Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
   2/*
   3 * Copyright(c) 2015 - 2020 Intel Corporation.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   4 */
   5
   6#include <linux/topology.h>
   7#include <linux/cpumask.h>
 
   8#include <linux/interrupt.h>
   9#include <linux/numa.h>
  10
  11#include "hfi.h"
  12#include "affinity.h"
  13#include "sdma.h"
  14#include "trace.h"
  15
  16struct hfi1_affinity_node_list node_affinity = {
  17	.list = LIST_HEAD_INIT(node_affinity.list),
  18	.lock = __MUTEX_INITIALIZER(node_affinity.lock)
  19};
  20
  21/* Name of IRQ types, indexed by enum irq_type */
  22static const char * const irq_type_names[] = {
  23	"SDMA",
  24	"RCVCTXT",
  25	"NETDEVCTXT",
  26	"GENERAL",
  27	"OTHER",
  28};
  29
  30/* Per NUMA node count of HFI devices */
  31static unsigned int *hfi1_per_node_cntr;
  32
  33static inline void init_cpu_mask_set(struct cpu_mask_set *set)
  34{
  35	cpumask_clear(&set->mask);
  36	cpumask_clear(&set->used);
  37	set->gen = 0;
  38}
  39
  40/* Increment generation of CPU set if needed */
  41static void _cpu_mask_set_gen_inc(struct cpu_mask_set *set)
  42{
  43	if (cpumask_equal(&set->mask, &set->used)) {
  44		/*
  45		 * We've used up all the CPUs, bump up the generation
  46		 * and reset the 'used' map
  47		 */
  48		set->gen++;
  49		cpumask_clear(&set->used);
  50	}
  51}
  52
  53static void _cpu_mask_set_gen_dec(struct cpu_mask_set *set)
  54{
  55	if (cpumask_empty(&set->used) && set->gen) {
  56		set->gen--;
  57		cpumask_copy(&set->used, &set->mask);
  58	}
  59}
  60
  61/* Get the first CPU from the list of unused CPUs in a CPU set data structure */
  62static int cpu_mask_set_get_first(struct cpu_mask_set *set, cpumask_var_t diff)
  63{
  64	int cpu;
  65
  66	if (!diff || !set)
  67		return -EINVAL;
  68
  69	_cpu_mask_set_gen_inc(set);
  70
  71	/* Find out CPUs left in CPU mask */
  72	cpumask_andnot(diff, &set->mask, &set->used);
  73
  74	cpu = cpumask_first(diff);
  75	if (cpu >= nr_cpu_ids) /* empty */
  76		cpu = -EINVAL;
  77	else
  78		cpumask_set_cpu(cpu, &set->used);
  79
  80	return cpu;
  81}
  82
  83static void cpu_mask_set_put(struct cpu_mask_set *set, int cpu)
  84{
  85	if (!set)
  86		return;
  87
  88	cpumask_clear_cpu(cpu, &set->used);
  89	_cpu_mask_set_gen_dec(set);
  90}
  91
  92/* Initialize non-HT cpu cores mask */
  93void init_real_cpu_mask(void)
  94{
  95	int possible, curr_cpu, i, ht;
  96
  97	cpumask_clear(&node_affinity.real_cpu_mask);
  98
  99	/* Start with cpu online mask as the real cpu mask */
 100	cpumask_copy(&node_affinity.real_cpu_mask, cpu_online_mask);
 101
 102	/*
 103	 * Remove HT cores from the real cpu mask.  Do this in two steps below.
 104	 */
 105	possible = cpumask_weight(&node_affinity.real_cpu_mask);
 106	ht = cpumask_weight(topology_sibling_cpumask(
 107				cpumask_first(&node_affinity.real_cpu_mask)));
 108	/*
 109	 * Step 1.  Skip over the first N HT siblings and use them as the
 110	 * "real" cores.  Assumes that HT cores are not enumerated in
 111	 * succession (except in the single core case).
 112	 */
 113	curr_cpu = cpumask_first(&node_affinity.real_cpu_mask);
 114	for (i = 0; i < possible / ht; i++)
 115		curr_cpu = cpumask_next(curr_cpu, &node_affinity.real_cpu_mask);
 116	/*
 117	 * Step 2.  Remove the remaining HT siblings.  Use cpumask_next() to
 118	 * skip any gaps.
 119	 */
 120	for (; i < possible; i++) {
 121		cpumask_clear_cpu(curr_cpu, &node_affinity.real_cpu_mask);
 122		curr_cpu = cpumask_next(curr_cpu, &node_affinity.real_cpu_mask);
 123	}
 124}
 125
 126int node_affinity_init(void)
 127{
 128	int node;
 129	struct pci_dev *dev = NULL;
 130	const struct pci_device_id *ids = hfi1_pci_tbl;
 131
 132	cpumask_clear(&node_affinity.proc.used);
 133	cpumask_copy(&node_affinity.proc.mask, cpu_online_mask);
 134
 135	node_affinity.proc.gen = 0;
 136	node_affinity.num_core_siblings =
 137				cpumask_weight(topology_sibling_cpumask(
 138					cpumask_first(&node_affinity.proc.mask)
 139					));
 140	node_affinity.num_possible_nodes = num_possible_nodes();
 141	node_affinity.num_online_nodes = num_online_nodes();
 142	node_affinity.num_online_cpus = num_online_cpus();
 143
 144	/*
 145	 * The real cpu mask is part of the affinity struct but it has to be
 146	 * initialized early. It is needed to calculate the number of user
 147	 * contexts in set_up_context_variables().
 148	 */
 149	init_real_cpu_mask();
 150
 151	hfi1_per_node_cntr = kcalloc(node_affinity.num_possible_nodes,
 152				     sizeof(*hfi1_per_node_cntr), GFP_KERNEL);
 153	if (!hfi1_per_node_cntr)
 154		return -ENOMEM;
 155
 156	while (ids->vendor) {
 157		dev = NULL;
 158		while ((dev = pci_get_device(ids->vendor, ids->device, dev))) {
 159			node = pcibus_to_node(dev->bus);
 160			if (node < 0)
 161				goto out;
 162
 163			hfi1_per_node_cntr[node]++;
 164		}
 165		ids++;
 166	}
 167
 168	return 0;
 169
 170out:
 171	/*
 172	 * Invalid PCI NUMA node information found, note it, and populate
 173	 * our database 1:1.
 174	 */
 175	pr_err("HFI: Invalid PCI NUMA node. Performance may be affected\n");
 176	pr_err("HFI: System BIOS may need to be upgraded\n");
 177	for (node = 0; node < node_affinity.num_possible_nodes; node++)
 178		hfi1_per_node_cntr[node] = 1;
 179
 180	pci_dev_put(dev);
 181
 182	return 0;
 183}
 184
 185static void node_affinity_destroy(struct hfi1_affinity_node *entry)
 186{
 187	free_percpu(entry->comp_vect_affinity);
 188	kfree(entry);
 189}
 190
 191void node_affinity_destroy_all(void)
 192{
 193	struct list_head *pos, *q;
 194	struct hfi1_affinity_node *entry;
 195
 196	mutex_lock(&node_affinity.lock);
 197	list_for_each_safe(pos, q, &node_affinity.list) {
 198		entry = list_entry(pos, struct hfi1_affinity_node,
 199				   list);
 200		list_del(pos);
 201		node_affinity_destroy(entry);
 202	}
 203	mutex_unlock(&node_affinity.lock);
 204	kfree(hfi1_per_node_cntr);
 205}
 206
 207static struct hfi1_affinity_node *node_affinity_allocate(int node)
 208{
 209	struct hfi1_affinity_node *entry;
 210
 211	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 212	if (!entry)
 213		return NULL;
 214	entry->node = node;
 215	entry->comp_vect_affinity = alloc_percpu(u16);
 216	INIT_LIST_HEAD(&entry->list);
 217
 218	return entry;
 219}
 220
 221/*
 222 * It appends an entry to the list.
 223 * It *must* be called with node_affinity.lock held.
 224 */
 225static void node_affinity_add_tail(struct hfi1_affinity_node *entry)
 226{
 227	list_add_tail(&entry->list, &node_affinity.list);
 228}
 229
 230/* It must be called with node_affinity.lock held */
 231static struct hfi1_affinity_node *node_affinity_lookup(int node)
 232{
 
 233	struct hfi1_affinity_node *entry;
 234
 235	list_for_each_entry(entry, &node_affinity.list, list) {
 
