_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE));
1351	/* set back to 0 for next time */
1352	write_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE), 0);
1353}
1354
1355/* allow all in-flight packets to drain on the context */
1356void sc_flush(struct send_context *sc)
1357{
1358	if (!sc)
1359		return;
1360
1361	sc_wait_for_packet_egress(sc, 1);
1362}
1363
1364/* drop all packets on the context, no waiting until they are sent */
1365void sc_drop(struct send_context *sc)
1366{
1367	if (!sc)
1368		return;
1369
1370	dd_dev_info(sc->dd, "%s: context %u(%u) - not implemented\n",
1371		    __func__, sc->sw_index, sc->hw_context);
1372}
1373
1374/*
1375 * Start the software reaction to a context halt or SPC freeze:
1376 *	- mark the context as halted or frozen
1377 *	- stop buffer allocations
1378 *
1379 * Called from the error interrupt.  Other work is deferred until
1380 * out of the interrupt.
1381 */
1382void sc_stop(struct send_context *sc, int flag)
1383{
1384	unsigned long flags;
1385
1386	/* stop buffer allocations */
1387	spin_lock_irqsave(&sc->alloc_lock, flags);
1388	/* mark the context */
1389	sc->flags |= flag;
1390	sc->flags &= ~SCF_ENABLED;
1391	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1392	wake_up(&sc->halt_wait);
1393}
1394
1395#define BLOCK_DWORDS (PIO_BLOCK_SIZE / sizeof(u32))
1396#define dwords_to_blocks(x) DIV_ROUND_UP(x, BLOCK_DWORDS)
1397
1398/*
1399 * The send context buffer "allocator".
1400 *
1401 * @sc: the PIO send context we are allocating from
1402 * @len: length of whole packet - including PBC - in dwords
1403 * @cb: optional callback to call when the buffer is finished sending
1404 * @arg: argument for cb
1405 *
1406 * Return a pointer to a PIO buffer, NULL if not enough room, -ECOMM
1407 * when link is down.
1408 */
1409struct pio_buf *sc_buffer_alloc(struct send_context *sc, u32 dw_len,
1410				pio_release_cb cb, void *arg)
1411{
1412	struct pio_buf *pbuf = NULL;
1413	unsigned long flags;
1414	unsigned long avail;
1415	unsigned long blocks = dwords_to_blocks(dw_len);
1416	u32 fill_wrap;
1417	int trycount = 0;
1418	u32 head, next;
1419
1420	spin_lock_irqsave(&sc->alloc_lock, flags);
1421	if (!(sc->flags & SCF_ENABLED)) {
1422		spin_unlock_irqrestore(&sc->alloc_lock, flags);
1423		return ERR_PTR(-ECOMM);
1424	}
1425
1426retry:
1427	avail = (unsigned long)sc->credits - (sc->fill - sc->alloc_free);
1428	if (blocks > avail) {
1429		/* not enough room */
1430		if (unlikely(trycount))	{ /* already tried to get more room */
1431			spin_unlock_irqrestore(&sc->alloc_lock, flags);
1432			goto done;
1433		}
1434		/* copy from receiver cache line and recalculate */
1435		sc->alloc_free = READ_ONCE(sc->free);
1436		avail =
1437			(unsigned long)sc->credits -
1438			(sc->fill - sc->alloc_free);
1439		if (blocks > avail) {
1440			/* still no room, actively update */
1441			sc_release_update(sc);
1442			sc->alloc_free = READ_ONCE(sc->free);
1443			trycount++;
1444			goto retry;
1445		}
1446	}
1447
1448	/* there is enough room */
1449
1450	preempt_disable();
1451	this_cpu_inc(*sc->buffers_allocated);
1452
1453	/* read this once */
1454	head = sc->sr_head;
1455
1456	/* "allocate" the buffer */
1457	sc->fill += blocks;
1458	fill_wrap = sc->fill_wrap;
1459	sc->fill_wrap += blocks;
1460	if (sc->fill_wrap >= sc->credits)
1461		sc->fill_wrap = sc->fill_wrap - sc->credits;
1462
1463	/*
1464	 * Fill the parts that the releaser looks at before moving the head.
