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>
50 #include <linux/cpumask.h>
57 struct hfi1_affinity_node_list node_affinity = {
58 .list = LIST_HEAD_INIT(node_affinity.list),
59 .lock = __SPIN_LOCK_UNLOCKED(&node_affinity.lock),
62 /* Name of IRQ types, indexed by enum irq_type */
63 static const char * const irq_type_names[] = {
70 /* Per NUMA node count of HFI devices */
71 static unsigned int *hfi1_per_node_cntr;
73 static inline void init_cpu_mask_set(struct cpu_mask_set *set)
75 cpumask_clear(&set->mask);
76 cpumask_clear(&set->used);
80 /* Initialize non-HT cpu cores mask */
81 void init_real_cpu_mask(void)
83 int possible, curr_cpu, i, ht;
85 cpumask_clear(&node_affinity.real_cpu_mask);
87 /* Start with cpu online mask as the real cpu mask */
88 cpumask_copy(&node_affinity.real_cpu_mask, cpu_online_mask);
91 * Remove HT cores from the real cpu mask. Do this in two steps below.
93 possible = cpumask_weight(&node_affinity.real_cpu_mask);
94 ht = cpumask_weight(topology_sibling_cpumask(
95 cpumask_first(&node_affinity.real_cpu_mask)));
97 * Step 1. Skip over the first N HT siblings and use them as the
98 * "real" cores. Assumes that HT cores are not enumerated in
99 * succession (except in the single core case).
101 curr_cpu = cpumask_first(&node_affinity.real_cpu_mask);
102 for (i = 0; i < possible / ht; i++)
103 curr_cpu = cpumask_next(curr_cpu, &node_affinity.real_cpu_mask);
105 * Step 2. Remove the remaining HT siblings. Use cpumask_next() to
108 for (; i < possible; i++) {
109 cpumask_clear_cpu(curr_cpu, &node_affinity.real_cpu_mask);
110 curr_cpu = cpumask_next(curr_cpu, &node_affinity.real_cpu_mask);
114 int node_affinity_init(void)
117 struct pci_dev *dev = NULL;
118 const struct pci_device_id *ids = hfi1_pci_tbl;
120 cpumask_clear(&node_affinity.proc.used);
121 cpumask_copy(&node_affinity.proc.mask, cpu_online_mask);
123 node_affinity.proc.gen = 0;
124 node_affinity.num_core_siblings =
125 cpumask_weight(topology_sibling_cpumask(
126 cpumask_first(&node_affinity.proc.mask)
128 node_affinity.num_online_nodes = num_online_nodes();
129 node_affinity.num_online_cpus = num_online_cpus();
132 * The real cpu mask is part of the affinity struct but it has to be
133 * initialized early. It is needed to calculate the number of user
134 * contexts in set_up_context_variables().
136 init_real_cpu_mask();
138 hfi1_per_node_cntr = kcalloc(num_possible_nodes(),
139 sizeof(*hfi1_per_node_cntr), GFP_KERNEL);
140 if (!hfi1_per_node_cntr)
143 while (ids->vendor) {
145 while ((dev = pci_get_device(ids->vendor, ids->device, dev))) {
146 node = pcibus_to_node(dev->bus);
148 node = numa_node_id();
150 hfi1_per_node_cntr[node]++;
158 void node_affinity_destroy(void)
160 struct list_head *pos, *q;
161 struct hfi1_affinity_node *entry;
163 spin_lock(&node_affinity.lock);
164 list_for_each_safe(pos, q, &node_affinity.list) {
165 entry = list_entry(pos, struct hfi1_affinity_node,
170 spin_unlock(&node_affinity.lock);
171 kfree(hfi1_per_node_cntr);
174 static struct hfi1_affinity_node *node_affinity_allocate(int node)
176 struct hfi1_affinity_node *entry;
178 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
182 INIT_LIST_HEAD(&entry->list);
188 * It appends an entry to the list.
189 * It *must* be called with node_affinity.lock held.
191 static void node_affinity_add_tail(struct hfi1_affinity_node *entry)
193 list_add_tail(&entry->list, &node_affinity.list);
196 /* It must be called with node_affinity.lock held */
197 static struct hfi1_affinity_node *node_affinity_lookup(int node)
199 struct list_head *pos;
200 struct hfi1_affinity_node *entry;
202 list_for_each(pos, &node_affinity.list) {
203 entry = list_entry(pos, struct hfi1_affinity_node, list);
204 if (entry->node == node)
212 * Interrupt affinity.