 236		if (entry->node == node)
 237			return entry;
 238	}
 239
 240	return NULL;
 241}
 242
 243static int per_cpu_affinity_get(cpumask_var_t possible_cpumask,
 244				u16 __percpu *comp_vect_affinity)
 245{
 246	int curr_cpu;
 247	u16 cntr;
 248	u16 prev_cntr;
 249	int ret_cpu;
 250
 251	if (!possible_cpumask) {
 252		ret_cpu = -EINVAL;
 253		goto fail;
 254	}
 255
 256	if (!comp_vect_affinity) {
 257		ret_cpu = -EINVAL;
 258		goto fail;
 259	}
 260
 261	ret_cpu = cpumask_first(possible_cpumask);
 262	if (ret_cpu >= nr_cpu_ids) {
 263		ret_cpu = -EINVAL;
 264		goto fail;
 265	}
 266
 267	prev_cntr = *per_cpu_ptr(comp_vect_affinity, ret_cpu);
 268	for_each_cpu(curr_cpu, possible_cpumask) {
 269		cntr = *per_cpu_ptr(comp_vect_affinity, curr_cpu);
 270
 271		if (cntr < prev_cntr) {
 272			ret_cpu = curr_cpu;
 273			prev_cntr = cntr;
 274		}
 275	}
 276
 277	*per_cpu_ptr(comp_vect_affinity, ret_cpu) += 1;
 278
 279fail:
 280	return ret_cpu;
 281}
 282
 283static int per_cpu_affinity_put_max(cpumask_var_t possible_cpumask,
 284				    u16 __percpu *comp_vect_affinity)
 285{
 286	int curr_cpu;
 287	int max_cpu;
 288	u16 cntr;
 289	u16 prev_cntr;
 290
 291	if (!possible_cpumask)
 292		return -EINVAL;
 293
 294	if (!comp_vect_affinity)
 295		return -EINVAL;
 296
 297	max_cpu = cpumask_first(possible_cpumask);
 298	if (max_cpu >= nr_cpu_ids)
 299		return -EINVAL;
 300
 301	prev_cntr = *per_cpu_ptr(comp_vect_affinity, max_cpu);
 302	for_each_cpu(curr_cpu, possible_cpumask) {
 303		cntr = *per_cpu_ptr(comp_vect_affinity, curr_cpu);
 304
 305		if (cntr > prev_cntr) {
 306			max_cpu = curr_cpu;
 307			prev_cntr = cntr;
 308		}
 309	}
 310
 311	*per_cpu_ptr(comp_vect_affinity, max_cpu) -= 1;
 312
 313	return max_cpu;
 314}
 315
 316/*
 317 * Non-interrupt CPUs are used first, then interrupt CPUs.
 318 * Two already allocated cpu masks must be passed.
 319 */
 320static int _dev_comp_vect_cpu_get(struct hfi1_devdata *dd,
 321				  struct hfi1_affinity_node *entry,
 322				  cpumask_var_t non_intr_cpus,
 323				  cpumask_var_t available_cpus)
 324	__must_hold(&node_affinity.lock)
 325{
 326	int cpu;
 327	struct cpu_mask_set *set = dd->comp_vect;
 328
 329	lockdep_assert_held(&node_affinity.lock);
 330	if (!non_intr_cpus) {
 331		cpu = -1;
 332		goto fail;
 333	}
 334
 335	if (!available_cpus) {
 336		cpu = -1;
 337		goto fail;
 338	}
 339
 340	/* Available CPUs for pinning completion vectors */
 341	_cpu_mask_set_gen_inc(set);
 342	cpumask_andnot(available_cpus, &set->mask, &set->used);
 343
 344	/* Available CPUs without SDMA engine interrupts */
 345	cpumask_andnot(non_intr_cpus, available_cpus,
 346		       &entry->def_intr.used);
 347
 348	/* If there are non-interrupt CPUs available, use them first */
 349	if (!cpumask_empty(non_intr_cpus))
 350		cpu = cpumask_first(non_intr_cpus);
 351	else /* Otherwise, use interrupt CPUs */
 352		cpu = cpumask_first(available_cpus);
 353
 354	if (cpu >= nr_cpu_ids) { /* empty */
 355		cpu = -1;
 356		goto fail;
 357	}
 358	cpumask_set_cpu(cpu, &set->used);
 359
 360fail:
 361	return cpu;
 362}
 363
 364static void _dev_comp_vect_cpu_put(struct hfi1_devdata *dd, int cpu)
 365{
 366	struct cpu_mask_set *set = dd->comp_vect;
 367
 368	if (cpu < 0)
 369		return;
 370
 371	cpu_mask_set_put(set, cpu);
 372}
 373
 374/* _dev_comp_vect_mappings_destroy() is reentrant */
 375static void _dev_comp_vect_mappings_destroy(struct hfi1_devdata *dd)
 376{
 377	int i, cpu;
 378
 379	if (!dd->comp_vect_mappings)
 380		return;
 381
 382	for (i = 0; i < dd->comp_vect_possible_cpus; i++) {
 383		cpu = dd->comp_vect_mappings[i];
 384		_dev_comp_vect_cpu_put(dd, cpu);
 385		dd->comp_vect_mappings[i] = -1;
 386		hfi1_cdbg(AFFINITY,
 387			  "[%s] Release CPU %d from completion vector %d",
 388			  rvt_get_ibdev_name(&(dd)->verbs_dev.rdi), cpu, i);
 389	}
 390
 391	kfree(dd->comp_vect_mappings);
 392	dd->comp_vect_mappings = NULL;
 393}
 394
 395/*
 396 * This function creates the table for looking up CPUs for completion vectors.
 397 * num_comp_vectors needs to have been initilized before calling this function.
 398 */
 399static int _dev_comp_vect_mappings_create(struct hfi1_devdata *dd,
 400					  struct hfi1_affinity_node *entry)
 401	__must_hold(&node_affinity.lock)
 402{
 403	int i, cpu, ret;
 404	cpumask_var_t non_intr_cpus;
 405	cpumask_var_t available_cpus;
 406
 407	lockdep_assert_held(&node_affinity.lock);
 408
 409	if (!zalloc_cpumask_var(&non_intr_cpus, GFP_KERNEL))
 410		return -ENOMEM;
 411
 412	if (!zalloc_cpumask_var(&available_cpus, GFP_KERNEL)) {
 413		free_cpumask_var(non_intr_cpus);
 414		return -ENOMEM;
 415	}
 416
 417	dd->comp_vect_mappings = kcalloc(dd->comp_vect_possible_cpus,
 418					 sizeof(*dd->comp_vect_mappings),
 419					 GFP_KERNEL);
 420	if (!dd->comp_vect_mappings) {
 421		ret = -ENOMEM;
 422		goto fail;
 423	}
 424	for (i = 0; i < dd->comp_vect_possible_cpus; i++)
 425		dd->comp_vect_mappings[i] = -1;
 426
 427	for (i = 0; i < dd->comp_vect_possible_cpus; i++) {
 428		cpu = _dev_comp_vect_cpu_get(dd, entry, non_intr_cpus,
 429					     available_cpus);
 430		if (cpu < 0) {
 431			ret = -EINVAL;
 432			goto fail;
 433		}
 434
 435		dd->comp_vect_mappings[i] = cpu;
 436		hfi1_cdbg(AFFINITY,
 437			  "[%s] Completion Vector %d -> CPU %d",
 438			  rvt_get_ibdev_name(&(dd)->verbs_dev.rdi), i, cpu);
 439	}
 440
 441	free_cpumask_var(available_cpus);
 442	free_cpumask_var(non_intr_cpus);
 443	return 0;
 444
 445fail:
 446	free_cpumask_var(available_cpus);
 447	free_cpumask_var(non_intr_cpus);
 448	_dev_comp_vect_mappings_destroy(dd);
 449
 450	return ret;
 451}
 452
 453int hfi1_comp_vectors_set_up(struct hfi1_devdata *dd)
 454{
 455	int ret;
 456	struct hfi1_affinity_node *entry;
 457
 458	mutex_lock(&node_affinity.lock);
 459	entry = node_affinity_lookup(dd->node);
 460	if (!entry) {
 461		ret = -EINVAL;
 462		goto unlock;
 463	}
 464	ret = _dev_comp_vect_mappings_create(dd, entry);
 465unlock:
 466	mutex_unlock(&node_affinity.lock);
 467
 468	return ret;
 469}
 470
 471void hfi1_comp_vectors_clean_up(struct hfi1_devdata *dd)
 472{
 473	_dev_comp_vect_mappings_destroy(dd);
 474}
 475
 476int hfi1_comp_vect_mappings_lookup(struct rvt_dev_info *rdi, int comp_vect)
 477{
 478	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
 479	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
 480
 481	if (!dd->comp_vect_mappings)
 482		return -EINVAL;
 483	if (comp_vect >= dd->comp_vect_possible_cpus)
 484		return -EINVAL;
 485
 486	return dd->comp_vect_mappings[comp_vect];
 487}
 488
 489/*
 490 * It assumes dd->comp_vect_possible_cpus is available.
 491 */
 492static int _dev_comp_vect_cpu_mask_init(struct hfi1_devdata *dd,
 493					struct hfi1_affinity_node *entry,
 494					bool first_dev_init)
 495	__must_hold(&node_affinity.lock)
 496{
 497	int i, j, curr_cpu;
 498	int possible_cpus_comp_vect = 0;
 499	struct cpumask *dev_comp_vect_mask = &dd->comp_vect->mask;
 500
 501	lockdep_assert_held(&node_affinity.lock);
 502	/*
 503	 * If there's only one CPU available for completion vectors, then
 504	 * there will only be one completion vector available. Othewise,
 505	 * the number of completion vector available will be the number of
 506	 * available CPUs divide it by the number of devices in the
 507	 * local NUMA node.
 508	 */
 509	if (cpumask_weight(&entry->comp_vect_mask) == 1) {
 510		possible_cpus_comp_vect = 1;
 511		dd_dev_warn(dd,
 512			    "Number of kernel receive queues is too large for completion vector affinity to be effective\n");
 513	} else {
 514		possible_cpus_comp_vect +=
 515			cpumask_weight(&entry->comp_vect_mask) /
 516				       hfi1_per_node_cntr[dd->node];
 517
 518		/*
 519		 * If the completion vector CPUs available doesn't divide
 520		 * evenly among devices, then the first device device to be
 521		 * initialized gets an extra CPU.
 522		 */
 523		if (first_dev_init &&
 524		    cpumask_weight(&entry->comp_vect_mask) %
 525		    hfi1_per_node_cntr[dd->node] != 0)
 526			possible_cpus_comp_vect++;
 527	}
 528
 529	dd->comp_vect_possible_cpus = possible_cpus_comp_vect;
 530
 531	/* Reserving CPUs for device completion vector */
 532	for (i = 0; i < dd->comp_vect_possible_cpus; i++) {
 533		curr_cpu = per_cpu_affinity_get(&entry->comp_vect_mask,
 534						entry->comp_vect_affinity);
 535		if (curr_cpu < 0)
 536			goto fail;
 537
 538		cpumask_set_cpu(curr_cpu, dev_comp_vect_mask);
 539	}
 540
 541	hfi1_cdbg(AFFINITY,
 542		  "[%s] Completion vector affinity CPU set(s) %*pbl",
 543		  rvt_get_ibdev_name(&(dd)->verbs_dev.rdi),
 544		  cpumask_pr_args(dev_comp_vect_mask));
 545
 546	return 0;
 547
 548fail:
 549	for (j = 0; j < i; j++)
 550		per_cpu_affinity_put_max(&entry->comp_vect_mask,
 551					 entry->comp_vect_affinity);
 552
 553	return curr_cpu;
 554}
 555
 556/*
 557 * It assumes dd->comp_vect_possible_cpus is available.
 558 */
 559static void _dev_comp_vect_cpu_mask_clean_up(struct hfi1_devdata *dd,
 560					     struct hfi1_affinity_node *entry)
 561	__must_hold(&node_affinity.lock)
 562{
 563	int i, cpu;
 564
 565	lockdep_assert_held(&node_affinity.lock);
 566	if (!dd->comp_vect_possible_cpus)
 567		return;
 568
 569	for (i = 0; i < dd->comp_vect_possible_cpus; i++) {
 570		cpu = per_cpu_affinity_put_max(&dd->comp_vect->mask,
 571					       entry->comp_vect_affinity);
 572		/* Clearing CPU in device completion vector cpu mask */
 573		if (cpu >= 0)
 574			cpumask_clear_cpu(cpu, &dd->comp_vect->mask);
 575	}
 576
 577	dd->comp_vect_possible_cpus = 0;
 578}
 579
 580/*
 581 * Interrupt affinity.
 582 *
 583 * non-rcv avail gets a default mask that
 584 * starts as possible cpus with threads reset
 585 * and each rcv avail reset.
 586 *
 587 * rcv avail gets node relative 1 wrapping back
 588 * to the node relative 1 as necessary.
 589 *
 590 */
 591int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
 592{
 
 593	struct hfi1_affinity_node *entry;
 594	const struct cpumask *local_mask;
 595	int curr_cpu, possible, i, ret;
 596	bool new_entry = false;
 597
 
 
 
 
 
 
 
 
 
