2 * Copyright(c) 2015, 2016 Intel Corporation.
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.
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.
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.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
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
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.
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.
47 #include <linux/topology.h>
48 #include <linux/cpumask.h>
49 #include <linux/module.h>
56 struct hfi1_affinity_node_list node_affinity = {
57 .list = LIST_HEAD_INIT(node_affinity.list),
58 .lock = __SPIN_LOCK_UNLOCKED(&node_affinity.lock),
61 /* Name of IRQ types, indexed by enum irq_type */
62 static const char * const irq_type_names[] = {
69 /* Per NUMA node count of HFI devices */
70 static unsigned int *hfi1_per_node_cntr;
72 static inline void init_cpu_mask_set(struct cpu_mask_set *set)
74 cpumask_clear(&set->mask);
75 cpumask_clear(&set->used);
79 /* Initialize non-HT cpu cores mask */
80 void init_real_cpu_mask(void)
82 int possible, curr_cpu, i, ht;
84 cpumask_clear(&node_affinity.real_cpu_mask);
86 /* Start with cpu online mask as the real cpu mask */
87 cpumask_copy(&node_affinity.real_cpu_mask, cpu_online_mask);
90 * Remove HT cores from the real cpu mask. Do this in two steps below.
92 possible = cpumask_weight(&node_affinity.real_cpu_mask);
93 ht = cpumask_weight(topology_sibling_cpumask(
94 cpumask_first(&node_affinity.real_cpu_mask)));
96 * Step 1. Skip over the first N HT siblings and use them as the
97 * "real" cores. Assumes that HT cores are not enumerated in
98 * succession (except in the single core case).
100 curr_cpu = cpumask_first(&node_affinity.real_cpu_mask);
101 for (i = 0; i < possible / ht; i++)
102 curr_cpu = cpumask_next(curr_cpu, &node_affinity.real_cpu_mask);
104 * Step 2. Remove the remaining HT siblings. Use cpumask_next() to
107 for (; i < possible; i++) {
108 cpumask_clear_cpu(curr_cpu, &node_affinity.real_cpu_mask);
109 curr_cpu = cpumask_next(curr_cpu, &node_affinity.real_cpu_mask);
113 int node_affinity_init(void)
116 struct pci_dev *dev = NULL;
117 const struct pci_device_id *ids = hfi1_pci_tbl;
119 cpumask_clear(&node_affinity.proc.used);
120 cpumask_copy(&node_affinity.proc.mask, cpu_online_mask);
122 node_affinity.proc.gen = 0;
123 node_affinity.num_core_siblings =
124 cpumask_weight(topology_sibling_cpumask(
125 cpumask_first(&node_affinity.proc.mask)
127 node_affinity.num_online_nodes = num_online_nodes();
128 node_affinity.num_online_cpus = num_online_cpus();
131 * The real cpu mask is part of the affinity struct but it has to be
132 * initialized early. It is needed to calculate the number of user
133 * contexts in set_up_context_variables().
135 init_real_cpu_mask();
137 hfi1_per_node_cntr = kcalloc(num_possible_nodes(),
138 sizeof(*hfi1_per_node_cntr), GFP_KERNEL);
139 if (!hfi1_per_node_cntr)
142 while (ids->vendor) {
144 while ((dev = pci_get_device(ids->vendor, ids->device, dev))) {
145 node = pcibus_to_node(dev->bus);
147 node = numa_node_id();
149 hfi1_per_node_cntr[node]++;
157 void node_affinity_destroy(void)
159 struct list_head *pos, *q;
160 struct hfi1_affinity_node *entry;
162 spin_lock(&node_affinity.lock);
163 list_for_each_safe(pos, q, &node_affinity.list) {
164 entry = list_entry(pos, struct hfi1_affinity_node,
169 spin_unlock(&node_affinity.lock);
170 kfree(hfi1_per_node_cntr);
173 static struct hfi1_affinity_node *node_affinity_allocate(int node)
175 struct hfi1_affinity_node *entry;
177 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
181 INIT_LIST_HEAD(&entry->list);
187 * It appends an entry to the list.
188 * It *must* be called with node_affinity.lock held.