214 * non-rcv avail gets a default mask that
215 * starts as possible cpus with threads reset
216 * and each rcv avail reset.
218 * rcv avail gets node relative 1 wrapping back
219 * to the node relative 1 as necessary.
222 int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
224 int node = pcibus_to_node(dd->pcidev->bus);
225 struct hfi1_affinity_node *entry;
226 const struct cpumask *local_mask;
227 int curr_cpu, possible, i;
230 node = numa_node_id();
233 local_mask = cpumask_of_node(dd->node);
234 if (cpumask_first(local_mask) >= nr_cpu_ids)
235 local_mask = topology_core_cpumask(0);
237 spin_lock(&node_affinity.lock);
238 entry = node_affinity_lookup(dd->node);
239 spin_unlock(&node_affinity.lock);
242 * If this is the first time this NUMA node's affinity is used,
243 * create an entry in the global affinity structure and initialize it.
246 entry = node_affinity_allocate(node);
249 "Unable to allocate global affinity node\n");
252 init_cpu_mask_set(&entry->def_intr);
253 init_cpu_mask_set(&entry->rcv_intr);
254 cpumask_clear(&entry->general_intr_mask);
255 /* Use the "real" cpu mask of this node as the default */
256 cpumask_and(&entry->def_intr.mask, &node_affinity.real_cpu_mask,
259 /* fill in the receive list */
260 possible = cpumask_weight(&entry->def_intr.mask);
261 curr_cpu = cpumask_first(&entry->def_intr.mask);
264 /* only one CPU, everyone will use it */
265 cpumask_set_cpu(curr_cpu, &entry->rcv_intr.mask);
266 cpumask_set_cpu(curr_cpu, &entry->general_intr_mask);
269 * The general/control context will be the first CPU in
270 * the default list, so it is removed from the default
271 * list and added to the general interrupt list.
273 cpumask_clear_cpu(curr_cpu, &entry->def_intr.mask);
274 cpumask_set_cpu(curr_cpu, &entry->general_intr_mask);
275 curr_cpu = cpumask_next(curr_cpu,
276 &entry->def_intr.mask);
279 * Remove the remaining kernel receive queues from
280 * the default list and add them to the receive list.
283 i < (dd->n_krcv_queues - 1) *
284 hfi1_per_node_cntr[dd->node];
286 cpumask_clear_cpu(curr_cpu,
287 &entry->def_intr.mask);
288 cpumask_set_cpu(curr_cpu,
289 &entry->rcv_intr.mask);
290 curr_cpu = cpumask_next(curr_cpu,
291 &entry->def_intr.mask);
292 if (curr_cpu >= nr_cpu_ids)
297 * If there ends up being 0 CPU cores leftover for SDMA
298 * engines, use the same CPU cores as general/control
301 if (cpumask_weight(&entry->def_intr.mask) == 0)
302 cpumask_copy(&entry->def_intr.mask,
303 &entry->general_intr_mask);
306 spin_lock(&node_affinity.lock);
307 node_affinity_add_tail(entry);
308 spin_unlock(&node_affinity.lock);
314 int hfi1_get_irq_affinity(struct hfi1_devdata *dd, struct hfi1_msix_entry *msix)
318 struct hfi1_affinity_node *entry;
319 struct cpu_mask_set *set = NULL;
320 struct sdma_engine *sde = NULL;
321 struct hfi1_ctxtdata *rcd = NULL;
326 cpumask_clear(&msix->mask);
328 ret = zalloc_cpumask_var(&diff, GFP_KERNEL);
332 spin_lock(&node_affinity.lock);
333 entry = node_affinity_lookup(dd->node);
334 spin_unlock(&node_affinity.lock);
336 switch (msix->type) {
338 sde = (struct sdma_engine *)msix->arg;
339 scnprintf(extra, 64, "engine %u", sde->this_idx);
340 set = &entry->def_intr;
343 cpu = cpumask_first(&entry->general_intr_mask);
346 rcd = (struct hfi1_ctxtdata *)msix->arg;
347 if (rcd->ctxt == HFI1_CTRL_CTXT)
348 cpu = cpumask_first(&entry->general_intr_mask);
350 set = &entry->rcv_intr;
351 scnprintf(extra, 64, "ctxt %u", rcd->ctxt);
354 dd_dev_err(dd, "Invalid IRQ type %d\n", msix->type);
359 * The general and control contexts are placed on a particular
360 * CPU, which is set above. Skip accounting for it. Everything else
361 * finds its CPU here.