 
 598	local_mask = cpumask_of_node(dd->node);
 599	if (cpumask_first(local_mask) >= nr_cpu_ids)
 600		local_mask = topology_core_cpumask(0);
 601
 602	mutex_lock(&node_affinity.lock);
 603	entry = node_affinity_lookup(dd->node);
 604
 605	/*
 606	 * If this is the first time this NUMA node's affinity is used,
 607	 * create an entry in the global affinity structure and initialize it.
 608	 */
 609	if (!entry) {
 610		entry = node_affinity_allocate(dd->node);
 611		if (!entry) {
 612			dd_dev_err(dd,
 613				   "Unable to allocate global affinity node\n");
 614			ret = -ENOMEM;
 615			goto fail;
 616		}
 617		new_entry = true;
 618
 619		init_cpu_mask_set(&entry->def_intr);
 620		init_cpu_mask_set(&entry->rcv_intr);
 621		cpumask_clear(&entry->comp_vect_mask);
 622		cpumask_clear(&entry->general_intr_mask);
 623		/* Use the "real" cpu mask of this node as the default */
 624		cpumask_and(&entry->def_intr.mask, &node_affinity.real_cpu_mask,
 625			    local_mask);
 626
 627		/* fill in the receive list */
 628		possible = cpumask_weight(&entry->def_intr.mask);
 629		curr_cpu = cpumask_first(&entry->def_intr.mask);
 630
 631		if (possible == 1) {
 632			/* only one CPU, everyone will use it */
 633			cpumask_set_cpu(curr_cpu, &entry->rcv_intr.mask);
 634			cpumask_set_cpu(curr_cpu, &entry->general_intr_mask);
 635		} else {
 636			/*
 637			 * The general/control context will be the first CPU in
 638			 * the default list, so it is removed from the default
 639			 * list and added to the general interrupt list.
 640			 */
 641			cpumask_clear_cpu(curr_cpu, &entry->def_intr.mask);
 642			cpumask_set_cpu(curr_cpu, &entry->general_intr_mask);
 643			curr_cpu = cpumask_next(curr_cpu,
 644						&entry->def_intr.mask);
 645
 646			/*
 647			 * Remove the remaining kernel receive queues from
 648			 * the default list and add them to the receive list.
 649			 */
 650			for (i = 0;
 651			     i < (dd->n_krcv_queues - 1) *
 652				  hfi1_per_node_cntr[dd->node];
 653			     i++) {
 654				cpumask_clear_cpu(curr_cpu,
 655						  &entry->def_intr.mask);
 656				cpumask_set_cpu(curr_cpu,
 657						&entry->rcv_intr.mask);
 658				curr_cpu = cpumask_next(curr_cpu,
 659							&entry->def_intr.mask);
 660				if (curr_cpu >= nr_cpu_ids)
 661					break;
 662			}
 663
 664			/*
 665			 * If there ends up being 0 CPU cores leftover for SDMA
 666			 * engines, use the same CPU cores as general/control
 667			 * context.
 668			 */
 669			if (cpumask_empty(&entry->def_intr.mask))
 670				cpumask_copy(&entry->def_intr.mask,
 671					     &entry->general_intr_mask);
 672		}
 673
 674		/* Determine completion vector CPUs for the entire node */
 675		cpumask_and(&entry->comp_vect_mask,
 676			    &node_affinity.real_cpu_mask, local_mask);
 677		cpumask_andnot(&entry->comp_vect_mask,
 678			       &entry->comp_vect_mask,
 679			       &entry->rcv_intr.mask);
 680		cpumask_andnot(&entry->comp_vect_mask,
 681			       &entry->comp_vect_mask,
 682			       &entry->general_intr_mask);
 683
 684		/*
 685		 * If there ends up being 0 CPU cores leftover for completion
 686		 * vectors, use the same CPU core as the general/control
 687		 * context.
 688		 */
 689		if (cpumask_empty(&entry->comp_vect_mask))
 690			cpumask_copy(&entry->comp_vect_mask,
 691				     &entry->general_intr_mask);
 692	}
 693
 694	ret = _dev_comp_vect_cpu_mask_init(dd, entry, new_entry);
 695	if (ret < 0)
 696		goto fail;
 697
 698	if (new_entry)
 699		node_affinity_add_tail(entry);
 700
 701	dd->affinity_entry = entry;
 702	mutex_unlock(&node_affinity.lock);
 703
 704	return 0;
 705
 706fail:
 707	if (new_entry)
 708		node_affinity_destroy(entry);
 709	mutex_unlock(&node_affinity.lock);
 710	return ret;
 711}
 712
 713void hfi1_dev_affinity_clean_up(struct hfi1_devdata *dd)
 714{
 715	struct hfi1_affinity_node *entry;
 716
 
 
 
 717	mutex_lock(&node_affinity.lock);
 718	if (!dd->affinity_entry)
 719		goto unlock;
 720	entry = node_affinity_lookup(dd->node);
 721	if (!entry)
 722		goto unlock;
 723
 724	/*
 725	 * Free device completion vector CPUs to be used by future
 726	 * completion vectors
 727	 */
 728	_dev_comp_vect_cpu_mask_clean_up(dd, entry);
 729unlock:
 730	dd->affinity_entry = NULL;
 731	mutex_unlock(&node_affinity.lock);
 
 732}
 733
 734/*
 735 * Function updates the irq affinity hint for msix after it has been changed
 736 * by the user using the /proc/irq interface. This function only accepts
 737 * one cpu in the mask.
 738 */
 739static void hfi1_update_sdma_affinity(struct hfi1_msix_entry *msix, int cpu)
 740{
 741	struct sdma_engine *sde = msix->arg;
 742	struct hfi1_devdata *dd = sde->dd;
 743	struct hfi1_affinity_node *entry;
 744	struct cpu_mask_set *set;
 745	int i, old_cpu;
 746
 747	if (cpu > num_online_cpus() || cpu == sde->cpu)
 748		return;
 749
 750	mutex_lock(&node_affinity.lock);
 751	entry = node_affinity_lookup(dd->node);
 752	if (!entry)
 753		goto unlock;
 754
 755	old_cpu = sde->cpu;
 756	sde->cpu = cpu;
 757	cpumask_clear(&msix->mask);
 758	cpumask_set_cpu(cpu, &msix->mask);
 759	dd_dev_dbg(dd, "IRQ: %u, type %s engine %u -> cpu: %d\n",
 760		   msix->irq, irq_type_names[msix->type],
 761		   sde->this_idx, cpu);
 762	irq_set_affinity_hint(msix->irq, &msix->mask);
 763
 764	/*
 765	 * Set the new cpu in the hfi1_affinity_node and clean
 766	 * the old cpu if it is not used by any other IRQ
 767	 */
 768	set = &entry->def_intr;
 769	cpumask_set_cpu(cpu, &set->mask);
 770	cpumask_set_cpu(cpu, &set->used);
 771	for (i = 0; i < dd->msix_info.max_requested; i++) {
 772		struct hfi1_msix_entry *other_msix;
 773
 774		other_msix = &dd->msix_info.msix_entries[i];
 775		if (other_msix->type != IRQ_SDMA || other_msix == msix)
 776			continue;
 777
 778		if (cpumask_test_cpu(old_cpu, &other_msix->mask))
 779			goto unlock;
 780	}
 781	cpumask_clear_cpu(old_cpu, &set->mask);
 782	cpumask_clear_cpu(old_cpu, &set->used);
 783unlock:
 784	mutex_unlock(&node_affinity.lock);
 785}
 786
 787static void hfi1_irq_notifier_notify(struct irq_affinity_notify *notify,
 788				     const cpumask_t *mask)
 789{
 790	int cpu = cpumask_first(mask);
 791	struct hfi1_msix_entry *msix = container_of(notify,
 792						    struct hfi1_msix_entry,
 793						    notify);
 794
 795	/* Only one CPU configuration supported currently */
 796	hfi1_update_sdma_affinity(msix, cpu);
 797}
 798
 799static void hfi1_irq_notifier_release(struct kref *ref)
 800{
 801	/*
 802	 * This is required by affinity notifier. We don't have anything to
 803	 * free here.
 804	 */
 805}
 806
 807static void hfi1_setup_sdma_notifier(struct hfi1_msix_entry *msix)
 808{
 809	struct irq_affinity_notify *notify = &msix->notify;
 810
 811	notify->irq = msix->irq;
 812	notify->notify = hfi1_irq_notifier_notify;
 813	notify->release = hfi1_irq_notifier_release;
 814
 815	if (irq_set_affinity_notifier(notify->irq, notify))
 816		pr_err("Failed to register sdma irq affinity notifier for irq %d\n",
 817		       notify->irq);
 818}
 819
 820static void hfi1_cleanup_sdma_notifier(struct hfi1_msix_entry *msix)
 821{
 822	struct irq_affinity_notify *notify = &msix->notify;
 823
 824	if (irq_set_affinity_notifier(notify->irq, NULL))
 825		pr_err("Failed to cleanup sdma irq affinity notifier for irq %d\n",
 826		       notify->irq);
 827}
 828
 829/*
 830 * Function sets the irq affinity for msix.
 831 * It *must* be called with node_affinity.lock held.
 832 */
 833static int get_irq_affinity(struct hfi1_devdata *dd,
 834			    struct hfi1_msix_entry *msix)
 835{
 836	cpumask_var_t diff;
 837	struct hfi1_affinity_node *entry;
 838	struct cpu_mask_set *set = NULL;
 839	struct sdma_engine *sde = NULL;
 840	struct hfi1_ctxtdata *rcd = NULL;
 841	char extra[64];
 842	int cpu = -1;
 843
 844	extra[0] = '\0';
 845	cpumask_clear(&msix->mask);
 846
 847	entry = node_affinity_lookup(dd->node);
 848
 849	switch (msix->type) {
 850	case IRQ_SDMA:
 851		sde = (struct sdma_engine *)msix->arg;
 852		scnprintf(extra, 64, "engine %u", sde->this_idx);
 853		set = &entry->def_intr;
 854		break;
 855	case IRQ_GENERAL:
 856		cpu = cpumask_first(&entry->general_intr_mask);
 857		break;
 858	case IRQ_RCVCTXT:
 859		rcd = (struct hfi1_ctxtdata *)msix->arg;
 860		if (rcd->ctxt == HFI1_CTRL_CTXT)
 861			cpu = cpumask_first(&entry->general_intr_mask);
 862		else
 863			set = &entry->rcv_intr;
 864		scnprintf(extra, 64, "ctxt %u", rcd->ctxt);
 865		break;
 866	case IRQ_NETDEVCTXT:
 867		rcd = (struct hfi1_ctxtdata *)msix->arg;
 868		set = &entry->def_intr;
 869		scnprintf(extra, 64, "ctxt %u", rcd->ctxt);
 870		break;
 871	default:
 872		dd_dev_err(dd, "Invalid IRQ type %d\n", msix->type);
 873		return -EINVAL;
 874	}
 875
 876	/*
 877	 * The general and control contexts are placed on a particular
 878	 * CPU, which is set above. Skip accounting for it. Everything else
 879	 * finds its CPU here.
 880	 */
 881	if (cpu == -1 && set) {
 882		if (!zalloc_cpumask_var(&diff, GFP_KERNEL))
 883			return -ENOMEM;
 884
 885		cpu = cpu_mask_set_get_first(set, diff);
 886		if (cpu < 0) {
 887			free_cpumask_var(diff);
 888			dd_dev_err(dd, "Failure to obtain CPU for IRQ\n");
 889			return cpu;
 890		}
 891
 892		free_cpumask_var(diff);
 893	}
 894
 895	cpumask_set_cpu(cpu, &msix->mask);
 896	dd_dev_info(dd, "IRQ: %u, type %s %s -> cpu: %d\n",
 897		    msix->irq, irq_type_names[msix->type],
 898		    extra, cpu);
 899	irq_set_affinity_hint(msix->irq, &msix->mask);
 900
 901	if (msix->type == IRQ_SDMA) {
 902		sde->cpu = cpu;
 903		hfi1_setup_sdma_notifier(msix);
 904	}
 905
 906	return 0;
 907}
 908
 909int hfi1_get_irq_affinity(struct hfi1_devdata *dd, struct hfi1_msix_entry *msix)
 910{
 911	int ret;
 912
 913	mutex_lock(&node_affinity.lock);
 914	ret = get_irq_affinity(dd, msix);
 915	mutex_unlock(&node_affinity.lock);
 916	return ret;
 917}
 918
 919void hfi1_put_irq_affinity(struct hfi1_devdata *dd,
 920			   struct hfi1_msix_entry *msix)
 921{
 922	struct cpu_mask_set *set = NULL;
 