1465	 * The only necessary piece is the sent_at field.  The credits
1466	 * we have just allocated cannot have been returned yet, so the
1467	 * cb and arg will not be looked at for a "while".  Put them
1468	 * on this side of the memory barrier anyway.
1469	 */
1470	pbuf = &sc->sr[head].pbuf;
1471	pbuf->sent_at = sc->fill;
1472	pbuf->cb = cb;
1473	pbuf->arg = arg;
1474	pbuf->sc = sc;	/* could be filled in at sc->sr init time */
1475	/* make sure this is in memory before updating the head */
1476
1477	/* calculate next head index, do not store */
1478	next = head + 1;
1479	if (next >= sc->sr_size)
1480		next = 0;
1481	/*
1482	 * update the head - must be last! - the releaser can look at fields
1483	 * in pbuf once we move the head
1484	 */
1485	smp_wmb();
1486	sc->sr_head = next;
1487	spin_unlock_irqrestore(&sc->alloc_lock, flags);
1488
1489	/* finish filling in the buffer outside the lock */
1490	pbuf->start = sc->base_addr + fill_wrap * PIO_BLOCK_SIZE;
1491	pbuf->end = sc->base_addr + sc->size;
1492	pbuf->qw_written = 0;
1493	pbuf->carry_bytes = 0;
1494	pbuf->carry.val64 = 0;
1495done:
1496	return pbuf;
1497}
1498
1499/*
1500 * There are at least two entities that can turn on credit return
1501 * interrupts and they can overlap.  Avoid problems by implementing
1502 * a count scheme that is enforced by a lock.  The lock is needed because
1503 * the count and CSR write must be paired.
1504 */
1505
1506/*
1507 * Start credit return interrupts.  This is managed by a count.  If already
1508 * on, just increment the count.
1509 */
1510void sc_add_credit_return_intr(struct send_context *sc)
1511{
1512	unsigned long flags;
1513
1514	/* lock must surround both the count change and the CSR update */
1515	spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
1516	if (sc->credit_intr_count == 0) {
1517		sc->credit_ctrl |= SC(CREDIT_CTRL_CREDIT_INTR_SMASK);
1518		write_kctxt_csr(sc->dd, sc->hw_context,
1519				SC(CREDIT_CTRL), sc->credit_ctrl);
1520	}
1521	sc->credit_intr_count++;
1522	spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
1523}
1524
1525/*
1526 * Stop credit return interrupts.  This is managed by a count.  Decrement the
1527 * count, if the last user, then turn the credit interrupts off.
1528 */
1529void sc_del_credit_return_intr(struct send_context *sc)
1530{
1531	unsigned long flags;
1532
1533	WARN_ON(sc->credit_intr_count == 0);
1534
1535	/* lock must surround both the count change and the CSR update */
1536	spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
1537	sc->credit_intr_count--;
1538	if (sc->credit_intr_count == 0) {
1539		sc->credit_ctrl &= ~SC(CREDIT_CTRL_CREDIT_INTR_SMASK);
1540		write_kctxt_csr(sc->dd, sc->hw_context,
1541				SC(CREDIT_CTRL), sc->credit_ctrl);
1542	}
1543	spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
1544}
1545
1546/*
1547 * The caller must be careful when calling this.  All needint calls
1548 * must be paired with !needint.
1549 */
1550void hfi1_sc_wantpiobuf_intr(struct send_context *sc, u32 needint)
1551{
1552	if (needint)
1553		sc_add_credit_return_intr(sc);
1554	else
1555		sc_del_credit_return_intr(sc);
1556	trace_hfi1_wantpiointr(sc, needint, sc->credit_ctrl);
1557	if (needint)
1558		sc_return_credits(sc);
1559}
1560
1561/**
1562 * sc_piobufavail - callback when a PIO buffer is available
1563 * @sc: the send context
1564 *
1565 * This is called from the interrupt handler when a PIO buffer is
1566 * available after hfi1_verbs_send() returned an error that no buffers were
1567 * available. Disable the interrupt if there are no more QPs waiting.