190 static void node_affinity_add_tail(struct hfi1_affinity_node *entry)
192 list_add_tail(&entry->list, &node_affinity.list);
195 /* It must be called with node_affinity.lock held */
196 static struct hfi1_affinity_node *node_affinity_lookup(int node)
198 struct list_head *pos;
199 struct hfi1_affinity_node *entry;
201 list_for_each(pos, &node_affinity.list) {
202 entry = list_entry(pos, struct hfi1_affinity_node, list);
203 if (entry->node == node)
211 * Interrupt affinity.
213 * non-rcv avail gets a default mask that
214 * starts as possible cpus with threads reset
215 * and each rcv avail reset.
217 * rcv avail gets node relative 1 wrapping back
218 * to the node relative 1 as necessary.
221 int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
223 int node = pcibus_to_node(dd->pcidev->bus);
224 struct hfi1_affinity_node *entry;
225 const struct cpumask *local_mask;
226 int curr_cpu, possible, i;
229 node = numa_node_id();
232 local_mask = cpumask_of_node(dd->node);
233 if (cpumask_first(local_mask) >= nr_cpu_ids)
234 local_mask = topology_core_cpumask(0);
236 spin_lock(&node_affinity.lock);
237 entry = node_affinity_lookup(dd->node);
238 spin_unlock(&node_affinity.lock);
241 * If this is the first time this NUMA node's affinity is used,
242 * create an entry in the global affinity structure and initialize it.
245 entry = node_affinity_allocate(node);
248 "Unable to allocate global affinity node\n");
251 init_cpu_mask_set(&entry->def_intr);
252 init_cpu_mask_set(&entry->rcv_intr);
253 cpumask_clear(&entry->general_intr_mask);
254 /* Use the "real" cpu mask of this node as the default */
255 cpumask_and(&entry->def_intr.mask, &node_affinity.real_cpu_mask,
258 /* fill in the receive list */
259 possible = cpumask_weight(&entry->def_intr.mask);
260 curr_cpu = cpumask_first(&entry->def_intr.mask);
263 /* only one CPU, everyone will use it */
264 cpumask_set_cpu(curr_cpu, &entry->rcv_intr.mask);
265 cpumask_set_cpu(curr_cpu, &entry->general_intr_mask);
268 * The general/control context will be the first CPU in
269 * the default list, so it is removed from the default
270 * list and added to the general interrupt list.
272 cpumask_clear_cpu(curr_cpu, &entry->def_intr.mask);
273 cpumask_set_cpu(curr_cpu, &entry->general_intr_mask);
274 curr_cpu = cpumask_next(curr_cpu,
275 &entry->def_intr.mask);
278 * Remove the remaining kernel receive queues from
279 * the default list and add them to the receive list.
282 i < (dd->n_krcv_queues - 1) *
283 hfi1_per_node_cntr[dd->node];
285 cpumask_clear_cpu(curr_cpu,
286 &entry->def_intr.mask);
287 cpumask_set_cpu(curr_cpu,
288 &entry->rcv_intr.mask);
289 curr_cpu = cpumask_next(curr_cpu,
290 &entry->def_intr.mask);
291 if (curr_cpu >= nr_cpu_ids)
296 * If there ends up being 0 CPU cores leftover for SDMA
297 * engines, use the same CPU cores as general/control
300 if (cpumask_weight(&entry->def_intr.mask) == 0)
301 cpumask_copy(&entry->def_intr.mask,
302 &entry->general_intr_mask);
305 spin_lock(&node_affinity.lock);
306 node_affinity_add_tail(entry);
307 spin_unlock(&node_affinity.lock);
313 int hfi1_get_irq_affinity(struct hfi1_devdata *dd, struct hfi1_msix_entry *msix)
317 struct hfi1_affinity_node *entry;
318 struct cpu_mask_set *set = NULL;
319 struct sdma_engine *sde = NULL;
320 struct hfi1_ctxtdata *rcd = NULL;
325 cpumask_clear(&msix->mask);
327 ret = zalloc_cpumask_var(&diff, GFP_KERNEL);
331 spin_lock(&node_affinity.lock);
332 entry = node_affinity_lookup(dd->node);
333 spin_unlock(&node_affinity.lock);
335 switch (msix->type) {
337 sde = (struct sdma_engine *)msix->arg;
338 scnprintf(extra, 64, "engine %u", sde->this_idx);
339 set = &entry->def_intr;
342 cpu = cpumask_first(&entry->general_intr_mask);
345 rcd = (struct hfi1_ctxtdata *)msix->arg;
346 if (rcd->ctxt == HFI1_CTRL_CTXT)
347 cpu = cpumask_first(&entry->general_intr_mask);
349 set = &entry->rcv_intr;
350 scnprintf(extra, 64, "ctxt %u", rcd->ctxt);
353 dd_dev_err(dd, "Invalid IRQ type %d\n", msix->type);
358 * The general and control contexts are placed on a particular
359 * CPU, which is set above. Skip accounting for it. Everything else
360 * finds its CPU here.