 923	struct hfi1_affinity_node *entry;
 924
 925	mutex_lock(&node_affinity.lock);
 926	entry = node_affinity_lookup(dd->node);
 927
 928	switch (msix->type) {
 929	case IRQ_SDMA:
 930		set = &entry->def_intr;
 931		hfi1_cleanup_sdma_notifier(msix);
 932		break;
 933	case IRQ_GENERAL:
 934		/* Don't do accounting for general contexts */
 935		break;
 936	case IRQ_RCVCTXT: {
 937		struct hfi1_ctxtdata *rcd = msix->arg;
 938
 939		/* Don't do accounting for control contexts */
 940		if (rcd->ctxt != HFI1_CTRL_CTXT)
 941			set = &entry->rcv_intr;
 942		break;
 943	}
 944	case IRQ_NETDEVCTXT:
 
 945		set = &entry->def_intr;
 946		break;
 947	default:
 948		mutex_unlock(&node_affinity.lock);
 949		return;
 950	}
 951
 952	if (set) {
 953		cpumask_andnot(&set->used, &set->used, &msix->mask);
 954		_cpu_mask_set_gen_dec(set);
 955	}
 956
 957	irq_set_affinity_hint(msix->irq, NULL);
 958	cpumask_clear(&msix->mask);
 959	mutex_unlock(&node_affinity.lock);
 960}
 961
 962/* This should be called with node_affinity.lock held */
 963static void find_hw_thread_mask(uint hw_thread_no, cpumask_var_t hw_thread_mask,
 964				struct hfi1_affinity_node_list *affinity)
 965{
 966	int possible, curr_cpu, i;
 967	uint num_cores_per_socket = node_affinity.num_online_cpus /
 968					affinity->num_core_siblings /
 969						node_affinity.num_online_nodes;
 970
 971	cpumask_copy(hw_thread_mask, &affinity->proc.mask);
 972	if (affinity->num_core_siblings > 0) {
 973		/* Removing other siblings not needed for now */
 974		possible = cpumask_weight(hw_thread_mask);
 975		curr_cpu = cpumask_first(hw_thread_mask);
 976		for (i = 0;
 977		     i < num_cores_per_socket * node_affinity.num_online_nodes;
 978		     i++)
 979			curr_cpu = cpumask_next(curr_cpu, hw_thread_mask);
 980
 981		for (; i < possible; i++) {
 982			cpumask_clear_cpu(curr_cpu, hw_thread_mask);
 983			curr_cpu = cpumask_next(curr_cpu, hw_thread_mask);
 984		}
 985
 986		/* Identifying correct HW threads within physical cores */
 987		cpumask_shift_left(hw_thread_mask, hw_thread_mask,
 988				   num_cores_per_socket *
 989				   node_affinity.num_online_nodes *
 990				   hw_thread_no);
 991	}
 992}
 993
 994int hfi1_get_proc_affinity(int node)
 995{
 996	int cpu = -1, ret, i;
 997	struct hfi1_affinity_node *entry;
 998	cpumask_var_t diff, hw_thread_mask, available_mask, intrs_mask;
 999	const struct cpumask *node_mask,
1000		*proc_mask = current->cpus_ptr;
1001	struct hfi1_affinity_node_list *affinity = &node_affinity;
1002	struct cpu_mask_set *set = &affinity->proc;
1003
1004	/*
1005	 * check whether process/context affinity has already
1006	 * been set
1007	 */
1008	if (current->nr_cpus_allowed == 1) {
1009		hfi1_cdbg(PROC, "PID %u %s affinity set to CPU %*pbl",
1010			  current->pid, current->comm,
1011			  cpumask_pr_args(proc_mask));
1012		/*
1013		 * Mark the pre-set CPU as used. This is atomic so we don't
1014		 * need the lock
1015		 */
1016		cpu = cpumask_first(proc_mask);
1017		cpumask_set_cpu(cpu, &set->used);
1018		goto done;
1019	} else if (current->nr_cpus_allowed < cpumask_weight(&set->mask)) {
1020		hfi1_cdbg(PROC, "PID %u %s affinity set to CPU set(s) %*pbl",
1021			  current->pid, current->comm,
1022			  cpumask_pr_args(proc_mask));
1023		goto done;
1024	}
1025
1026	/*
1027	 * The process does not have a preset CPU affinity so find one to
1028	 * recommend using the following algorithm:
1029	 *
1030	 * For each user process that is opening a context on HFI Y:
1031	 *  a) If all cores are filled, reinitialize the bitmask
1032	 *  b) Fill real cores first, then HT cores (First set of HT
1033	 *     cores on all physical cores, then second set of HT core,
1034	 *     and, so on) in the following order:
1035	 *
1036	 *     1. Same NUMA node as HFI Y and not running an IRQ
1037	 *        handler
1038	 *     2. Same NUMA node as HFI Y and running an IRQ handler
1039	 *     3. Different NUMA node to HFI Y and not running an IRQ
1040	 *        handler
1041	 *     4. Different NUMA node to HFI Y and running an IRQ
1042	 *        handler
1043	 *  c) Mark core as filled in the bitmask. As user processes are
1044	 *     done, clear cores from the bitmask.
1045	 */
1046
1047	ret = zalloc_cpumask_var(&diff, GFP_KERNEL);
1048	if (!ret)
1049		goto done;
1050	ret = zalloc_cpumask_var(&hw_thread_mask, GFP_KERNEL);
1051	if (!ret)
1052		goto free_diff;
1053	ret = zalloc_cpumask_var(&available_mask, GFP_KERNEL);
1054	if (!ret)
1055		goto free_hw_thread_mask;
1056	ret = zalloc_cpumask_var(&intrs_mask, GFP_KERNEL);
1057	if (!ret)
1058		goto free_available_mask;
1059
1060	mutex_lock(&affinity->lock);
1061	/*
1062	 * If we've used all available HW threads, clear the mask and start
1063	 * overloading.
1064	 */
1065	_cpu_mask_set_gen_inc(set);
1066
1067	/*
1068	 * If NUMA node has CPUs used by interrupt handlers, include them in the
1069	 * interrupt handler mask.
1070	 */
1071	entry = node_affinity_lookup(node);
1072	if (entry) {
1073		cpumask_copy(intrs_mask, (entry->def_intr.gen ?
1074					  &entry->def_intr.mask :
1075					  &entry->def_intr.used));
1076		cpumask_or(intrs_mask, intrs_mask, (entry->rcv_intr.gen ?
1077						    &entry->rcv_intr.mask :
1078						    &entry->rcv_intr.used));
1079		cpumask_or(intrs_mask, intrs_mask, &entry->general_intr_mask);
1080	}
1081	hfi1_cdbg(PROC, "CPUs used by interrupts: %*pbl",
1082		  cpumask_pr_args(intrs_mask));
1083
1084	cpumask_copy(hw_thread_mask, &set->mask);
1085
1086	/*
1087	 * If HT cores are enabled, identify which HW threads within the
1088	 * physical cores should be used.
1089	 */
1090	if (affinity->num_core_siblings > 0) {
1091		for (i = 0; i < affinity->num_core_siblings; i++) {
1092			find_hw_thread_mask(i, hw_thread_mask, affinity);
1093
1094			/*
1095			 * If there's at least one available core for this HW
1096			 * thread number, stop looking for a core.
1097			 *
1098			 * diff will always be not empty at least once in this
1099			 * loop as the used mask gets reset when
1100			 * (set->mask == set->used) before this loop.
1101			 */
1102			cpumask_andnot(diff, hw_thread_mask, &set->used);
1103			if (!cpumask_empty(diff))
1104				break;
1105		}
1106	}
1107	hfi1_cdbg(PROC, "Same available HW thread on all physical CPUs: %*pbl",
1108		  cpumask_pr_args(hw_thread_mask));
1109
1110	node_mask = cpumask_of_node(node);
1111	hfi1_cdbg(PROC, "Device on NUMA %u, CPUs %*pbl", node,
1112		  cpumask_pr_args(node_mask));
1113
1114	/* Get cpumask of available CPUs on preferred NUMA */
1115	cpumask_and(available_mask, hw_thread_mask, node_mask);
1116	cpumask_andnot(available_mask, available_mask, &set->used);
1117	hfi1_cdbg(PROC, "Available CPUs on NUMA %u: %*pbl", node,
1118		  cpumask_pr_args(available_mask));
1119
1120	/*
1121	 * At first, we don't want to place processes on the same
1122	 * CPUs as interrupt handlers. Then, CPUs running interrupt
1123	 * handlers are used.
1124	 *
1125	 * 1) If diff is not empty, then there are CPUs not running
1126	 *    non-interrupt handlers available, so diff gets copied
1127	 *    over to available_mask.
1128	 * 2) If diff is empty, then all CPUs not running interrupt
1129	 *    handlers are taken, so available_mask contains all
1130	 *    available CPUs running interrupt handlers.
1131	 * 3) If available_mask is empty, then all CPUs on the
1132	 *    preferred NUMA node are taken, so other NUMA nodes are
1133	 *    used for process assignments using the same method as
1134	 *    the preferred NUMA node.
1135	 */
1136	cpumask_andnot(diff, available_mask, intrs_mask);
1137	if (!cpumask_empty(diff))
1138		cpumask_copy(available_mask, diff);
1139
1140	/* If we don't have CPUs on the preferred node, use other NUMA nodes */
1141	if (cpumask_empty(available_mask)) {
1142		cpumask_andnot(available_mask, hw_thread_mask, &set->used);
1143		/* Excluding preferred NUMA cores */
1144		cpumask_andnot(available_mask, available_mask, node_mask);
1145		hfi1_cdbg(PROC,
1146			  "Preferred NUMA node cores are taken, cores available in other NUMA nodes: %*pbl",
1147			  cpumask_pr_args(available_mask));
1148
1149		/*
1150		 * At first, we don't want to place processes on the same
1151		 * CPUs as interrupt handlers.
1152		 */
1153		cpumask_andnot(diff, available_mask, intrs_mask);
1154		if (!cpumask_empty(diff))
1155			cpumask_copy(available_mask, diff);
1156	}
1157	hfi1_cdbg(PROC, "Possible CPUs for process: %*pbl",
1158		  cpumask_pr_args(available_mask));
1159
1160	cpu = cpumask_first(available_mask);
1161	if (cpu >= nr_cpu_ids) /* empty */
1162		cpu = -1;
1163	else
1164		cpumask_set_cpu(cpu, &set->used);
1165
1166	mutex_unlock(&affinity->lock);
1167	hfi1_cdbg(PROC, "Process assigned to CPU %d", cpu);
1168
1169	free_cpumask_var(intrs_mask);
1170free_available_mask:
1171	free_cpumask_var(available_mask);
1172free_hw_thread_mask:
1173	free_cpumask_var(hw_thread_mask);
1174free_diff:
1175	free_cpumask_var(diff);
1176done:
1177	return cpu;
1178}
1179
1180void hfi1_put_proc_affinity(int cpu)
1181{
1182	struct hfi1_affinity_node_list *affinity = &node_affinity;
1183	struct cpu_mask_set *set = &affinity->proc;
1184
1185	if (cpu < 0)
1186		return;
1187
1188	mutex_lock(&affinity->lock);
1189	cpu_mask_set_put(set, cpu);
1190	hfi1_cdbg(PROC, "Returning CPU %d for future process assignment", cpu);
1191	mutex_unlock(&affinity->lock);
1192}
v5.9
 
