arch/x86/kernel/apic/x2apic_cluster.c

   1 #include <linux/threads.h>
   2 #include <linux/cpumask.h>
   3 #include <linux/string.h>
   4 #include <linux/kernel.h>
   5 #include <linux/ctype.h>
   6 #include <linux/dmar.h>
   7 #include <linux/cpu.h>
   8
   9 #include <asm/smp.h>
  10 #include <asm/x2apic.h>
  11
  12 static DEFINE_PER_CPU(u32, x86_cpu_to_logical_apicid);
  13 static DEFINE_PER_CPU(cpumask_var_t, cpus_in_cluster);
  14 static DEFINE_PER_CPU(cpumask_var_t, ipi_mask);
  15
  16 static int x2apic_acpi_madt_oem_check(char *oem_id, char *oem_table_id)
  17 {
  18         return x2apic_enabled();
  19 }
  20
  21 static inline u32 x2apic_cluster(int cpu)
  22 {
  23         return per_cpu(x86_cpu_to_logical_apicid, cpu) >> 16;
  24 }
  25
  26 static void
  27 __x2apic_send_IPI_mask(const struct cpumask *mask, int vector, int apic_dest)
  28 {
  29         struct cpumask *cpus_in_cluster_ptr;
  30         struct cpumask *ipi_mask_ptr;
  31         unsigned int cpu, this_cpu;
  32         unsigned long flags;
  33         u32 dest;
  34
  35         x2apic_wrmsr_fence();
  36
  37         local_irq_save(flags);
  38
  39         this_cpu = smp_processor_id();
  40
  41         /*
  42          * We are to modify mask, so we need an own copy
  43          * and be sure it's manipulated with irq off.
  44          */
  45         ipi_mask_ptr = this_cpu_cpumask_var_ptr(ipi_mask);
  46
  47         /*
  48          * The idea is to send one IPI per cluster.
  49          */
  50         for_each_cpu(cpu, ipi_mask_ptr) {
  51                 unsigned long i;
  52
  53                 cpus_in_cluster_ptr = per_cpu(cpus_in_cluster, cpu);
  54                 dest = 0;
  55
  56                 /* Collect cpus in cluster. */
  57                 for_each_cpu_and(i, ipi_mask_ptr, cpus_in_cluster_ptr) {
  58                         if (apic_dest == APIC_DEST_ALLINC || i != this_cpu)
  59                                 dest |= per_cpu(x86_cpu_to_logical_apicid, i);
  60                 }
  61
  62                 if (!dest)
  63                         continue;
  64
  65                 __x2apic_send_IPI_dest(dest, vector, apic->dest_logical);
  66                 /*
  67                  * Cluster sibling cpus should be discared now so
  68                  * we would not send IPI them second time.
  69                  */
  70                 cpumask_andnot(ipi_mask_ptr, ipi_mask_ptr, cpus_in_cluster_ptr);
  71         }
  72
  73         local_irq_restore(flags);
  74 }
  75
  76 static void x2apic_send_IPI_mask(const struct cpumask *mask, int vector)
  77 {
  78         __x2apic_send_IPI_mask(mask, vector, APIC_DEST_ALLINC);
  79 }
  80
  81 static void
  82 x2apic_send_IPI_mask_allbutself(const struct cpumask *mask, int vector)
  83 {
  84         __x2apic_send_IPI_mask(mask, vector, APIC_DEST_ALLBUT);
  85 }
  86
  87 static void x2apic_send_IPI_allbutself(int vector)
  88 {
  89         __x2apic_send_IPI_mask(cpu_online_mask, vector, APIC_DEST_ALLBUT);
  90 }
  91
  92 static void x2apic_send_IPI_all(int vector)
  93 {
  94         __x2apic_send_IPI_mask(cpu_online_mask, vector, APIC_DEST_ALLINC);
  95 }
  96
  97 static int
  98 x2apic_cpu_mask_to_apicid_and(const struct cpumask *cpumask,
  99                               const struct cpumask *andmask,
 100                               unsigned int *apicid)
 101 {
 102         u32 dest = 0;
 103         u16 cluster;
 104         int i;
 105
 106         for_each_cpu_and(i, cpumask, andmask) {
 107                 if (!cpumask_test_cpu(i, cpu_online_mask))
 108                         continue;
 109                 dest = per_cpu(x86_cpu_to_logical_apicid, i);
 110                 cluster = x2apic_cluster(i);