1568 */
1569static void sc_piobufavail(struct send_context *sc)
1570{
1571	struct hfi1_devdata *dd = sc->dd;
1572	struct list_head *list;
1573	struct rvt_qp *qps[PIO_WAIT_BATCH_SIZE];
1574	struct rvt_qp *qp;
1575	struct hfi1_qp_priv *priv;
1576	unsigned long flags;
1577	uint i, n = 0, top_idx = 0;
1578
1579	if (dd->send_contexts[sc->sw_index].type != SC_KERNEL &&
1580	    dd->send_contexts[sc->sw_index].type != SC_VL15)
1581		return;
1582	list = &sc->piowait;
1583	/*
1584	 * Note: checking that the piowait list is empty and clearing
1585	 * the buffer available interrupt needs to be atomic or we
1586	 * could end up with QPs on the wait list with the interrupt
1587	 * disabled.
1588	 */
1589	write_seqlock_irqsave(&sc->waitlock, flags);
1590	while (!list_empty(list)) {
1591		struct iowait *wait;
1592
1593		if (n == ARRAY_SIZE(qps))
1594			break;
1595		wait = list_first_entry(list, struct iowait, list);
1596		iowait_get_priority(wait);
1597		qp = iowait_to_qp(wait);
1598		priv = qp->priv;
1599		list_del_init(&priv->s_iowait.list);
1600		priv->s_iowait.lock = NULL;
1601		if (n) {
1602			priv = qps[top_idx]->priv;
1603			top_idx = iowait_priority_update_top(wait,
1604							     &priv->s_iowait,
1605							     n, top_idx);
1606		}
1607
1608		/* refcount held until actual wake up */
1609		qps[n++] = qp;
1610	}
1611	/*
1612	 * If there had been waiters and there are more
1613	 * insure that we redo the force to avoid a potential hang.
1614	 */
1615	if (n) {
1616		hfi1_sc_wantpiobuf_intr(sc, 0);
1617		if (!list_empty(list))
1618			hfi1_sc_wantpiobuf_intr(sc, 1);
1619	}
1620	write_sequnlock_irqrestore(&sc->waitlock, flags);
1621
1622	/* Wake up the top-priority one first */
1623	if (n)
1624		hfi1_qp_wakeup(qps[top_idx],
1625			       RVT_S_WAIT_PIO | HFI1_S_WAIT_PIO_DRAIN);
1626	for (i = 0; i < n; i++)
1627		if (i != top_idx)
1628			hfi1_qp_wakeup(qps[i],
1629				       RVT_S_WAIT_PIO | HFI1_S_WAIT_PIO_DRAIN);
1630}
1631
1632/* translate a send credit update to a bit code of reasons */
1633static inline int fill_code(u64 hw_free)
1634{
1635	int code = 0;
1636
1637	if (hw_free & CR_STATUS_SMASK)
1638		code |= PRC_STATUS_ERR;
1639	if (hw_free & CR_CREDIT_RETURN_DUE_TO_PBC_SMASK)
1640		code |= PRC_PBC;
1641	if (hw_free & CR_CREDIT_RETURN_DUE_TO_THRESHOLD_SMASK)
1642		code |= PRC_THRESHOLD;
1643	if (hw_free & CR_CREDIT_RETURN_DUE_TO_ERR_SMASK)
1644		code |= PRC_FILL_ERR;
1645	if (hw_free & CR_CREDIT_RETURN_DUE_TO_FORCE_SMASK)
1646		code |= PRC_SC_DISABLE;
1647	return code;
1648}
1649
1650/* use the jiffies compare to get the wrap right */
1651#define sent_before(a, b) time_before(a, b)	/* a < b */
1652
1653/*
1654 * The send context buffer "releaser".