   1/*
   2 * Copyright(c) 2015 - 2020 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#include <linux/topology.h>
  48#include <linux/cpumask.h>
  49#include <linux/module.h>
  50#include <linux/interrupt.h>
  51#include <linux/numa.h>
  52
  53#include "hfi.h"
  54#include "affinity.h"
  55#include "sdma.h"
  56#include "trace.h"
  57
  58struct hfi1_affinity_node_list node_affinity = {
  59	.list = LIST_HEAD_INIT(node_affinity.list),
  60	.lock = __MUTEX_INITIALIZER(node_affinity.lock)
  61};
  62
  63/* Name of IRQ types, indexed by enum irq_type */
  64static const char * const irq_type_names[] = {
  65	"SDMA",
  66	"RCVCTXT",
  67	"NETDEVCTXT",
  68	"GENERAL",
  69	"OTHER",
  70};
  71
  72/* Per NUMA node count of HFI devices */
  73static unsigned int *hfi1_per_node_cntr;
  74
  75static inline void init_cpu_mask_set(struct cpu_mask_set *set)
  76{
  77	cpumask_clear(&set->mask);
  78	cpumask_clear(&set->used);
  79	set->gen = 0;
  80}
  81
  82/* Increment generation of CPU set if needed */
  83static void _cpu_mask_set_gen_inc(struct cpu_mask_set *set)
  84{
  85	if (cpumask_equal(&set->mask, &set->used)) {
  86		/*
  87		 * We've used up all the CPUs, bump up the generation
  88		 * and reset the 'used' map
  89		 */
  90		set->gen++;
  91		cpumask_clear(&set->used);
  92	}
  93}
  94
  95static void _cpu_mask_set_gen_dec(struct cpu_mask_set *set)
  96{
  97	if (cpumask_empty(&set->used) && set->gen) {
  98		set->gen--;
  99		cpumask_copy(&set->used, &set->mask);
 100	}
 101}
 102
 103/* Get the first CPU from the list of unused CPUs in a CPU set data structure */
 104static int cpu_mask_set_get_first(struct cpu_mask_set *set, cpumask_var_t diff)
 105{
 106	int cpu;
 107
 108	if (!diff || !set)
 109		return -EINVAL;
 110
 111	_cpu_mask_set_gen_inc(set);
 112
 113	/* Find out CPUs left in CPU mask */
 114	cpumask_andnot(diff, &set->mask, &set->used);
 115
 116	cpu = cpumask_first(diff);
 117	if (cpu >= nr_cpu_ids) /* empty */
 118		cpu = -EINVAL;
 119	else
 120		cpumask_set_cpu(cpu, &set->used);
 121
 122	return cpu;
 123}
 124
 125static void cpu_mask_set_put(struct cpu_mask_set *set, int cpu)
 126{
 127	if (!set)
 128		return;
 129
 130	cpumask_clear_cpu(cpu, &set->used);
 131	_cpu_mask_set_gen_dec(set);
 132}
 133
 134/* Initialize non-HT cpu cores mask */
 135void init_real_cpu_mask(void)
 136{
 137	int possible, curr_cpu, i, ht;
 138
 139	cpumask_clear(&node_affinity.real_cpu_mask);
 140
 141	/* Start with cpu online mask as the real cpu mask */
 142	cpumask_copy(&node_affinity.real_cpu_mask, cpu_online_mask);
 143
 144	/*
 145	 * Remove HT cores from the real cpu mask.  Do this in two steps below.
 146	 */
 147	possible = cpumask_weight(&node_affinity.real_cpu_mask);
 148	ht = cpumask_weight(topology_sibling_cpumask(
 149				cpumask_first(&node_affinity.real_cpu_mask)));
 150	/*
 151	 * Step 1.  Skip over the first N HT siblings and use them as the
 152	 * "real" cores.  Assumes that HT cores are not enumerated in
 153	 * succession (except in the single core case).
 154	 */
 155	curr_cpu = cpumask_first(&node_affinity.real_cpu_mask);
 156	for (i = 0; i < possible / ht; i++)
 157		curr_cpu = cpumask_next(curr_cpu, &node_affinity.real_cpu_mask);
 158	/*
 159	 * Step 2.  Remove the remaining HT siblings.  Use cpumask_next() to
 160	 * skip any gaps.
 161	 */
 162	for (; i < possible; i++) {
 163		cpumask_clear_cpu(curr_cpu, &node_affinity.real_cpu_mask);
 164		curr_cpu = cpumask_next(curr_cpu, &node_affinity.real_cpu_mask);
 165	}
 166}
 167
 168int node_affinity_init(void)
 169{
 170	int node;
 171	struct pci_dev *dev = NULL;
 172	const struct pci_device_id *ids = hfi1_pci_tbl;
 173
 174	cpumask_clear(&node_affinity.proc.used);
 175	cpumask_copy(&node_affinity.proc.mask, cpu_online_mask);
 176
 177	node_affinity.proc.gen = 0;
 178	node_affinity.num_core_siblings =
 179				cpumask_weight(topology_sibling_cpumask(
 180					cpumask_first(&node_affinity.proc.mask)
 181					));
 182	node_affinity.num_possible_nodes = num_possible_nodes();
 183	node_affinity.num_online_nodes = num_online_nodes();
 184	node_affinity.num_online_cpus = num_online_cpus();
 185
 186	/*
 187	 * The real cpu mask is part of the affinity struct but it has to be
 188	 * initialized early. It is needed to calculate the number of user
 189	 * contexts in set_up_context_variables().
 190	 */
 191	init_real_cpu_mask();
 192
 193	hfi1_per_node_cntr = kcalloc(node_affinity.num_possible_nodes,
 194				     sizeof(*hfi1_per_node_cntr), GFP_KERNEL);
 195	if (!hfi1_per_node_cntr)
 196		return -ENOMEM;
 197
 198	while (ids->vendor) {
 199		dev = NULL;
 200		while ((dev = pci_get_device(ids->vendor, ids->device, dev))) {
 201			node = pcibus_to_node(dev->bus);
 202			if (node < 0)
 203				goto out;
 204
 205			hfi1_per_node_cntr[node]++;
 206		}
 207		ids++;
 208	}
 209
 210	return 0;
 211
 212out:
 213	/*
 214	 * Invalid PCI NUMA node information found, note it, and populate
 215	 * our database 1:1.
 216	 */
 217	pr_err("HFI: Invalid PCI NUMA node. Performance may be affected\n");
 218	pr_err("HFI: System BIOS may need to be upgraded\n");
 219	for (node = 0; node < node_affinity.num_possible_nodes; node++)
 220		hfi1_per_node_cntr[node] = 1;
 221
 