362 if (cpu == -1 && set) {
363 spin_lock(&node_affinity.lock);
364 if (cpumask_equal(&set->mask, &set->used)) {
366 * We've used up all the CPUs, bump up the generation
367 * and reset the 'used' map
370 cpumask_clear(&set->used);
372 cpumask_andnot(diff, &set->mask, &set->used);
373 cpu = cpumask_first(diff);
374 cpumask_set_cpu(cpu, &set->used);
375 spin_unlock(&node_affinity.lock);
378 switch (msix->type) {
388 cpumask_set_cpu(cpu, &msix->mask);
389 dd_dev_info(dd, "IRQ vector: %u, type %s %s -> cpu: %d\n",
390 msix->msix.vector, irq_type_names[msix->type],
392 irq_set_affinity_hint(msix->msix.vector, &msix->mask);
394 free_cpumask_var(diff);
398 void hfi1_put_irq_affinity(struct hfi1_devdata *dd,
399 struct hfi1_msix_entry *msix)
401 struct cpu_mask_set *set = NULL;
402 struct hfi1_ctxtdata *rcd;
403 struct hfi1_affinity_node *entry;
405 spin_lock(&node_affinity.lock);
406 entry = node_affinity_lookup(dd->node);
407 spin_unlock(&node_affinity.lock);
409 switch (msix->type) {
411 set = &entry->def_intr;
414 /* Don't do accounting for general contexts */
417 rcd = (struct hfi1_ctxtdata *)msix->arg;
418 /* Don't do accounting for control contexts */
419 if (rcd->ctxt != HFI1_CTRL_CTXT)
420 set = &entry->rcv_intr;
427 spin_lock(&node_affinity.lock);
428 cpumask_andnot(&set->used, &set->used, &msix->mask);
429 if (cpumask_empty(&set->used) && set->gen) {
431 cpumask_copy(&set->used, &set->mask);
433 spin_unlock(&node_affinity.lock);
436 irq_set_affinity_hint(msix->msix.vector, NULL);
437 cpumask_clear(&msix->mask);
440 /* This should be called with node_affinity.lock held */
441 static void find_hw_thread_mask(uint hw_thread_no, cpumask_var_t hw_thread_mask,
442 struct hfi1_affinity_node_list *affinity)
444 int possible, curr_cpu, i;
445 uint num_cores_per_socket = node_affinity.num_online_cpus /
446 affinity->num_core_siblings /
447 node_affinity.num_online_nodes;
449 cpumask_copy(hw_thread_mask, &affinity->proc.mask);
450 if (affinity->num_core_siblings > 0) {
451 /* Removing other siblings not needed for now */
452 possible = cpumask_weight(hw_thread_mask);
453 curr_cpu = cpumask_first(hw_thread_mask);
455 i < num_cores_per_socket * node_affinity.num_online_nodes;
457 curr_cpu = cpumask_next(curr_cpu, hw_thread_mask);
459 for (; i < possible; i++) {
460 cpumask_clear_cpu(curr_cpu, hw_thread_mask);
461 curr_cpu = cpumask_next(curr_cpu, hw_thread_mask);
464 /* Identifying correct HW threads within physical cores */
465 cpumask_shift_left(hw_thread_mask, hw_thread_mask,
466 num_cores_per_socket *
467 node_affinity.num_online_nodes *
472 int hfi1_get_proc_affinity(int node)
474 int cpu = -1, ret, i;
475 struct hfi1_affinity_node *entry;
476 cpumask_var_t diff, hw_thread_mask, available_mask, intrs_mask;
477 const struct cpumask *node_mask,
478 *proc_mask = tsk_cpus_allowed(current);
479 struct hfi1_affinity_node_list *affinity = &node_affinity;
480 struct cpu_mask_set *set = &affinity->proc;
483 * check whether process/context affinity has already
486 if (cpumask_weight(proc_mask) == 1) {
487 hfi1_cdbg(PROC, "PID %u %s affinity set to CPU %*pbl",
488 current->pid, current->comm,
489 cpumask_pr_args(proc_mask));
491 * Mark the pre-set CPU as used. This is atomic so we don't
494 cpu = cpumask_first(proc_mask);
495 cpumask_set_cpu(cpu, &set->used);
497 } else if (cpumask_weight(proc_mask) < cpumask_weight(&set->mask)) {