 111                 break;
 112         }
 113
 114         if (!dest)
 115                 return -EINVAL;
 116
 117         for_each_cpu_and(i, cpumask, andmask) {
 118                 if (!cpumask_test_cpu(i, cpu_online_mask))
 119                         continue;
 120                 if (cluster != x2apic_cluster(i))
 121                         continue;
 122                 dest |= per_cpu(x86_cpu_to_logical_apicid, i);
 123         }
 124
 125         *apicid = dest;
 126
 127         return 0;
 128 }
 129
 130 static void init_x2apic_ldr(void)
 131 {
 132         unsigned int this_cpu = smp_processor_id();
 133         unsigned int cpu;
 134
 135         per_cpu(x86_cpu_to_logical_apicid, this_cpu) = apic_read(APIC_LDR);
 136
 137         __cpu_set(this_cpu, per_cpu(cpus_in_cluster, this_cpu));
 138         for_each_online_cpu(cpu) {
 139                 if (x2apic_cluster(this_cpu) != x2apic_cluster(cpu))
 140                         continue;
 141                 __cpu_set(this_cpu, per_cpu(cpus_in_cluster, cpu));
 142                 __cpu_set(cpu, per_cpu(cpus_in_cluster, this_cpu));
 143         }
 144 }
 145
 146  /*
 147   * At CPU state changes, update the x2apic cluster sibling info.
 148   */
 149 static int
 150 update_clusterinfo(struct notifier_block *nfb, unsigned long action, void *hcpu)
 151 {
 152         unsigned int this_cpu = (unsigned long)hcpu;
 153         unsigned int cpu;
 154         int err = 0;
 155
 156         switch (action) {
 157         case CPU_UP_PREPARE:
 158                 if (!zalloc_cpumask_var(&per_cpu(cpus_in_cluster, this_cpu),
 159                                         GFP_KERNEL)) {
 160                         err = -ENOMEM;
 161                 } else if (!zalloc_cpumask_var(&per_cpu(ipi_mask, this_cpu),
 162                                                GFP_KERNEL)) {
 163                         free_cpumask_var(per_cpu(cpus_in_cluster, this_cpu));
 164                         err = -ENOMEM;
 165                 }
 166                 break;
 167         case CPU_UP_CANCELED:
 168         case CPU_UP_CANCELED_FROZEN:
 169         case CPU_DEAD:
 170                 for_each_online_cpu(cpu) {
 171                         if (x2apic_cluster(this_cpu) != x2apic_cluster(cpu))
 172                                 continue;
 173                         __cpu_clear(this_cpu, per_cpu(cpus_in_cluster, cpu));
 174                         __cpu_clear(cpu, per_cpu(cpus_in_cluster, this_cpu));
 175                 }
 176                 free_cpumask_var(per_cpu(cpus_in_cluster, this_cpu));
 177                 free_cpumask_var(per_cpu(ipi_mask, this_cpu));
 178                 break;
 179         }
 180
 181         return notifier_from_errno(err);
 182 }
 183
 184 static struct notifier_block __refdata x2apic_cpu_notifier = {
 185         .notifier_call = update_clusterinfo,
 186 };
 187
 188 static int x2apic_init_cpu_notifier(void)
 189 {
 190         int cpu = smp_processor_id();
 191
 192         zalloc_cpumask_var(&per_cpu(cpus_in_cluster, cpu), GFP_KERNEL);
 193         zalloc_cpumask_var(&per_cpu(ipi_mask, cpu), GFP_KERNEL);
 194
 195         BUG_ON(!per_cpu(cpus_in_cluster, cpu) || !per_cpu(ipi_mask, cpu));
 196
 197         __cpu_set(cpu, per_cpu(cpus_in_cluster, cpu));
 198         register_hotcpu_notifier(&x2apic_cpu_notifier);
 199         return 1;
 200 }
 201
 202 static int x2apic_cluster_probe(void)
 203 {
 204         if (x2apic_mode)
 205                 return x2apic_init_cpu_notifier();
 206         else
 207                 return 0;
 208 }
 209
 210 static const struct cpumask *x2apic_cluster_target_cpus(void)
 211 {
 212         return cpu_all_mask;
 213 }
 214
 215 /*
 216  * Each x2apic cluster is an allocation domain.
