kernel/irq/cpuhotplug.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Generic cpu hotunplug interrupt migration code copied from the
   4  * arch/arm implementation
   5  *
   6  * Copyright (C) Russell King
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 as
  10  * published by the Free Software Foundation.
  11  */
  12 #include <linux/interrupt.h>
  13 #include <linux/ratelimit.h>
  14 #include <linux/irq.h>
  15 #include <linux/sched/isolation.h>
  16
  17 #include "internals.h"
  18
  19 /* For !GENERIC_IRQ_EFFECTIVE_AFF_MASK this looks at general affinity mask */
  20 static inline bool irq_needs_fixup(struct irq_data *d)
  21 {
  22         const struct cpumask *m = irq_data_get_effective_affinity_mask(d);
  23         unsigned int cpu = smp_processor_id();
  24
  25 #ifdef CONFIG_GENERIC_IRQ_EFFECTIVE_AFF_MASK
  26         /*
  27          * The cpumask_empty() check is a workaround for interrupt chips,
  28          * which do not implement effective affinity, but the architecture has
  29          * enabled the config switch. Use the general affinity mask instead.
  30          */
  31         if (cpumask_empty(m))
  32                 m = irq_data_get_affinity_mask(d);
  33
  34         /*
  35          * Sanity check. If the mask is not empty when excluding the outgoing
  36          * CPU then it must contain at least one online CPU. The outgoing CPU
  37          * has been removed from the online mask already.
  38          */
  39         if (cpumask_any_but(m, cpu) < nr_cpu_ids &&
  40             !cpumask_intersects(m, cpu_online_mask)) {
  41                 /*
  42                  * If this happens then there was a missed IRQ fixup at some
  43                  * point. Warn about it and enforce fixup.
  44                  */
  45                 pr_warn("Eff. affinity %*pbl of IRQ %u contains only offline CPUs after offlining CPU %u\n",
  46                         cpumask_pr_args(m), d->irq, cpu);
  47                 return true;
  48         }
  49 #endif
  50         return cpumask_test_cpu(cpu, m);
  51 }
  52
  53 static bool migrate_one_irq(struct irq_desc *desc)
  54 {
  55         struct irq_data *d = irq_desc_get_irq_data(desc);
  56         struct irq_chip *chip = irq_data_get_irq_chip(d);
  57         bool maskchip = !irq_can_move_pcntxt(d) && !irqd_irq_masked(d);
  58         const struct cpumask *affinity;
  59         bool brokeaff = false;
  60         int err;
  61
  62         /*
  63          * IRQ chip might be already torn down, but the irq descriptor is
  64          * still in the radix tree. Also if the chip has no affinity setter,
  65          * nothing can be done here.
  66          */
  67         if (!chip || !chip->irq_set_affinity) {
  68                 pr_debug("IRQ %u: Unable to migrate away\n", d->irq);
  69                 return false;
  70         }
  71
  72         /*
  73          * Complete an eventually pending irq move cleanup. If this
  74          * interrupt was moved in hard irq context, then the vectors need
  75          * to be cleaned up. It can't wait until this interrupt actually
  76          * happens and this CPU was involved.
  77          */
  78         irq_force_complete_move(desc);
  79
  80         /*
  81          * No move required, if:
  82          * - Interrupt is per cpu
  83          * - Interrupt is not started
  84          * - Affinity mask does not include this CPU.
  85          *
  86          * Note: Do not check desc->action as this might be a chained
  87          * interrupt.
  88          */
  89         if (irqd_is_per_cpu(d) || !irqd_is_started(d) || !irq_needs_fixup(d)) {
  90                 /*
  91                  * If an irq move is pending, abort it if the dying CPU is
  92                  * the sole target.
  93                  */
  94                 irq_fixup_move_pending(desc, false);
  95                 return false;
  96         }
  97
  98         /*
  99          * If there is a setaffinity pending, then try to reuse the pending
 100          * mask, so the last change of the affinity does not get lost. If
 101          * there is no move pending or the pending mask does not contain
 102          * any online CPU, use the current affinity mask.
 103          */
 104         if (irq_fixup_move_pending(desc, true))
 105                 affinity = irq_desc_get_pending_mask(desc);
 106         else
 107                 affinity = irq_data_get_affinity_mask(d);
 108
 109         /* Mask the chip for interrupts which cannot move in process context */
 110         if (maskchip && chip->irq_mask)
 111                 chip->irq_mask(d);
 112
 113         if (!cpumask_intersects(affinity, cpu_online_mask)) {
 114                 /*
 115                  * If the interrupt is managed, then shut it down and leave
 116                  * the affinity untouched.
 117                  */
 118                 if (irqd_affinity_is_managed(d)) {
 119                         irqd_set_managed_shutdown(d);
 120                         irq_shutdown_and_deactivate(desc);
 121                         return false;
 122                 }
 123                 affinity = cpu_online_mask;
 124                 brokeaff = true;
 125         }
 126         /*
 127          * Do not set the force argument of irq_do_set_affinity() as this
 128          * disables the masking of offline CPUs from the supplied affinity
 129          * mask and therefore might keep/reassign the irq to the outgoing
 130          * CPU.