498 hfi1_cdbg(PROC, "PID %u %s affinity set to CPU set(s) %*pbl",
499 current->pid, current->comm,
500 cpumask_pr_args(proc_mask));
505 * The process does not have a preset CPU affinity so find one to
506 * recommend using the following algorithm:
508 * For each user process that is opening a context on HFI Y:
509 * a) If all cores are filled, reinitialize the bitmask
510 * b) Fill real cores first, then HT cores (First set of HT
511 * cores on all physical cores, then second set of HT core,
512 * and, so on) in the following order:
514 * 1. Same NUMA node as HFI Y and not running an IRQ
516 * 2. Same NUMA node as HFI Y and running an IRQ handler
517 * 3. Different NUMA node to HFI Y and not running an IRQ
519 * 4. Different NUMA node to HFI Y and running an IRQ
521 * c) Mark core as filled in the bitmask. As user processes are
522 * done, clear cores from the bitmask.
525 ret = zalloc_cpumask_var(&diff, GFP_KERNEL);
528 ret = zalloc_cpumask_var(&hw_thread_mask, GFP_KERNEL);
531 ret = zalloc_cpumask_var(&available_mask, GFP_KERNEL);
533 goto free_hw_thread_mask;
534 ret = zalloc_cpumask_var(&intrs_mask, GFP_KERNEL);
536 goto free_available_mask;
538 spin_lock(&affinity->lock);
540 * If we've used all available HW threads, clear the mask and start
543 if (cpumask_equal(&set->mask, &set->used)) {
545 cpumask_clear(&set->used);
549 * If NUMA node has CPUs used by interrupt handlers, include them in the
550 * interrupt handler mask.
552 entry = node_affinity_lookup(node);
554 cpumask_copy(intrs_mask, (entry->def_intr.gen ?
555 &entry->def_intr.mask :
556 &entry->def_intr.used));
557 cpumask_or(intrs_mask, intrs_mask, (entry->rcv_intr.gen ?
558 &entry->rcv_intr.mask :
559 &entry->rcv_intr.used));
560 cpumask_or(intrs_mask, intrs_mask, &entry->general_intr_mask);
562 hfi1_cdbg(PROC, "CPUs used by interrupts: %*pbl",
563 cpumask_pr_args(intrs_mask));
565 cpumask_copy(hw_thread_mask, &set->mask);
568 * If HT cores are enabled, identify which HW threads within the
569 * physical cores should be used.
571 if (affinity->num_core_siblings > 0) {
572 for (i = 0; i < affinity->num_core_siblings; i++) {
573 find_hw_thread_mask(i, hw_thread_mask, affinity);
576 * If there's at least one available core for this HW
577 * thread number, stop looking for a core.
579 * diff will always be not empty at least once in this
580 * loop as the used mask gets reset when
581 * (set->mask == set->used) before this loop.
583 cpumask_andnot(diff, hw_thread_mask, &set->used);
584 if (!cpumask_empty(diff))
588 hfi1_cdbg(PROC, "Same available HW thread on all physical CPUs: %*pbl",
589 cpumask_pr_args(hw_thread_mask));
591 node_mask = cpumask_of_node(node);
592 hfi1_cdbg(PROC, "Device on NUMA %u, CPUs %*pbl", node,
593 cpumask_pr_args(node_mask));
595 /* Get cpumask of available CPUs on preferred NUMA */
596 cpumask_and(available_mask, hw_thread_mask, node_mask);
597 cpumask_andnot(available_mask, available_mask, &set->used);
598 hfi1_cdbg(PROC, "Available CPUs on NUMA %u: %*pbl", node,
599 cpumask_pr_args(available_mask));
602 * At first, we don't want to place processes on the same
603 * CPUs as interrupt handlers. Then, CPUs running interrupt
606 * 1) If diff is not empty, then there are CPUs not running
607 * non-interrupt handlers available, so diff gets copied
608 * over to available_mask.