 
 222	return 0;
 223}
 224
 225static void node_affinity_destroy(struct hfi1_affinity_node *entry)
 226{
 227	free_percpu(entry->comp_vect_affinity);
 228	kfree(entry);
 229}
 230
 231void node_affinity_destroy_all(void)
 232{
 233	struct list_head *pos, *q;
 234	struct hfi1_affinity_node *entry;
 235
 236	mutex_lock(&node_affinity.lock);
 237	list_for_each_safe(pos, q, &node_affinity.list) {
 238		entry = list_entry(pos, struct hfi1_affinity_node,
 239				   list);
 240		list_del(pos);
 241		node_affinity_destroy(entry);
 242	}
 243	mutex_unlock(&node_affinity.lock);
 244	kfree(hfi1_per_node_cntr);
 245}
 246
 247static struct hfi1_affinity_node *node_affinity_allocate(int node)
 248{
 249	struct hfi1_affinity_node *entry;
 250
 251	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
 252	if (!entry)
 253		return NULL;
 254	entry->node = node;
 255	entry->comp_vect_affinity = alloc_percpu(u16);
 256	INIT_LIST_HEAD(&entry->list);
 257
 258	return entry;
 259}
 260
 261/*
 262 * It appends an entry to the list.
 263 * It *must* be called with node_affinity.lock held.
 264 */
 265static void node_affinity_add_tail(struct hfi1_affinity_node *entry)
 266{
 267	list_add_tail(&entry->list, &node_affinity.list);
 268}
 269
 270/* It must be called with node_affinity.lock held */
 271static struct hfi1_affinity_node *node_affinity_lookup(int node)
 272{
 273	struct list_head *pos;
 274	struct hfi1_affinity_node *entry;
 275
 276	list_for_each(pos, &node_affinity.list) {
 277		entry = list_entry(pos, struct hfi1_affinity_node, list);
 278		if (entry->node == node)
 279			return entry;
 280	}
 281
 282	return NULL;
 283}
 284
 285static int per_cpu_affinity_get(cpumask_var_t possible_cpumask,
 286				u16 __percpu *comp_vect_affinity)
 287{
 288	int curr_cpu;
 289	u16 cntr;
 290	u16 prev_cntr;
 291	int ret_cpu;
 292
 293	if (!possible_cpumask) {
 294		ret_cpu = -EINVAL;
 295		goto fail;
 296	}
 297
 298	if (!comp_vect_affinity) {
 299		ret_cpu = -EINVAL;
 300		goto fail;
 301	}
 302
 303	ret_cpu = cpumask_first(possible_cpumask);
 304	if (ret_cpu >= nr_cpu_ids) {
 305		ret_cpu = -EINVAL;
 306		goto fail;
 307	}
 308
 309	prev_cntr = *per_cpu_ptr(comp_vect_affinity, ret_cpu);
 310	for_each_cpu(curr_cpu, possible_cpumask) {
 311		cntr = *per_cpu_ptr(comp_vect_affinity, curr_cpu);
 312
 313		if (cntr < prev_cntr) {
 314			ret_cpu = curr_cpu;
 315			prev_cntr = cntr;
 316		}
 317	}
 318
 319	*per_cpu_ptr(comp_vect_affinity, ret_cpu) += 1;
 320
 321fail:
 322	return ret_cpu;
 323}
 324
 325static int per_cpu_affinity_put_max(cpumask_var_t possible_cpumask,
 326				    u16 __percpu *comp_vect_affinity)
 327{
 328	int curr_cpu;
 329	int max_cpu;
 330	u16 cntr;
 331	u16 prev_cntr;
 332
 333	if (!possible_cpumask)
 334		return -EINVAL;
 335
 336	if (!comp_vect_affinity)
 337		return -EINVAL;
 338
 339	max_cpu = cpumask_first(possible_cpumask);
 340	if (max_cpu >= nr_cpu_ids)
 341		return -EINVAL;
 342
 343	prev_cntr = *per_cpu_ptr(comp_vect_affinity, max_cpu);
 344	for_each_cpu(curr_cpu, possible_cpumask) {
 345		cntr = *per_cpu_ptr(comp_vect_affinity, curr_cpu);
 346
 347		if (cntr > prev_cntr) {
 348			max_cpu = curr_cpu;
 349			prev_cntr = cntr;
 350		}
 351	}
 352
 353	*per_cpu_ptr(comp_vect_affinity, max_cpu) -= 1;
 354
 355	return max_cpu;
 356}
 357
 358/*
 359 * Non-interrupt CPUs are used first, then interrupt CPUs.
 360 * Two already allocated cpu masks must be passed.
 361 */
 362static int _dev_comp_vect_cpu_get(struct hfi1_devdata *dd,
 363				  struct hfi1_affinity_node *entry,
 364				  cpumask_var_t non_intr_cpus,
 365				  cpumask_var_t available_cpus)
 366	__must_hold(&node_affinity.lock)
 367{
 368	int cpu;
 369	struct cpu_mask_set *set = dd->comp_vect;
 370
 371	lockdep_assert_held(&node_affinity.lock);
 372	if (!non_intr_cpus) {
 373		cpu = -1;
 374		goto fail;
 375	}
 376
 377	if (!available_cpus) {
 378		cpu = -1;
 379		goto fail;
 380	}
 381
 382	/* Available CPUs for pinning completion vectors */
 383	_cpu_mask_set_gen_inc(set);
 384	cpumask_andnot(available_cpus, &set->mask, &set->used);
 385
 386	/* Available CPUs without SDMA engine interrupts */
 387	cpumask_andnot(non_intr_cpus, available_cpus,
 388		       &entry->def_intr.used);
 389
 390	/* If there are non-interrupt CPUs available, use them first */
 391	if (!cpumask_empty(non_intr_cpus))
 392		cpu = cpumask_first(non_intr_cpus);
 393	else /* Otherwise, use interrupt CPUs */
 394		cpu = cpumask_first(available_cpus);
 395
 396	if (cpu >= nr_cpu_ids) { /* empty */
 397		cpu = -1;
 398		goto fail;
 399	}
 400	cpumask_set_cpu(cpu, &set->used);
 401
 402fail:
 403	return cpu;
 404}
 405
 406static void _dev_comp_vect_cpu_put(struct hfi1_devdata *dd, int cpu)
 407{
 408	struct cpu_mask_set *set = dd->comp_vect;
 409
 410	if (cpu < 0)
 411		return;
 412
 413	cpu_mask_set_put(set, cpu);
 414}
 415
 416/* _dev_comp_vect_mappings_destroy() is reentrant */
 417static void _dev_comp_vect_mappings_destroy(struct hfi1_devdata *dd)
 418{
 419	int i, cpu;
 420
 421	if (!dd->comp_vect_mappings)
 422		return;
 423
 424	for (i = 0; i < dd->comp_vect_possible_cpus; i++) {
 425		cpu = dd->comp_vect_mappings[i];
 426		_dev_comp_vect_cpu_put(dd, cpu);
 427		dd->comp_vect_mappings[i] = -1;
 428		hfi1_cdbg(AFFINITY,
 429			  "[%s] Release CPU %d from completion vector %d",
 430			  rvt_get_ibdev_name(&(dd)->verbs_dev.rdi), cpu, i);
 431	}
 432
 433	kfree(dd->comp_vect_mappings);
 434	dd->comp_vect_mappings = NULL;
 435}
 436
 437/*
 438 * This function creates the table for looking up CPUs for completion vectors.
 439 * num_comp_vectors needs to have been initilized before calling this function.
 440 */
 441static int _dev_comp_vect_mappings_create(struct hfi1_devdata *dd,
 442					  struct hfi1_affinity_node *entry)
 443	__must_hold(&node_affinity.lock)
 444{
 445	int i, cpu, ret;
 446	cpumask_var_t non_intr_cpus;
 447	cpumask_var_t available_cpus;
 448
 449	lockdep_assert_held(&node_affinity.lock);
 450
 451	if (!zalloc_cpumask_var(&non_intr_cpus, GFP_KERNEL))
 452		return -ENOMEM;
 453
 454	if (!zalloc_cpumask_var(&available_cpus, GFP_KERNEL)) {
 455		free_cpumask_var(non_intr_cpus);
 456		return -ENOMEM;
 457	}
 458
 459	dd->comp_vect_mappings = kcalloc(dd->comp_vect_possible_cpus,
 460					 sizeof(*dd->comp_vect_mappings),
 461					 GFP_KERNEL);
 462	if (!dd->comp_vect_mappings) {
 463		ret = -ENOMEM;
 464		goto fail;
 465	}
 466	for (i = 0; i < dd->comp_vect_possible_cpus; i++)
 467		dd->comp_vect_mappings[i] = -1;
 468
 469	for (i = 0; i < dd->comp_vect_possible_cpus; i++) {
 470		cpu = _dev_comp_vect_cpu_get(dd, entry, non_intr_cpus,
 471					     available_cpus);
 472		if (cpu < 0) {
 473			ret = -EINVAL;
 474			goto fail;
 475		}
 476
 477		dd->comp_vect_mappings[i] = cpu;
 478		hfi1_cdbg(AFFINITY,
 479			  "[%s] Completion Vector %d -> CPU %d",
 480			  rvt_get_ibdev_name(&(dd)->verbs_dev.rdi), i, cpu);
 481	}
 482
 483	free_cpumask_var(available_cpus);
 484	free_cpumask_var(non_intr_cpus);
 485	return 0;
 486
 487fail:
 488	free_cpumask_var(available_cpus);
 489	free_cpumask_var(non_intr_cpus);
 490	_dev_comp_vect_mappings_destroy(dd);
 491
 492	return ret;
 493}
 494
 495int hfi1_comp_vectors_set_up(struct hfi1_devdata *dd)
 496{
 497	int ret;
 498	struct hfi1_affinity_node *entry;
 499
 500	mutex_lock(&node_affinity.lock);
 501	entry = node_affinity_lookup(dd->node);
 502	if (!entry) {
 503		ret = -EINVAL;
 504		goto unlock;
 505	}
 506	ret = _dev_comp_vect_mappings_create(dd, entry);
 507unlock:
 508	mutex_unlock(&node_affinity.lock);
 509
 510	return ret;
 511}
 512
 513void hfi1_comp_vectors_clean_up(struct hfi1_devdata *dd)
 514{
 515	_dev_comp_vect_mappings_destroy(dd);
 516}
 517
 518int hfi1_comp_vect_mappings_lookup(struct rvt_dev_info *rdi, int comp_vect)
 519{
 520	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
 521	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
 522
 523	if (!dd->comp_vect_mappings)
 524		return -EINVAL;
 525	if (comp_vect >= dd->comp_vect_possible_cpus)
 526		return -EINVAL;
 527
 528	return dd->comp_vect_mappings[comp_vect];
 529}
 530
 531/*
 532 * It assumes dd->comp_vect_possible_cpus is available.
 533 */
 534static int _dev_comp_vect_cpu_mask_init(struct hfi1_devdata *dd,
 535					struct hfi1_affinity_node *entry,
 536					bool first_dev_init)
 537	__must_hold(&node_affinity.lock)
 538{
 539	int i, j, curr_cpu;
 540	int possible_cpus_comp_vect = 0;
 541	struct cpumask *dev_comp_vect_mask = &dd->comp_vect->mask;
 542
 543	lockdep_assert_held(&node_affinity.lock);
 544	/*
 545	 * If there's only one CPU available for completion vectors, then
 546	 * there will only be one completion vector available. Othewise,
 547	 * the number of completion vector available will be the number of
 548	 * available CPUs divide it by the number of devices in the
 549	 * local NUMA node.
 550	 */
 551	if (cpumask_weight(&entry->comp_vect_mask) == 1) {
 552		possible_cpus_comp_vect = 1;
 553		dd_dev_warn(dd,
 554			    "Number of kernel receive queues is too large for completion vector affinity to be effective\n");
 555	} else {
 556		possible_cpus_comp_vect +=
 557			cpumask_weight(&entry->comp_vect_mask) /
 558				       hfi1_per_node_cntr[dd->node];
 559
 560		/*
 561		 * If the completion vector CPUs available doesn't divide
 562		 * evenly among devices, then the first device device to be
 563		 * initialized gets an extra CPU.
 564		 */
 565		if (first_dev_init &&
 566		    cpumask_weight(&entry->comp_vect_mask) %
 567		    hfi1_per_node_cntr[dd->node] != 0)
 568			possible_cpus_comp_vect++;
 569	}
 570
 571	dd->comp_vect_possible_cpus = possible_cpus_comp_vect;
 572
 573	/* Reserving CPUs for device completion vector */
 574	for (i = 0; i < dd->comp_vect_possible_cpus; i++) {
 575		curr_cpu = per_cpu_affinity_get(&entry->comp_vect_mask,
 576						entry->comp_vect_affinity);
 577		if (curr_cpu < 0)
 578			goto fail;
 579
 580		cpumask_set_cpu(curr_cpu, dev_comp_vect_mask);
 581	}
 582
 583	hfi1_cdbg(AFFINITY,
 584		  "[%s] Completion vector affinity CPU set(s) %*pbl",
 585		  rvt_get_ibdev_name(&(dd)->verbs_dev.rdi),
 586		  cpumask_pr_args(dev_comp_vect_mask));
 587
 588	return 0;
 589
 590fail:
 591	for (j = 0; j < i; j++)
 592		per_cpu_affinity_put_max(&entry->comp_vect_mask,
 593					 entry->comp_vect_affinity);
 594
 595	return curr_cpu;
 596}
 597
 598/*
 599 * It assumes dd->comp_vect_possible_cpus is available.
 600 */
 601static void _dev_comp_vect_cpu_mask_clean_up(struct hfi1_devdata *dd,
 602					     struct hfi1_affinity_node *entry)
 603	__must_hold(&node_affinity.lock)
 604{
 605	int i, cpu;
 606
 607	lockdep_assert_held(&node_affinity.lock);
 608	if (!dd->comp_vect_possible_cpus)
 609		return;
 610
 611	for (i = 0; i < dd->comp_vect_possible_cpus; i++) {
 612		cpu = per_cpu_affinity_put_max(&dd->comp_vect->mask,
 613					       entry->comp_vect_affinity);
 614		/* Clearing CPU in device completion vector cpu mask */
 615		if (cpu >= 0)
 616			cpumask_clear_cpu(cpu, &dd->comp_vect->mask);
 617	}
 618
 619	dd->comp_vect_possible_cpus = 0;
 620}
 621
 622/*
 623 * Interrupt affinity.
 624 *
 625 * non-rcv avail gets a default mask that
 626 * starts as possible cpus with threads reset
 627 * and each rcv avail reset.
 628 *
 629 * rcv avail gets node relative 1 wrapping back
 630 * to the node relative 1 as necessary.
 631 *
 632 */
 633int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
 634{
 635	int node = pcibus_to_node(dd->pcidev->bus);
 636	struct hfi1_affinity_node *entry;
 637	const struct cpumask *local_mask;
 638	int curr_cpu, possible, i, ret;
 639	bool new_entry = false;
 640
 641	/*
 642	 * If the BIOS does not have the NUMA node information set, select
 643	 * NUMA 0 so we get consistent performance.
 644	 */
 645	if (node < 0) {
 646		dd_dev_err(dd, "Invalid PCI NUMA node. Performance may be affected\n");
 647		node = 0;
 648	}
 649	dd->node = node;
 650
 651	local_mask = cpumask_of_node(dd->node);
 652	if (cpumask_first(local_mask) >= nr_cpu_ids)
 653		local_mask = topology_core_cpumask(0);
 654
 655	mutex_lock(&node_affinity.lock);
 656	entry = node_affinity_lookup(dd->node);
 657
 658	/*
 659	 * If this is the first time this NUMA node's affinity is used,
 660	 * create an entry in the global affinity structure and initialize it.
 661	 */
 662	if (!entry) {
 663		entry = node_affinity_allocate(node);
 664		if (!entry) {
 665			dd_dev_err(dd,
 666				   "Unable to allocate global affinity node\n");
 667			ret = -ENOMEM;
 668			goto fail;
 669		}
 670		new_entry = true;
 671
 672		init_cpu_mask_set(&entry->def_intr);
 673		init_cpu_mask_set(&entry->rcv_intr);
 674		cpumask_clear(&entry->comp_vect_mask);
 675		cpumask_clear(&entry->general_intr_mask);
 676		/* Use the "real" cpu mask of this node as the default */
 677		cpumask_and(&entry->def_intr.mask, &node_affinity.real_cpu_mask,
 678			    local_mask);
 679
 680		/* fill in the receive list */
 681		possible = cpumask_weight(&entry->def_intr.mask);
 682		curr_cpu = cpumask_first(&entry->def_intr.mask);
 683
 684		if (possible == 1) {
 685			/* only one CPU, everyone will use it */
 686			cpumask_set_cpu(curr_cpu, &entry->rcv_intr.mask);
 687			cpumask_set_cpu(curr_cpu, &entry->general_intr_mask);
 688		} else {
 689			/*
 690			 * The general/control context will be the first CPU in
 691			 * the default list, so it is removed from the default
 692			 * list and added to the general interrupt list.
 693			 */
 694			cpumask_clear_cpu(curr_cpu, &entry->def_intr.mask);
 695			cpumask_set_cpu(curr_cpu, &entry->general_intr_mask);
 696			curr_cpu = cpumask_next(curr_cpu,
 697						&entry->def_intr.mask);
 698
 699			/*
 700			 * Remove the remaining kernel receive queues from
 701			 * the default list and add them to the receive list.
 702			 */
 703			for (i = 0;
 704			     i < (dd->n_krcv_queues - 1) *
 705				  hfi1_per_node_cntr[dd->node];
 706			     i++) {
 707				cpumask_clear_cpu(curr_cpu,
 708						  &entry->def_intr.mask);
 709				cpumask_set_cpu(curr_cpu,
 710						&entry->rcv_intr.mask);
 711				curr_cpu = cpumask_next(curr_cpu,
 712							&entry->def_intr.mask);
 713				if (curr_cpu >= nr_cpu_ids)
 714					break;
 715			}
 716
 717			/*
 718			 * If there ends up being 0 CPU cores leftover for SDMA
 719			 * engines, use the same CPU cores as general/control
 720			 * context.
 721			 */
 722			if (cpumask_weight(&entry->def_intr.mask) == 0)
 723				cpumask_copy(&entry->def_intr.mask,
 724					     &entry->general_intr_mask);
 725		}
 726
 727		/* Determine completion vector CPUs for the entire node */
 728		cpumask_and(&entry->comp_vect_mask,
 729			    &node_affinity.real_cpu_mask, local_mask);
 730		cpumask_andnot(&entry->comp_vect_mask,
 731			       &entry->comp_vect_mask,
 732			       &entry->rcv_intr.mask);
 733		cpumask_andnot(&entry->comp_vect_mask,
 734			       &entry->comp_vect_mask,
 735			       &entry->general_intr_mask);
 736
 737		/*
 738		 * If there ends up being 0 CPU cores leftover for completion
 739		 * vectors, use the same CPU core as the general/control
 740		 * context.
 741		 */
 742		if (cpumask_weight(&entry->comp_vect_mask) == 0)
 743			cpumask_copy(&entry->comp_vect_mask,
 744				     &entry->general_intr_mask);
 745	}
 746
 747	ret = _dev_comp_vect_cpu_mask_init(dd, entry, new_entry);
 748	if (ret < 0)
 749		goto fail;
 750
 751	if (new_entry)
 752		node_affinity_add_tail(entry);
 753
 