 217  */
 218 static void cluster_vector_allocation_domain(int cpu, struct cpumask *retmask,
 219                                              const struct cpumask *mask)
 220 {
 221         /*
 222          * To minimize vector pressure, default case of boot, device bringup
 223          * etc will use a single cpu for the interrupt destination.
 224          *
 225          * On explicit migration requests coming from irqbalance etc,
 226          * interrupts will be routed to the x2apic cluster (cluster-id
 227          * derived from the first cpu in the mask) members specified
 228          * in the mask.
 229          */
 230         if (mask == x2apic_cluster_target_cpus())
 231                 cpumask_copy(retmask, cpumask_of(cpu));
 232         else
 233                 cpumask_and(retmask, mask, per_cpu(cpus_in_cluster, cpu));
 234 }
 235
 236 static struct apic apic_x2apic_cluster = {
 237
 238         .name                           = "cluster x2apic",
 239         .probe                          = x2apic_cluster_probe,
 240         .acpi_madt_oem_check            = x2apic_acpi_madt_oem_check,
 241         .apic_id_valid                  = x2apic_apic_id_valid,
 242         .apic_id_registered             = x2apic_apic_id_registered,
 243
 244         .irq_delivery_mode              = dest_LowestPrio,
 245         .irq_dest_mode                  = 1, /* logical */
 246
 247         .target_cpus                    = x2apic_cluster_target_cpus,
 248         .disable_esr                    = 0,
 249         .dest_logical                   = APIC_DEST_LOGICAL,
 250         .check_apicid_used              = NULL,
 251
 252         .vector_allocation_domain       = cluster_vector_allocation_domain,
 253         .init_apic_ldr                  = init_x2apic_ldr,
 254
 255         .ioapic_phys_id_map             = NULL,
 256         .setup_apic_routing             = NULL,
 257         .cpu_present_to_apicid          = default_cpu_present_to_apicid,
 258         .apicid_to_cpu_present          = NULL,
 259         .check_phys_apicid_present      = default_check_phys_apicid_present,
 260         .phys_pkg_id                    = x2apic_phys_pkg_id,
 261
 262         .get_apic_id                    = x2apic_get_apic_id,
 263         .set_apic_id                    = x2apic_set_apic_id,
 264         .apic_id_mask                   = 0xFFFFFFFFu,
 265
 266         .cpu_mask_to_apicid_and         = x2apic_cpu_mask_to_apicid_and,
 267
 268         .send_IPI_mask                  = x2apic_send_IPI_mask,
 269         .send_IPI_mask_allbutself       = x2apic_send_IPI_mask_allbutself,
 270         .send_IPI_allbutself            = x2apic_send_IPI_allbutself,
 271         .send_IPI_all                   = x2apic_send_IPI_all,
 272         .send_IPI_self                  = x2apic_send_IPI_self,
 273
 274         .wait_for_init_deassert         = false,
 275         .inquire_remote_apic            = NULL,
 276
 277         .read                           = native_apic_msr_read,
 278         .write                          = native_apic_msr_write,
 279         .eoi_write                      = native_apic_msr_eoi_write,
 280         .icr_read                       = native_x2apic_icr_read,
 281         .icr_write                      = native_x2apic_icr_write,
 282         .wait_icr_idle                  = native_x2apic_wait_icr_idle,
 283         .safe_wait_icr_idle             = native_safe_x2apic_wait_icr_idle,
 284 };
 285
 286 apic_driver(apic_x2apic_cluster);