1655 */
1656void sc_release_update(struct send_context *sc)
1657{
1658	struct pio_buf *pbuf;
1659	u64 hw_free;
1660	u32 head, tail;
1661	unsigned long old_free;
1662	unsigned long free;
1663	unsigned long extra;
1664	unsigned long flags;
1665	int code;
1666
1667	if (!sc)
1668		return;
1669
1670	spin_lock_irqsave(&sc->release_lock, flags);
1671	/* update free */
1672	hw_free = le64_to_cpu(*sc->hw_free);		/* volatile read */
1673	old_free = sc->free;
1674	extra = (((hw_free & CR_COUNTER_SMASK) >> CR_COUNTER_SHIFT)
1675			- (old_free & CR_COUNTER_MASK))
1676				& CR_COUNTER_MASK;
1677	free = old_free + extra;
1678	trace_hfi1_piofree(sc, extra);
1679
1680	/* call sent buffer callbacks */
1681	code = -1;				/* code not yet set */
1682	head = READ_ONCE(sc->sr_head);	/* snapshot the head */
1683	tail = sc->sr_tail;
1684	while (head != tail) {
1685		pbuf = &sc->sr[tail].pbuf;
1686
1687		if (sent_before(free, pbuf->sent_at)) {
1688			/* not sent yet */
1689			break;
1690		}
1691		if (pbuf->cb) {
1692			if (code < 0) /* fill in code on first user */
1693				code = fill_code(hw_free);
1694			(*pbuf->cb)(pbuf->arg, code);
1695		}
1696
1697		tail++;
1698		if (tail >= sc->sr_size)
1699			tail = 0;
1700	}
1701	sc->sr_tail = tail;
1702	/* make sure tail is updated before free */
1703	smp_wmb();
1704	sc->free = free;
1705	spin_unlock_irqrestore(&sc->release_lock, flags);
1706	sc_piobufavail(sc);
1707}
1708
1709/*
1710 * Send context group releaser.  Argument is the send context that caused
1711 * the interrupt.  Called from the send context interrupt handler.
1712 *
1713 * Call release on all contexts in the group.
1714 *
1715 * This routine takes the sc_lock without an irqsave because it is only
1716 * called from an interrupt handler.  Adjust if that changes.
1717 */
1718void sc_group_release_update(struct hfi1_devdata *dd, u32 hw_context)
1719{
1720	struct send_context *sc;
1721	u32 sw_index;
1722	u32 gc, gc_end;
1723
1724	spin_lock(&dd->sc_lock);
1725	sw_index = dd->hw_to_sw[hw_context];
1726	if (unlikely(sw_index >= dd->num_send_contexts)) {
1727		dd_dev_err(dd, "%s: invalid hw (%u) to sw (%u) mapping\n",
1728			   __func__, hw_context, sw_index);
1729		goto done;
1730	}
1731	sc = dd->send_contexts[sw_index].sc;
1732	if (unlikely(!sc))
1733		goto done;
1734
1735	gc = group_context(hw_context, sc->group);
1736	gc_end = gc + group_size(sc->group);
1737	for (; gc < gc_end; gc++) {
1738		sw_index = dd->hw_to_sw[gc];
1739		if (unlikely(sw_index >= dd->num_send_contexts)) {
1740			dd_dev_err(dd,
1741				   "%s: invalid hw (%u) to sw (%u) mapping\n",
1742				   __func__, hw_context, sw_index);
1743			continue;
1744		}
1745		sc_release_update(dd->send_contexts[sw_index].sc);
1746	}
1747done:
1748	spin_unlock(&dd->sc_lock);
1749}
1750
1751/*
1752 * pio_select_send_context_vl() - select send context
1753 * @dd: devdata
1754 * @selector: a spreading factor
1755 * @vl: this vl
1756 *
1757 * This function returns a send context based on the selector and a vl.