 131          */
 132         err = irq_do_set_affinity(d, affinity, false);
 133
 134         /*
 135          * If there are online CPUs in the affinity mask, but they have no
 136          * vectors left to make the migration work, try to break the
 137          * affinity by migrating to any online CPU.
 138          */
 139         if (err == -ENOSPC && !irqd_affinity_is_managed(d) && affinity != cpu_online_mask) {
 140                 pr_debug("IRQ%u: set affinity failed for %*pbl, re-try with online CPUs\n",
 141                          d->irq, cpumask_pr_args(affinity));
 142
 143                 affinity = cpu_online_mask;
 144                 brokeaff = true;
 145
 146                 err = irq_do_set_affinity(d, affinity, false);
 147         }
 148
 149         if (err) {
 150                 pr_warn_ratelimited("IRQ%u: set affinity failed(%d).\n",
 151                                     d->irq, err);
 152                 brokeaff = false;
 153         }
 154
 155         if (maskchip && chip->irq_unmask)
 156                 chip->irq_unmask(d);
 157
 158         return brokeaff;
 159 }
 160
 161 /**
 162  * irq_migrate_all_off_this_cpu - Migrate irqs away from offline cpu
 163  *
 164  * The current CPU has been marked offline.  Migrate IRQs off this CPU.
 165  * If the affinity settings do not allow other CPUs, force them onto any
 166  * available CPU.
 167  *
 168  * Note: we must iterate over all IRQs, whether they have an attached
 169  * action structure or not, as we need to get chained interrupts too.
 170  */
 171 void irq_migrate_all_off_this_cpu(void)
 172 {
 173         struct irq_desc *desc;
 174         unsigned int irq;
 175
 176         for_each_active_irq(irq) {
 177                 bool affinity_broken;
 178
 179                 desc = irq_to_desc(irq);
 180                 raw_spin_lock(&desc->lock);
 181                 affinity_broken = migrate_one_irq(desc);
 182                 raw_spin_unlock(&desc->lock);
 183
 184                 if (affinity_broken) {
 185                         pr_debug_ratelimited("IRQ %u: no longer affine to CPU%u\n",
 186                                             irq, smp_processor_id());
 187                 }
 188         }
 189 }
 190
 191 static bool hk_should_isolate(struct irq_data *data, unsigned int cpu)
 192 {
 193         const struct cpumask *hk_mask;
 194
 195         if (!housekeeping_enabled(HK_TYPE_MANAGED_IRQ))
 196                 return false;
 197
 198         hk_mask = housekeeping_cpumask(HK_TYPE_MANAGED_IRQ);
 199         if (cpumask_subset(irq_data_get_effective_affinity_mask(data), hk_mask))
 200                 return false;
 201
 202         return cpumask_test_cpu(cpu, hk_mask);
 203 }
 204
 205 static void irq_restore_affinity_of_irq(struct irq_desc *desc, unsigned int cpu)
 206 {
 207         struct irq_data *data = irq_desc_get_irq_data(desc);
 208         const struct cpumask *affinity = irq_data_get_affinity_mask(data);
 209
 210         if (!irqd_affinity_is_managed(data) || !desc->action ||
 211             !irq_data_get_irq_chip(data) || !cpumask_test_cpu(cpu, affinity))
 212                 return;
 213
 214         /*
 215          * Don't restore suspended interrupts here when a system comes back
 216          * from S3. They are reenabled via resume_device_irqs().
 217          */
 218         if (desc->istate & IRQS_SUSPENDED)
 219                 return;
 220
 221         if (irqd_is_managed_and_shutdown(data))
 222                 irq_startup(desc, IRQ_RESEND, IRQ_START_COND);
 223
 224         /*
 225          * If the interrupt can only be directed to a single target
 226          * CPU then it is already assigned to a CPU in the affinity
 227          * mask. No point in trying to move it around unless the
 228          * isolation mechanism requests to move it to an upcoming
 229          * housekeeping CPU.
 230          */
 231         if (!irqd_is_single_target(data) || hk_should_isolate(data, cpu))
 232                 irq_set_affinity_locked(data, affinity, false);
 233 }
 234
 235 /**
 236  * irq_affinity_online_cpu - Restore affinity for managed interrupts
 237  * @cpu:        Upcoming CPU for which interrupts should be restored
 238  */
 239 int irq_affinity_online_cpu(unsigned int cpu)
 240 {
 241         struct irq_desc *desc;
 242         unsigned int irq;
 243
 244         irq_lock_sparse();
 245         for_each_active_irq(irq) {
 246                 desc = irq_to_desc(irq);
 247                 raw_spin_lock_irq(&desc->lock);
 248                 irq_restore_affinity_of_irq(desc, cpu);
 249                 raw_spin_unlock_irq(&desc->lock);
 250         }
 251         irq_unlock_sparse();
 252
 253         return 0;
 254 }