609 * 2) If diff is empty, then all CPUs not running interrupt
610 * handlers are taken, so available_mask contains all
611 * available CPUs running interrupt handlers.
612 * 3) If available_mask is empty, then all CPUs on the
613 * preferred NUMA node are taken, so other NUMA nodes are
614 * used for process assignments using the same method as
615 * the preferred NUMA node.
617 cpumask_andnot(diff, available_mask, intrs_mask);
618 if (!cpumask_empty(diff))
619 cpumask_copy(available_mask, diff);
621 /* If we don't have CPUs on the preferred node, use other NUMA nodes */
622 if (cpumask_empty(available_mask)) {
623 cpumask_andnot(available_mask, hw_thread_mask, &set->used);
624 /* Excluding preferred NUMA cores */
625 cpumask_andnot(available_mask, available_mask, node_mask);
627 "Preferred NUMA node cores are taken, cores available in other NUMA nodes: %*pbl",
628 cpumask_pr_args(available_mask));
631 * At first, we don't want to place processes on the same
632 * CPUs as interrupt handlers.
634 cpumask_andnot(diff, available_mask, intrs_mask);
635 if (!cpumask_empty(diff))
636 cpumask_copy(available_mask, diff);
638 hfi1_cdbg(PROC, "Possible CPUs for process: %*pbl",
639 cpumask_pr_args(available_mask));
641 cpu = cpumask_first(available_mask);
642 if (cpu >= nr_cpu_ids) /* empty */
645 cpumask_set_cpu(cpu, &set->used);
646 spin_unlock(&affinity->lock);
647 hfi1_cdbg(PROC, "Process assigned to CPU %d", cpu);
649 free_cpumask_var(intrs_mask);
651 free_cpumask_var(available_mask);
653 free_cpumask_var(hw_thread_mask);
655 free_cpumask_var(diff);
660 void hfi1_put_proc_affinity(int cpu)
662 struct hfi1_affinity_node_list *affinity = &node_affinity;
663 struct cpu_mask_set *set = &affinity->proc;
667 spin_lock(&affinity->lock);
668 cpumask_clear_cpu(cpu, &set->used);
669 hfi1_cdbg(PROC, "Returning CPU %d for future process assignment", cpu);
670 if (cpumask_empty(&set->used) && set->gen) {
672 cpumask_copy(&set->used, &set->mask);
674 spin_unlock(&affinity->lock);
677 /* Prevents concurrent reads and writes of the sdma_affinity attrib */
678 static DEFINE_MUTEX(sdma_affinity_mutex);
680 int hfi1_set_sdma_affinity(struct hfi1_devdata *dd, const char *buf,
683 struct hfi1_affinity_node *entry;
687 spin_lock(&node_affinity.lock);
688 entry = node_affinity_lookup(dd->node);
689 spin_unlock(&node_affinity.lock);
694 ret = zalloc_cpumask_var(&mask, GFP_KERNEL);
698 ret = cpulist_parse(buf, mask);
702 if (!cpumask_subset(mask, cpu_online_mask) || cpumask_empty(mask)) {
703 dd_dev_warn(dd, "Invalid CPU mask\n");
708 mutex_lock(&sdma_affinity_mutex);
709 /* reset the SDMA interrupt affinity details */
710 init_cpu_mask_set(&entry->def_intr);
711 cpumask_copy(&entry->def_intr.mask, mask);
713 * Reassign the affinity for each SDMA interrupt.
715 for (i = 0; i < dd->num_msix_entries; i++) {
716 struct hfi1_msix_entry *msix;
718 msix = &dd->msix_entries[i];
719 if (msix->type != IRQ_SDMA)
722 ret = hfi1_get_irq_affinity(dd, msix);
727 mutex_unlock(&sdma_affinity_mutex);
729 free_cpumask_var(mask);
730 return ret ? ret : strnlen(buf, PAGE_SIZE);
733 int hfi1_get_sdma_affinity(struct hfi1_devdata *dd, char *buf)
735 struct hfi1_affinity_node *entry;
737 spin_lock(&node_affinity.lock);
738 entry = node_affinity_lookup(dd->node);
739 spin_unlock(&node_affinity.lock);
744 mutex_lock(&sdma_affinity_mutex);
745 cpumap_print_to_pagebuf(true, buf, &entry->def_intr.mask);
746 mutex_unlock(&sdma_affinity_mutex);
747 return strnlen(buf, PAGE_SIZE);