1758 * The mapping fields are protected by RCU
1759 */
1760struct send_context *pio_select_send_context_vl(struct hfi1_devdata *dd,
1761						u32 selector, u8 vl)
1762{
1763	struct pio_vl_map *m;
1764	struct pio_map_elem *e;
1765	struct send_context *rval;
1766
1767	/*
1768	 * NOTE This should only happen if SC->VL changed after the initial
1769	 * checks on the QP/AH
1770	 * Default will return VL0's send context below
1771	 */
1772	if (unlikely(vl >= num_vls)) {
1773		rval = NULL;
1774		goto done;
1775	}
1776
1777	rcu_read_lock();
1778	m = rcu_dereference(dd->pio_map);
1779	if (unlikely(!m)) {
1780		rcu_read_unlock();
1781		return dd->vld[0].sc;
1782	}
1783	e = m->map[vl & m->mask];
1784	rval = e->ksc[selector & e->mask];
1785	rcu_read_unlock();
1786
1787done:
1788	rval = !rval ? dd->vld[0].sc : rval;
1789	return rval;
1790}
1791
1792/*
1793 * pio_select_send_context_sc() - select send context
1794 * @dd: devdata
1795 * @selector: a spreading factor
1796 * @sc5: the 5 bit sc
1797 *
1798 * This function returns an send context based on the selector and an sc
1799 */
1800struct send_context *pio_select_send_context_sc(struct hfi1_devdata *dd,
1801						u32 selector, u8 sc5)
1802{
1803	u8 vl = sc_to_vlt(dd, sc5);
1804
1805	return pio_select_send_context_vl(dd, selector, vl);
1806}
1807
1808/*
1809 * Free the indicated map struct
1810 */
1811static void pio_map_free(struct pio_vl_map *m)
1812{
1813	int i;
1814
1815	for (i = 0; m && i < m->actual_vls; i++)
1816		kfree(m->map[i]);
1817	kfree(m);
1818}
1819
1820/*
1821 * Handle RCU callback
1822 */
1823static void pio_map_rcu_callback(struct rcu_head *list)
1824{
1825	struct pio_vl_map *m = container_of(list, struct pio_vl_map, list);
1826
1827	pio_map_free(m);
1828}
1829
1830/*
1831 * Set credit return threshold for the kernel send context
1832 */
1833static void set_threshold(struct hfi1_devdata *dd, int scontext, int i)
1834{
1835	u32 thres;
1836
1837	thres = min(sc_percent_to_threshold(dd->kernel_send_context[scontext],
1838					    50),
1839		    sc_mtu_to_threshold(dd->kernel_send_context[scontext],
1840					dd->vld[i].mtu,
1841					dd->rcd[0]->rcvhdrqentsize));
1842	sc_set_cr_threshold(dd->kernel_send_context[scontext], thres);
1843}
1844
1845/*
1846 * pio_map_init - called when #vls change
1847 * @dd: hfi1_devdata
1848 * @port: port number
1849 * @num_vls: number of vls
1850 * @vl_scontexts: per vl send context mapping (optional)
1851 *
1852 * This routine changes the mapping based on the number of vls.
1853 *
1854 * vl_scontexts is used to specify a non-uniform vl/send context
1855 * loading. NULL implies auto computing the loading and giving each
1856 * VL an uniform distribution of send contexts per VL.
1857 *
1858 * The auto algorithm computers the sc_per_vl and the number of extra
1859 * send contexts. Any extra send contexts are added from the last VL
1860 * on down
1861 *
1862 * rcu locking is used here to control access to the mapping fields.
1863 *
1864 * If either the num_vls or num_send_contexts are non-power of 2, the
1865 * array sizes in the struct pio_vl_map and the struct pio_map_elem are
1866 * rounded up to the next highest power of 2 and the first entry is
1867 * reused in a round robin fashion.