 754	mutex_unlock(&node_affinity.lock);
 755
 756	return 0;
 757
 758fail:
 759	if (new_entry)
 760		node_affinity_destroy(entry);
 761	mutex_unlock(&node_affinity.lock);
 762	return ret;
 763}
 764
 765void hfi1_dev_affinity_clean_up(struct hfi1_devdata *dd)
 766{
 767	struct hfi1_affinity_node *entry;
 768
 769	if (dd->node < 0)
 770		return;
 771
 772	mutex_lock(&node_affinity.lock);
 
 
 773	entry = node_affinity_lookup(dd->node);
 774	if (!entry)
 775		goto unlock;
 776
 777	/*
 778	 * Free device completion vector CPUs to be used by future
 779	 * completion vectors
 780	 */
 781	_dev_comp_vect_cpu_mask_clean_up(dd, entry);
 782unlock:
 
 783	mutex_unlock(&node_affinity.lock);
 784	dd->node = NUMA_NO_NODE;
 785}
 786
 787/*
 788 * Function updates the irq affinity hint for msix after it has been changed
 789 * by the user using the /proc/irq interface. This function only accepts
 790 * one cpu in the mask.
 791 */
 792static void hfi1_update_sdma_affinity(struct hfi1_msix_entry *msix, int cpu)
 793{
 794	struct sdma_engine *sde = msix->arg;
 795	struct hfi1_devdata *dd = sde->dd;
 796	struct hfi1_affinity_node *entry;
 797	struct cpu_mask_set *set;
 798	int i, old_cpu;
 799
 800	if (cpu > num_online_cpus() || cpu == sde->cpu)
 801		return;
 802
 803	mutex_lock(&node_affinity.lock);
 804	entry = node_affinity_lookup(dd->node);
 805	if (!entry)
 806		goto unlock;
 807
 808	old_cpu = sde->cpu;
 809	sde->cpu = cpu;
 810	cpumask_clear(&msix->mask);
 811	cpumask_set_cpu(cpu, &msix->mask);
 812	dd_dev_dbg(dd, "IRQ: %u, type %s engine %u -> cpu: %d\n",
 813		   msix->irq, irq_type_names[msix->type],
 814		   sde->this_idx, cpu);
 815	irq_set_affinity_hint(msix->irq, &msix->mask);
 816
 817	/*
 818	 * Set the new cpu in the hfi1_affinity_node and clean
 819	 * the old cpu if it is not used by any other IRQ
 820	 */
 821	set = &entry->def_intr;
 822	cpumask_set_cpu(cpu, &set->mask);
 823	cpumask_set_cpu(cpu, &set->used);
 824	for (i = 0; i < dd->msix_info.max_requested; i++) {
 825		struct hfi1_msix_entry *other_msix;
 826
 827		other_msix = &dd->msix_info.msix_entries[i];
 828		if (other_msix->type != IRQ_SDMA || other_msix == msix)
 829			continue;
 830
 831		if (cpumask_test_cpu(old_cpu, &other_msix->mask))
 832			goto unlock;
 833	}
 834	cpumask_clear_cpu(old_cpu, &set->mask);
 835	cpumask_clear_cpu(old_cpu, &set->used);
 836unlock:
 837	mutex_unlock(&node_affinity.lock);
 838}
 839
 840static void hfi1_irq_notifier_notify(struct irq_affinity_notify *notify,
 841				     const cpumask_t *mask)
 842{
 843	int cpu = cpumask_first(mask);
 844	struct hfi1_msix_entry *msix = container_of(notify,
 845						    struct hfi1_msix_entry,
 846						    notify);
 847
 848	/* Only one CPU configuration supported currently */
 849	hfi1_update_sdma_affinity(msix, cpu);
 850}
 851
 852static void hfi1_irq_notifier_release(struct kref *ref)
 853{
 854	/*
 855	 * This is required by affinity notifier. We don't have anything to
 856	 * free here.
 857	 */
 858}
 859
 860static void hfi1_setup_sdma_notifier(struct hfi1_msix_entry *msix)
 861{
 862	struct irq_affinity_notify *notify = &msix->notify;
 863
 864	notify->irq = msix->irq;
 865	notify->notify = hfi1_irq_notifier_notify;
 866	notify->release = hfi1_irq_notifier_release;
 867
 868	if (irq_set_affinity_notifier(notify->irq, notify))
 869		pr_err("Failed to register sdma irq affinity notifier for irq %d\n",
 870		       notify->irq);
 871}
 872
 873static void hfi1_cleanup_sdma_notifier(struct hfi1_msix_entry *msix)
 874{
 875	struct irq_affinity_notify *notify = &msix->notify;
 876
 877	if (irq_set_affinity_notifier(notify->irq, NULL))
 878		pr_err("Failed to cleanup sdma irq affinity notifier for irq %d\n",
 879		       notify->irq);
 880}
 881
 882/*
 883 * Function sets the irq affinity for msix.
 884 * It *must* be called with node_affinity.lock held.
 885 */
 886static int get_irq_affinity(struct hfi1_devdata *dd,
 887			    struct hfi1_msix_entry *msix)
 888{
 889	cpumask_var_t diff;
 890	struct hfi1_affinity_node *entry;
 891	struct cpu_mask_set *set = NULL;
 892	struct sdma_engine *sde = NULL;
 893	struct hfi1_ctxtdata *rcd = NULL;
 894	char extra[64];
 895	int cpu = -1;
 896
 897	extra[0] = '\0';
 898	cpumask_clear(&msix->mask);
 899
 900	entry = node_affinity_lookup(dd->node);
 901
 902	switch (msix->type) {
 903	case IRQ_SDMA:
 904		sde = (struct sdma_engine *)msix->arg;
 905		scnprintf(extra, 64, "engine %u", sde->this_idx);
 906		set = &entry->def_intr;
 907		break;
 908	case IRQ_GENERAL:
 909		cpu = cpumask_first(&entry->general_intr_mask);
 910		break;
 911	case IRQ_RCVCTXT:
 912		rcd = (struct hfi1_ctxtdata *)msix->arg;
 913		if (rcd->ctxt == HFI1_CTRL_CTXT)
 914			cpu = cpumask_first(&entry->general_intr_mask);
 915		else
 916			set = &entry->rcv_intr;
 917		scnprintf(extra, 64, "ctxt %u", rcd->ctxt);
 918		break;
 919	case IRQ_NETDEVCTXT:
 920		rcd = (struct hfi1_ctxtdata *)msix->arg;
 921		set = &entry->def_intr;
 922		scnprintf(extra, 64, "ctxt %u", rcd->ctxt);
 923		break;
 924	default:
 925		dd_dev_err(dd, "Invalid IRQ type %d\n", msix->type);
 926		return -EINVAL;
 927	}
 928
 929	/*
 930	 * The general and control contexts are placed on a particular
 931	 * CPU, which is set above. Skip accounting for it. Everything else
 932	 * finds its CPU here.
 933	 */
 934	if (cpu == -1 && set) {
 935		if (!zalloc_cpumask_var(&diff, GFP_KERNEL))
 936			return -ENOMEM;
 937
 938		cpu = cpu_mask_set_get_first(set, diff);
 939		if (cpu < 0) {
 940			free_cpumask_var(diff);
 941			dd_dev_err(dd, "Failure to obtain CPU for IRQ\n");
 942			return cpu;
 943		}
 944
 945		free_cpumask_var(diff);
 946	}
 947
 948	cpumask_set_cpu(cpu, &msix->mask);
 949	dd_dev_info(dd, "IRQ: %u, type %s %s -> cpu: %d\n",
 950		    msix->irq, irq_type_names[msix->type],
 951		    extra, cpu);
 952	irq_set_affinity_hint(msix->irq, &msix->mask);
 953
 954	if (msix->type == IRQ_SDMA) {
 955		sde->cpu = cpu;
 956		hfi1_setup_sdma_notifier(msix);
 957	}
 958
 959	return 0;
 960}
 961
 962int hfi1_get_irq_affinity(struct hfi1_devdata *dd, struct hfi1_msix_entry *msix)
 963{
 964	int ret;
 965
 966	mutex_lock(&node_affinity.lock);
 967	ret = get_irq_affinity(dd, msix);
 968	mutex_unlock(&node_affinity.lock);
 969	return ret;
 970}
 971
 972void hfi1_put_irq_affinity(struct hfi1_devdata *dd,
 973			   struct hfi1_msix_entry *msix)
 974{
 975	struct cpu_mask_set *set = NULL;
 976	struct hfi1_ctxtdata *rcd;
 977	struct hfi1_affinity_node *entry;
 978
 979	mutex_lock(&node_affinity.lock);
 980	entry = node_affinity_lookup(dd->node);
 981
 982	switch (msix->type) {
 983	case IRQ_SDMA:
 984		set = &entry->def_intr;
 985		hfi1_cleanup_sdma_notifier(msix);
 986		break;
 987	case IRQ_GENERAL:
 988		/* Don't do accounting for general contexts */
 989		break;
 990	case IRQ_RCVCTXT:
 991		rcd = (struct hfi1_ctxtdata *)msix->arg;
 