363 if (cpu == -1 && set) {
364 spin_lock(&node_affinity.lock);
365 if (cpumask_equal(&set->mask, &set->used)) {
367 * We've used up all the CPUs, bump up the generation
368 * and reset the 'used' map
371 cpumask_clear(&set->used);
373 cpumask_andnot(diff, &set->mask, &set->used);
374 cpu = cpumask_first(diff);
375 cpumask_set_cpu(cpu, &set->used);
376 spin_unlock(&node_affinity.lock);
379 switch (msix->type) {
389 cpumask_set_cpu(cpu, &msix->mask);
390 dd_dev_info(dd, "IRQ vector: %u, type %s %s -> cpu: %d\n",
391 msix->msix.vector, irq_type_names[msix->type],
393 irq_set_affinity_hint(msix->msix.vector, &msix->mask);
395 free_cpumask_var(diff);
399 void hfi1_put_irq_affinity(struct hfi1_devdata *dd,
400 struct hfi1_msix_entry *msix)
402 struct cpu_mask_set *set = NULL;
403 struct hfi1_ctxtdata *rcd;
404 struct hfi1_affinity_node *entry;
406 spin_lock(&node_affinity.lock);
407 entry = node_affinity_lookup(dd->node);
408 spin_unlock(&node_affinity.lock);
410 switch (msix->type) {
412 set = &entry->def_intr;
415 /* Don't do accounting for general contexts */
418 rcd = (struct hfi1_ctxtdata *)msix->arg;
419 /* Don't do accounting for control contexts */
420 if (rcd->ctxt != HFI1_CTRL_CTXT)
421 set = &entry->rcv_intr;
428 spin_lock(&node_affinity.lock);
429 cpumask_andnot(&set->used, &set->used, &msix->mask);
430 if (cpumask_empty(&set->used) && set->gen) {
432 cpumask_copy(&set->used, &set->mask);
434 spin_unlock(&node_affinity.lock);
437 irq_set_affinity_hint(msix->msix.vector, NULL);
438 cpumask_clear(&msix->mask);
441 /* This should be called with node_affinity.lock held */
442 static void find_hw_thread_mask(uint hw_thread_no, cpumask_var_t hw_thread_mask,
443 struct hfi1_affinity_node_list *affinity)
445 int possible, curr_cpu, i;
446 uint num_cores_per_socket = node_affinity.num_online_cpus /
447 affinity->num_core_siblings /
448 node_affinity.num_online_nodes;
450 cpumask_copy(hw_thread_mask, &affinity->proc.mask);
451 if (affinity->num_core_siblings > 0) {
452 /* Removing other siblings not needed for now */
453 possible = cpumask_weight(hw_thread_mask);
454 curr_cpu = cpumask_first(hw_thread_mask);
456 i < num_cores_per_socket * node_affinity.num_online_nodes;
458 curr_cpu = cpumask_next(curr_cpu, hw_thread_mask);
460 for (; i < possible; i++) {
461 cpumask_clear_cpu(curr_cpu, hw_thread_mask);
462 curr_cpu = cpumask_next(curr_cpu, hw_thread_mask);
465 /* Identifying correct HW threads within physical cores */
466 cpumask_shift_left(hw_thread_mask, hw_thread_mask,
467 num_cores_per_socket *
468 node_affinity.num_online_nodes *
473 int hfi1_get_proc_affinity(int node)
475 int cpu = -1, ret, i;
476 struct hfi1_affinity_node *entry;
477 cpumask_var_t diff, hw_thread_mask, available_mask, intrs_mask;
478 const struct cpumask *node_mask,
479 *proc_mask = tsk_cpus_allowed(current);
480 struct hfi1_affinity_node_list *affinity = &node_affinity;
481 struct cpu_mask_set *set = &affinity->proc;
484 * check whether process/context affinity has already
487 if (cpumask_weight(proc_mask) == 1) {
488 hfi1_cdbg(PROC, "PID %u %s affinity set to CPU %*pbl",
489 current->pid, current->comm,
490 cpumask_pr_args(proc_mask));
492 * Mark the pre-set CPU as used. This is atomic so we don't
495 cpu = cpumask_first(proc_mask);
496 cpumask_set_cpu(cpu, &set->used);
498 } else if (cpumask_weight(proc_mask) < cpumask_weight(&set->mask)) {