1868 *
1869 * If an error occurs the map change is not done and the mapping is not
1870 * chaged.
1871 *
1872 */
1873int pio_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_scontexts)
1874{
1875	int i, j;
1876	int extra, sc_per_vl;
1877	int scontext = 1;
1878	int num_kernel_send_contexts = 0;
1879	u8 lvl_scontexts[OPA_MAX_VLS];
1880	struct pio_vl_map *oldmap, *newmap;
1881
1882	if (!vl_scontexts) {
1883		for (i = 0; i < dd->num_send_contexts; i++)
1884			if (dd->send_contexts[i].type == SC_KERNEL)
1885				num_kernel_send_contexts++;
1886		/* truncate divide */
1887		sc_per_vl = num_kernel_send_contexts / num_vls;
1888		/* extras */
1889		extra = num_kernel_send_contexts % num_vls;
1890		vl_scontexts = lvl_scontexts;
1891		/* add extras from last vl down */
1892		for (i = num_vls - 1; i >= 0; i--, extra--)
1893			vl_scontexts[i] = sc_per_vl + (extra > 0 ? 1 : 0);
1894	}
1895	/* build new map */
1896	newmap = kzalloc(struct_size(newmap, map, roundup_pow_of_two(num_vls)),
 
 
1897			 GFP_KERNEL);
1898	if (!newmap)
1899		goto bail;
1900	newmap->actual_vls = num_vls;
1901	newmap->vls = roundup_pow_of_two(num_vls);
1902	newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1903	for (i = 0; i < newmap->vls; i++) {
1904		/* save for wrap around */
1905		int first_scontext = scontext;
1906
1907		if (i < newmap->actual_vls) {
1908			int sz = roundup_pow_of_two(vl_scontexts[i]);
1909
1910			/* only allocate once */
1911			newmap->map[i] = kzalloc(struct_size(newmap->map[i],
1912							     ksc, sz),
 
1913						 GFP_KERNEL);
1914			if (!newmap->map[i])
1915				goto bail;
1916			newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1917			/*
1918			 * assign send contexts and
1919			 * adjust credit return threshold
1920			 */
1921			for (j = 0; j < sz; j++) {
1922				if (dd->kernel_send_context[scontext]) {
1923					newmap->map[i]->ksc[j] =
1924					dd->kernel_send_context[scontext];
1925					set_threshold(dd, scontext, i);
1926				}
1927				if (++scontext >= first_scontext +
1928						  vl_scontexts[i])
1929					/* wrap back to first send context */
1930					scontext = first_scontext;
1931			}
1932		} else {
1933			/* just re-use entry without allocating */
1934			newmap->map[i] = newmap->map[i % num_vls];
1935		}
1936		scontext = first_scontext + vl_scontexts[i];
1937	}
1938	/* newmap in hand, save old map */
1939	spin_lock_irq(&dd->pio_map_lock);
1940	oldmap = rcu_dereference_protected(dd->pio_map,
1941					   lockdep_is_held(&dd->pio_map_lock));
1942
1943	/* publish newmap */
1944	rcu_assign_pointer(dd->pio_map, newmap);
1945
1946	spin_unlock_irq(&dd->pio_map_lock);
1947	/* success, free any old map after grace period */
1948	if (oldmap)
1949		call_rcu(&oldmap->list, pio_map_rcu_callback);
1950	return 0;
1951bail:
1952	/* free any partial allocation */
1953	pio_map_free(newmap);
1954	return -ENOMEM;
1955}
1956
1957void free_pio_map(struct hfi1_devdata *dd)
1958{
1959	/* Free PIO map if allocated */
1960	if (rcu_access_pointer(dd->pio_map)) {