 992		/* Don't do accounting for control contexts */
 993		if (rcd->ctxt != HFI1_CTRL_CTXT)
 994			set = &entry->rcv_intr;
 995		break;
 
 996	case IRQ_NETDEVCTXT:
 997		rcd = (struct hfi1_ctxtdata *)msix->arg;
 998		set = &entry->def_intr;
 999		break;
1000	default:
1001		mutex_unlock(&node_affinity.lock);
1002		return;
1003	}
1004
1005	if (set) {
1006		cpumask_andnot(&set->used, &set->used, &msix->mask);
1007		_cpu_mask_set_gen_dec(set);
1008	}
1009
1010	irq_set_affinity_hint(msix->irq, NULL);
1011	cpumask_clear(&msix->mask);
1012	mutex_unlock(&node_affinity.lock);
1013}
1014
1015/* This should be called with node_affinity.lock held */
1016static void find_hw_thread_mask(uint hw_thread_no, cpumask_var_t hw_thread_mask,
1017				struct hfi1_affinity_node_list *affinity)
1018{
1019	int possible, curr_cpu, i;
1020	uint num_cores_per_socket = node_affinity.num_online_cpus /
1021					affinity->num_core_siblings /
1022						node_affinity.num_online_nodes;
1023
1024	cpumask_copy(hw_thread_mask, &affinity->proc.mask);
1025	if (affinity->num_core_siblings > 0) {
1026		/* Removing other siblings not needed for now */
1027		possible = cpumask_weight(hw_thread_mask);
1028		curr_cpu = cpumask_first(hw_thread_mask);
1029		for (i = 0;
1030		     i < num_cores_per_socket * node_affinity.num_online_nodes;
1031		     i++)
1032			curr_cpu = cpumask_next(curr_cpu, hw_thread_mask);
1033
1034		for (; i < possible; i++) {
1035			cpumask_clear_cpu(curr_cpu, hw_thread_mask);
1036			curr_cpu = cpumask_next(curr_cpu, hw_thread_mask);
1037		}
1038
1039		/* Identifying correct HW threads within physical cores */
1040		cpumask_shift_left(hw_thread_mask, hw_thread_mask,
1041				   num_cores_per_socket *
1042				   node_affinity.num_online_nodes *
1043				   hw_thread_no);
1044	}
1045}
1046
1047int hfi1_get_proc_affinity(int node)
1048{
1049	int cpu = -1, ret, i;
1050	struct hfi1_affinity_node *entry;
1051	cpumask_var_t diff, hw_thread_mask, available_mask, intrs_mask;
1052	const struct cpumask *node_mask,
1053		*proc_mask = current->cpus_ptr;
1054	struct hfi1_affinity_node_list *affinity = &node_affinity;
1055	struct cpu_mask_set *set = &affinity->proc;
1056
1057	/*
1058	 * check whether process/context affinity has already
1059	 * been set
1060	 */
1061	if (current->nr_cpus_allowed == 1) {
1062		hfi1_cdbg(PROC, "PID %u %s affinity set to CPU %*pbl",
1063			  current->pid, current->comm,
1064			  cpumask_pr_args(proc_mask));
1065		/*
1066		 * Mark the pre-set CPU as used. This is atomic so we don't
1067		 * need the lock
1068		 */
1069		cpu = cpumask_first(proc_mask);
1070		cpumask_set_cpu(cpu, &set->used);
1071		goto done;
1072	} else if (current->nr_cpus_allowed < cpumask_weight(&set->mask)) {
1073		hfi1_cdbg(PROC, "PID %u %s affinity set to CPU set(s) %*pbl",
1074			  current->pid, current->comm,
1075			  cpumask_pr_args(proc_mask));
1076		goto done;
1077	}
1078
1079	/*
1080	 * The process does not have a preset CPU affinity so find one to
1081	 * recommend using the following algorithm:
1082	 *
1083	 * For each user process that is opening a context on HFI Y:
1084	 *  a) If all cores are filled, reinitialize the bitmask
1085	 *  b) Fill real cores first, then HT cores (First set of HT
1086	 *     cores on all physical cores, then second set of HT core,
1087	 *     and, so on) in the following order:
1088	 *
1089	 *     1. Same NUMA node as HFI Y and not running an IRQ
1090	 *        handler
1091	 *     2. Same NUMA node as HFI Y and running an IRQ handler
1092	 *     3. Different NUMA node to HFI Y and not running an IRQ
1093	 *        handler
1094	 *     4. Different NUMA node to HFI Y and running an IRQ
1095	 *        handler
1096	 *  c) Mark core as filled in the bitmask. As user processes are
1097	 *     done, clear cores from the bitmask.
1098	 */
1099
1100	ret = zalloc_cpumask_var(&diff, GFP_KERNEL);
1101	if (!ret)
1102		goto done;
1103	ret = zalloc_cpumask_var(&hw_thread_mask, GFP_KERNEL);
1104	if (!ret)
1105		goto free_diff;
1106	ret = zalloc_cpumask_var(&available_mask, GFP_KERNEL);
1107	if (!ret)
1108		goto free_hw_thread_mask;
1109	ret = zalloc_cpumask_var(&intrs_mask, GFP_KERNEL);
1110	if (!ret)
1111		goto free_available_mask;
1112
1113	mutex_lock(&affinity->lock);
1114	/*
1115	 * If we've used all available HW threads, clear the mask and start
1116	 * overloading.
1117	 */
1118	_cpu_mask_set_gen_inc(set);
1119
1120	/*
1121	 * If NUMA node has CPUs used by interrupt handlers, include them in the
1122	 * interrupt handler mask.
1123	 */
1124	entry = node_affinity_lookup(node);
1125	if (entry) {
1126		cpumask_copy(intrs_mask, (entry->def_intr.gen ?
1127					  &entry->def_intr.mask :
1128					  &entry->def_intr.used));
1129		cpumask_or(intrs_mask, intrs_mask, (entry->rcv_intr.gen ?
1130						    &entry->rcv_intr.mask :
1131						    &entry->rcv_intr.used));
1132		cpumask_or(intrs_mask, intrs_mask, &entry->general_intr_mask);
1133	}
1134	hfi1_cdbg(PROC, "CPUs used by interrupts: %*pbl",
1135		  cpumask_pr_args(intrs_mask));
1136
1137	cpumask_copy(hw_thread_mask, &set->mask);
1138
1139	/*
1140	 * If HT cores are enabled, identify which HW threads within the
1141	 * physical cores should be used.
1142	 */
1143	if (affinity->num_core_siblings > 0) {
1144		for (i = 0; i < affinity->num_core_siblings; i++) {
1145			find_hw_thread_mask(i, hw_thread_mask, affinity);
1146
1147			/*
1148			 * If there's at least one available core for this HW
1149			 * thread number, stop looking for a core.
1150			 *
1151			 * diff will always be not empty at least once in this
1152			 * loop as the used mask gets reset when
1153			 * (set->mask == set->used) before this loop.
1154			 */
1155			cpumask_andnot(diff, hw_thread_mask, &set->used);
1156			if (!cpumask_empty(diff))
1157				break;
1158		}
1159	}
1160	hfi1_cdbg(PROC, "Same available HW thread on all physical CPUs: %*pbl",
1161		  cpumask_pr_args(hw_thread_mask));
1162
1163	node_mask = cpumask_of_node(node);
1164	hfi1_cdbg(PROC, "Device on NUMA %u, CPUs %*pbl", node,
1165		  cpumask_pr_args(node_mask));
1166
1167	/* Get cpumask of available CPUs on preferred NUMA */
1168	cpumask_and(available_mask, hw_thread_mask, node_mask);
1169	cpumask_andnot(available_mask, available_mask, &set->used);
1170	hfi1_cdbg(PROC, "Available CPUs on NUMA %u: %*pbl", node,
1171		  cpumask_pr_args(available_mask));
1172
1173	/*
1174	 * At first, we don't want to place processes on the same
1175	 * CPUs as interrupt handlers. Then, CPUs running interrupt
1176	 * handlers are used.
1177	 *
1178	 * 1) If diff is not empty, then there are CPUs not running
1179	 *    non-interrupt handlers available, so diff gets copied
1180	 *    over to available_mask.
1181	 * 2) If diff is empty, then all CPUs not running interrupt
1182	 *    handlers are taken, so available_mask contains all
1183	 *    available CPUs running interrupt handlers.
1184	 * 3) If available_mask is empty, then all CPUs on the
1185	 *    preferred NUMA node are taken, so other NUMA nodes are
1186	 *    used for process assignments using the same method as
1187	 *    the preferred NUMA node.
1188	 */
1189	cpumask_andnot(diff, available_mask, intrs_mask);
1190	if (!cpumask_empty(diff))
1191		cpumask_copy(available_mask, diff);
1192
1193	/* If we don't have CPUs on the preferred node, use other NUMA nodes */
1194	if (cpumask_empty(available_mask)) {
1195		cpumask_andnot(available_mask, hw_thread_mask, &set->used);
1196		/* Excluding preferred NUMA cores */
1197		cpumask_andnot(available_mask, available_mask, node_mask);
1198		hfi1_cdbg(PROC,
1199			  "Preferred NUMA node cores are taken, cores available in other NUMA nodes: %*pbl",
1200			  cpumask_pr_args(available_mask));
1201
1202		/*
1203		 * At first, we don't want to place processes on the same
1204		 * CPUs as interrupt handlers.
1205		 */
1206		cpumask_andnot(diff, available_mask, intrs_mask);
1207		if (!cpumask_empty(diff))
1208			cpumask_copy(available_mask, diff);
1209	}
1210	hfi1_cdbg(PROC, "Possible CPUs for process: %*pbl",
1211		  cpumask_pr_args(available_mask));
1212
1213	cpu = cpumask_first(available_mask);
1214	if (cpu >= nr_cpu_ids) /* empty */
1215		cpu = -1;
1216	else
1217		cpumask_set_cpu(cpu, &set->used);
1218
1219	mutex_unlock(&affinity->lock);
1220	hfi1_cdbg(PROC, "Process assigned to CPU %d", cpu);
1221
1222	free_cpumask_var(intrs_mask);
1223free_available_mask:
1224	free_cpumask_var(available_mask);
1225free_hw_thread_mask:
1226	free_cpumask_var(hw_thread_mask);
1227free_diff:
1228	free_cpumask_var(diff);
1229done:
1230	return cpu;
1231}
1232
1233void hfi1_put_proc_affinity(int cpu)
1234{
1235	struct hfi1_affinity_node_list *affinity = &node_affinity;
1236	struct cpu_mask_set *set = &affinity->proc;
1237
1238	if (cpu < 0)
1239		return;
1240
1241	mutex_lock(&affinity->lock);
1242	cpu_mask_set_put(set, cpu);
1243	hfi1_cdbg(PROC, "Returning CPU %d for future process assignment", cpu);
1244	mutex_unlock(&affinity->lock);
1245}