499 hfi1_cdbg(PROC, "PID %u %s affinity set to CPU set(s) %*pbl",
500 current->pid, current->comm,
501 cpumask_pr_args(proc_mask));
506 * The process does not have a preset CPU affinity so find one to
507 * recommend using the following algorithm:
509 * For each user process that is opening a context on HFI Y:
510 * a) If all cores are filled, reinitialize the bitmask
511 * b) Fill real cores first, then HT cores (First set of HT
512 * cores on all physical cores, then second set of HT core,
513 * and, so on) in the following order:
515 * 1. Same NUMA node as HFI Y and not running an IRQ
517 * 2. Same NUMA node as HFI Y and running an IRQ handler
518 * 3. Different NUMA node to HFI Y and not running an IRQ
520 * 4. Different NUMA node to HFI Y and running an IRQ
522 * c) Mark core as filled in the bitmask. As user processes are
523 * done, clear cores from the bitmask.
526 ret = zalloc_cpumask_var(&diff, GFP_KERNEL);
529 ret = zalloc_cpumask_var(&hw_thread_mask, GFP_KERNEL);
532 ret = zalloc_cpumask_var(&available_mask, GFP_KERNEL);
534 goto free_hw_thread_mask;
535 ret = zalloc_cpumask_var(&intrs_mask, GFP_KERNEL);
537 goto free_available_mask;
539 spin_lock(&affinity->lock);
541 * If we've used all available HW threads, clear the mask and start
544 if (cpumask_equal(&set->mask, &set->used)) {
546 cpumask_clear(&set->used);
550 * If NUMA node has CPUs used by interrupt handlers, include them in the
551 * interrupt handler mask.
553 entry = node_affinity_lookup(node);
555 cpumask_copy(intrs_mask, (entry->def_intr.gen ?
556 &entry->def_intr.mask :
557 &entry->def_intr.used));
558 cpumask_or(intrs_mask, intrs_mask, (entry->rcv_intr.gen ?
559 &entry->rcv_intr.mask :
560 &entry->rcv_intr.used));
561 cpumask_or(intrs_mask, intrs_mask, &entry->general_intr_mask);
563 hfi1_cdbg(PROC, "CPUs used by interrupts: %*pbl",
564 cpumask_pr_args(intrs_mask));
566 cpumask_copy(hw_thread_mask, &set->mask);
569 * If HT cores are enabled, identify which HW threads within the
570 * physical cores should be used.
572 if (affinity->num_core_siblings > 0) {
573 for (i = 0; i < affinity->num_core_siblings; i++) {
574 find_hw_thread_mask(i, hw_thread_mask, affinity);
577 * If there's at least one available core for this HW
578 * thread number, stop looking for a core.
580 * diff will always be not empty at least once in this
581 * loop as the used mask gets reset when
582 * (set->mask == set->used) before this loop.
584 cpumask_andnot(diff, hw_thread_mask, &set->used);
585 if (!cpumask_empty(diff))
589 hfi1_cdbg(PROC, "Same available HW thread on all physical CPUs: %*pbl",
590 cpumask_pr_args(hw_thread_mask));
592 node_mask = cpumask_of_node(node);
593 hfi1_cdbg(PROC, "Device on NUMA %u, CPUs %*pbl", node,
594 cpumask_pr_args(node_mask));
596 /* Get cpumask of available CPUs on preferred NUMA */
597 cpumask_and(available_mask, hw_thread_mask, node_mask);
598 cpumask_andnot(available_mask, available_mask, &set->used);
599 hfi1_cdbg(PROC, "Available CPUs on NUMA %u: %*pbl", node,
600 cpumask_pr_args(available_mask));
603 * At first, we don't want to place processes on the same
604 * CPUs as interrupt handlers. Then, CPUs running interrupt
607 * 1) If diff is not empty, then there are CPUs not running
608 * non-interrupt handlers available, so diff gets copied
609 * over to available_mask.