1961		spin_lock_irq(&dd->pio_map_lock);
1962		pio_map_free(rcu_access_pointer(dd->pio_map));
1963		RCU_INIT_POINTER(dd->pio_map, NULL);
1964		spin_unlock_irq(&dd->pio_map_lock);
1965		synchronize_rcu();
1966	}
1967	kfree(dd->kernel_send_context);
1968	dd->kernel_send_context = NULL;
1969}
1970
1971int init_pervl_scs(struct hfi1_devdata *dd)
1972{
1973	int i;
1974	u64 mask, all_vl_mask = (u64)0x80ff; /* VLs 0-7, 15 */
1975	u64 data_vls_mask = (u64)0x00ff; /* VLs 0-7 */
1976	u32 ctxt;
1977	struct hfi1_pportdata *ppd = dd->pport;
1978
1979	dd->vld[15].sc = sc_alloc(dd, SC_VL15,
1980				  dd->rcd[0]->rcvhdrqentsize, dd->node);
1981	if (!dd->vld[15].sc)
1982		return -ENOMEM;
1983
1984	hfi1_init_ctxt(dd->vld[15].sc);
1985	dd->vld[15].mtu = enum_to_mtu(OPA_MTU_2048);
1986
1987	dd->kernel_send_context = kcalloc_node(dd->num_send_contexts,
1988					       sizeof(struct send_context *),
1989					       GFP_KERNEL, dd->node);
1990	if (!dd->kernel_send_context)
1991		goto freesc15;
1992
1993	dd->kernel_send_context[0] = dd->vld[15].sc;
1994
1995	for (i = 0; i < num_vls; i++) {
1996		/*
1997		 * Since this function does not deal with a specific
1998		 * receive context but we need the RcvHdrQ entry size,
1999		 * use the size from rcd[0]. It is guaranteed to be
2000		 * valid at this point and will remain the same for all
2001		 * receive contexts.
2002		 */
2003		dd->vld[i].sc = sc_alloc(dd, SC_KERNEL,
2004					 dd->rcd[0]->rcvhdrqentsize, dd->node);
2005		if (!dd->vld[i].sc)
2006			goto nomem;
2007		dd->kernel_send_context[i + 1] = dd->vld[i].sc;
2008		hfi1_init_ctxt(dd->vld[i].sc);
2009		/* non VL15 start with the max MTU */
2010		dd->vld[i].mtu = hfi1_max_mtu;
2011	}
2012	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++) {
2013		dd->kernel_send_context[i + 1] =
2014		sc_alloc(dd, SC_KERNEL, dd->rcd[0]->rcvhdrqentsize, dd->node);
2015		if (!dd->kernel_send_context[i + 1])
2016			goto nomem;
2017		hfi1_init_ctxt(dd->kernel_send_context[i + 1]);
2018	}
2019
2020	sc_enable(dd->vld[15].sc);
2021	ctxt = dd->vld[15].sc->hw_context;
2022	mask = all_vl_mask & ~(1LL << 15);
2023	write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2024	dd_dev_info(dd,
2025		    "Using send context %u(%u) for VL15\n",
2026		    dd->vld[15].sc->sw_index, ctxt);
2027
2028	for (i = 0; i < num_vls; i++) {
2029		sc_enable(dd->vld[i].sc);
2030		ctxt = dd->vld[i].sc->hw_context;
2031		mask = all_vl_mask & ~(data_vls_mask);
2032		write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2033	}
2034	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++) {
2035		sc_enable(dd->kernel_send_context[i + 1]);
2036		ctxt = dd->kernel_send_context[i + 1]->hw_context;
2037		mask = all_vl_mask & ~(data_vls_mask);
2038		write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2039	}
2040
2041	if (pio_map_init(dd, ppd->port - 1, num_vls, NULL))
2042		goto nomem;
2043	return 0;
2044
2045nomem:
2046	for (i = 0; i < num_vls; i++) {
2047		sc_free(dd->vld[i].sc);
2048		dd->vld[i].sc = NULL;
2049	}
2050
2051	for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++)
2052		sc_free(dd->kernel_send_context[i + 1]);
2053
2054	kfree(dd->kernel_send_context);
2055	dd->kernel_send_context = NULL;
2056
2057freesc15:
2058	sc_free(dd->vld[15].sc);
2059	return -ENOMEM;
2060}
2061
2062int init_credit_return(struct hfi1_devdata *dd)
2063{
2064	int ret;
2065	int i;
2066
2067	dd->cr_base = kcalloc(
2068		node_affinity.num_possible_nodes,
2069		sizeof(struct credit_return_base),
2070		GFP_KERNEL);
2071	if (!dd->cr_base) {
2072		ret = -ENOMEM;
2073		goto done;
2074	}
2075	for_each_node_with_cpus(i) {
2076		int bytes = TXE_NUM_CONTEXTS * sizeof(struct credit_return);
2077
2078		set_dev_node(&dd->pcidev->dev, i);
2079		dd->cr_base[i].va = dma_alloc_coherent(&dd->pcidev->dev,
2080						       bytes,
2081						       &dd->cr_base[i].dma,
2082						       GFP_KERNEL);
2083		if (!dd->cr_base[i].va) {
2084			set_dev_node(&dd->pcidev->dev, dd->node);
2085			dd_dev_err(dd,
2086				   "Unable to allocate credit return DMA range for NUMA %d\n",
2087				   i);
2088			ret = -ENOMEM;
2089			goto free_cr_base;
2090		}
2091	}
2092	set_dev_node(&dd->pcidev->dev, dd->node);
2093
2094	ret = 0;
2095done:
2096	return ret;
2097
2098free_cr_base:
2099	free_credit_return(dd);
2100	goto done;
2101}
2102
2103void free_credit_return(struct hfi1_devdata *dd)
2104{
2105	int i;
2106
2107	if (!dd->cr_base)
2108		return;
2109	for (i = 0; i < node_affinity.num_possible_nodes; i++) {
2110		if (dd->cr_base[i].va) {
2111			dma_free_coherent(&dd->pcidev->dev,
2112					  TXE_NUM_CONTEXTS *
2113					  sizeof(struct credit_return),
2114					  dd->cr_base[i].va,
2115					  dd->cr_base[i].dma);
2116		}
2117	}
2118	kfree(dd->cr_base);
2119	dd->cr_base = NULL;
2120}
2121
2122void seqfile_dump_sci(struct seq_file *s, u32 i,
2123		      struct send_context_info *sci)
2124{
2125	struct send_context *sc = sci->sc;
2126	u64 reg;
2127
2128	seq_printf(s, "SCI %u: type %u base %u credits %u\n",
2129		   i, sci->type, sci->base, sci->credits);
2130	seq_printf(s, "  flags 0x%x sw_inx %u hw_ctxt %u grp %u\n",
2131		   sc->flags,  sc->sw_index, sc->hw_context, sc->group);
2132	seq_printf(s, "  sr_size %u credits %u sr_head %u sr_tail %u\n",
2133		   sc->sr_size, sc->credits, sc->sr_head, sc->sr_tail);
2134	seq_printf(s, "  fill %lu free %lu fill_wrap %u alloc_free %lu\n",
2135		   sc->fill, sc->free, sc->fill_wrap, sc->alloc_free);
2136	seq_printf(s, "  credit_intr_count %u credit_ctrl 0x%llx\n",
2137		   sc->credit_intr_count, sc->credit_ctrl);
2138	reg = read_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_STATUS));
2139	seq_printf(s, "  *hw_free %llu CurrentFree %llu LastReturned %llu\n",
2140		   (le64_to_cpu(*sc->hw_free) & CR_COUNTER_SMASK) >>
2141		    CR_COUNTER_SHIFT,
2142		   (reg >> SC(CREDIT_STATUS_CURRENT_FREE_COUNTER_SHIFT)) &
2143		    SC(CREDIT_STATUS_CURRENT_FREE_COUNTER_MASK),
2144		   reg & SC(CREDIT_STATUS_LAST_RETURNED_COUNTER_SMASK));
2145}