610 * 2) If diff is empty, then all CPUs not running interrupt
611 * handlers are taken, so available_mask contains all
612 * available CPUs running interrupt handlers.
613 * 3) If available_mask is empty, then all CPUs on the
614 * preferred NUMA node are taken, so other NUMA nodes are
615 * used for process assignments using the same method as
616 * the preferred NUMA node.
618 cpumask_andnot(diff, available_mask, intrs_mask);
619 if (!cpumask_empty(diff))
620 cpumask_copy(available_mask, diff);
622 /* If we don't have CPUs on the preferred node, use other NUMA nodes */
623 if (cpumask_empty(available_mask)) {
624 cpumask_andnot(available_mask, hw_thread_mask, &set->used);
625 /* Excluding preferred NUMA cores */
626 cpumask_andnot(available_mask, available_mask, node_mask);
628 "Preferred NUMA node cores are taken, cores available in other NUMA nodes: %*pbl",
629 cpumask_pr_args(available_mask));
632 * At first, we don't want to place processes on the same
633 * CPUs as interrupt handlers.
635 cpumask_andnot(diff, available_mask, intrs_mask);
636 if (!cpumask_empty(diff))
637 cpumask_copy(available_mask, diff);
639 hfi1_cdbg(PROC, "Possible CPUs for process: %*pbl",
640 cpumask_pr_args(available_mask));
642 cpu = cpumask_first(available_mask);
643 if (cpu >= nr_cpu_ids) /* empty */
646 cpumask_set_cpu(cpu, &set->used);
647 spin_unlock(&affinity->lock);
648 hfi1_cdbg(PROC, "Process assigned to CPU %d", cpu);
650 free_cpumask_var(intrs_mask);
652 free_cpumask_var(available_mask);
654 free_cpumask_var(hw_thread_mask);
656 free_cpumask_var(diff);
661 void hfi1_put_proc_affinity(int cpu)
663 struct hfi1_affinity_node_list *affinity = &node_affinity;
664 struct cpu_mask_set *set = &affinity->proc;
668 spin_lock(&affinity->lock);
669 cpumask_clear_cpu(cpu, &set->used);
670 hfi1_cdbg(PROC, "Returning CPU %d for future process assignment", cpu);
671 if (cpumask_empty(&set->used) && set->gen) {
673 cpumask_copy(&set->used, &set->mask);
675 spin_unlock(&affinity->lock);
678 /* Prevents concurrent reads and writes of the sdma_affinity attrib */
679 static DEFINE_MUTEX(sdma_affinity_mutex);
681 int hfi1_set_sdma_affinity(struct hfi1_devdata *dd, const char *buf,
684 struct hfi1_affinity_node *entry;
688 spin_lock(&node_affinity.lock);
689 entry = node_affinity_lookup(dd->node);
690 spin_unlock(&node_affinity.lock);
695 ret = cpulist_parse(buf, &mask);
699 if (!cpumask_subset(&mask, cpu_online_mask) || cpumask_empty(&mask)) {
700 dd_dev_warn(dd, "Invalid CPU mask\n");
704 mutex_lock(&sdma_affinity_mutex);
705 /* reset the SDMA interrupt affinity details */
706 init_cpu_mask_set(&entry->def_intr);
707 cpumask_copy(&entry->def_intr.mask, &mask);
709 * Reassign the affinity for each SDMA interrupt.
711 for (i = 0; i < dd->num_msix_entries; i++) {
712 struct hfi1_msix_entry *msix;
714 msix = &dd->msix_entries[i];
715 if (msix->type != IRQ_SDMA)
718 ret = hfi1_get_irq_affinity(dd, msix);
724 mutex_unlock(&sdma_affinity_mutex);
725 return ret ? ret : strnlen(buf, PAGE_SIZE);
728 int hfi1_get_sdma_affinity(struct hfi1_devdata *dd, char *buf)
730 struct hfi1_affinity_node *entry;
732 spin_lock(&node_affinity.lock);
733 entry = node_affinity_lookup(dd->node);
734 spin_unlock(&node_affinity.lock);
739 mutex_lock(&sdma_affinity_mutex);
740 cpumap_print_to_pagebuf(true, buf, &entry->def_intr.mask);
741 mutex_unlock(&sdma_affinity_mutex);
742 return strnlen(buf, PAGE_SIZE);