arch/x86/kernel/apic/x2apic_cluster.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/cpuhotplug.h>
   4 #include <linux/cpumask.h>
   5 #include <linux/slab.h>
   6 #include <linux/mm.h>
   7
   8 #include <asm/apic.h>
   9
  10 #include "local.h"
  11
  12 struct cluster_mask {
  13         unsigned int    clusterid;
  14         int             node;
  15         struct cpumask  mask;
  16 };
  17
  18 /*
  19  * __x2apic_send_IPI_mask() possibly needs to read
  20  * x86_cpu_to_logical_apicid for all online cpus in a sequential way.
  21  * Using per cpu variable would cost one cache line per cpu.
  22  */
  23 static u32 *x86_cpu_to_logical_apicid __read_mostly;
  24
  25 static DEFINE_PER_CPU(cpumask_var_t, ipi_mask);
  26 static DEFINE_PER_CPU_READ_MOSTLY(struct cluster_mask *, cluster_masks);
  27 static struct cluster_mask *cluster_hotplug_mask;
  28
  29 static int x2apic_acpi_madt_oem_check(char *oem_id, char *oem_table_id)
  30 {
  31         return x2apic_enabled();
  32 }
  33
  34 static void x2apic_send_IPI(int cpu, int vector)
  35 {
  36         u32 dest = x86_cpu_to_logical_apicid[cpu];
  37
  38         /* x2apic MSRs are special and need a special fence: */
  39         weak_wrmsr_fence();
  40         __x2apic_send_IPI_dest(dest, vector, APIC_DEST_LOGICAL);
  41 }
  42
  43 static void
  44 __x2apic_send_IPI_mask(const struct cpumask *mask, int vector, int apic_dest)
  45 {
  46         unsigned int cpu, clustercpu;
  47         struct cpumask *tmpmsk;
  48         unsigned long flags;
  49         u32 dest;
  50
  51         /* x2apic MSRs are special and need a special fence: */
  52         weak_wrmsr_fence();
  53         local_irq_save(flags);
  54
  55         tmpmsk = this_cpu_cpumask_var_ptr(ipi_mask);
  56         cpumask_copy(tmpmsk, mask);
  57         /* If IPI should not be sent to self, clear current CPU */
  58         if (apic_dest != APIC_DEST_ALLINC)
  59                 __cpumask_clear_cpu(smp_processor_id(), tmpmsk);
  60
  61         /* Collapse cpus in a cluster so a single IPI per cluster is sent */
  62         for_each_cpu(cpu, tmpmsk) {
  63                 struct cluster_mask *cmsk = per_cpu(cluster_masks, cpu);
  64
  65                 dest = 0;
  66                 for_each_cpu_and(clustercpu, tmpmsk, &cmsk->mask)
  67                         dest |= x86_cpu_to_logical_apicid[clustercpu];
  68
  69                 if (!dest)
  70                         continue;
  71
  72                 __x2apic_send_IPI_dest(dest, vector, APIC_DEST_LOGICAL);
  73                 /* Remove cluster CPUs from tmpmask */
  74                 cpumask_andnot(tmpmsk, tmpmsk, &cmsk->mask);
  75         }
  76
  77         local_irq_restore(flags);
  78 }
  79
  80 static void x2apic_send_IPI_mask(const struct cpumask *mask, int vector)
  81 {
  82         __x2apic_send_IPI_mask(mask, vector, APIC_DEST_ALLINC);
  83 }
  84
  85 static void
  86 x2apic_send_IPI_mask_allbutself(const struct cpumask *mask, int vector)
  87 {
  88         __x2apic_send_IPI_mask(mask, vector, APIC_DEST_ALLBUT);
  89 }
  90
  91 static void x2apic_send_IPI_allbutself(int vector)
  92 {
  93         __x2apic_send_IPI_shorthand(vector, APIC_DEST_ALLBUT);
  94 }
  95
  96 static void x2apic_send_IPI_all(int vector)
  97 {
  98         __x2apic_send_IPI_shorthand(vector, APIC_DEST_ALLINC);
  99 }
 100
 101 static u32 x2apic_calc_apicid(unsigned int cpu)
 102 {
 103         return x86_cpu_to_logical_apicid[cpu];
 104 }
 105
 106 static void init_x2apic_ldr(void)
 107 {
 108         struct cluster_mask *cmsk = this_cpu_read(cluster_masks);
 109         u32 cluster, apicid = apic_read(APIC_LDR);
 110         unsigned int cpu;
 111
 112         x86_cpu_to_logical_apicid[smp_processor_id()] = apicid;
 113
 114         if (cmsk)
 115                 goto update;
 116
 117         cluster = apicid >> 16;
 118         for_each_online_cpu(cpu) {
 119                 cmsk = per_cpu(cluster_masks, cpu);
 120                 /* Matching cluster found. Link and update it. */
 121                 if (cmsk && cmsk->clusterid == cluster)
 122                         goto update;
 123         }
 124         cmsk = cluster_hotplug_mask;
 125         cmsk->clusterid = cluster;
 126         cluster_hotplug_mask = NULL;
 127 update:
 128         this_cpu_write(cluster_masks, cmsk);
 129         cpumask_set_cpu(smp_processor_id(), &cmsk->mask);
 130 }
 131
 132 static int alloc_clustermask(unsigned int cpu, int node)
 133 {
 134         if (per_cpu(cluster_masks, cpu))
 135                 return 0;
 136         /*
 137          * If a hotplug spare mask exists, check whether it's on the right
 138          * node. If not, free it and allocate a new one.
 139          */
 140         if (cluster_hotplug_mask) {
 141                 if (cluster_hotplug_mask->node == node)
 142                         return 0;
 143                 kfree(cluster_hotplug_mask);
 144         }
 145
 146         cluster_hotplug_mask = kzalloc_node(sizeof(*cluster_hotplug_mask),
 147                                             GFP_KERNEL, node);
 148         if (!cluster_hotplug_mask)
 149                 return -ENOMEM;
 150         cluster_hotplug_mask->node = node;
 151         return 0;
 152 }
 153
 154 static int x2apic_prepare_cpu(unsigned int cpu)
 155 {
 156         if (alloc_clustermask(cpu, cpu_to_node(cpu)) < 0)
 157                 return -ENOMEM;
 158         if (!zalloc_cpumask_var(&per_cpu(ipi_mask, cpu), GFP_KERNEL))
 159                 return -ENOMEM;
 160         return 0;
 161 }
 162
 163 static int x2apic_dead_cpu(unsigned int dead_cpu)
 164 {
 165         struct cluster_mask *cmsk = per_cpu(cluster_masks, dead_cpu);
 166
 167         if (cmsk)
 168                 cpumask_clear_cpu(dead_cpu, &cmsk->mask);
 169         free_cpumask_var(per_cpu(ipi_mask, dead_cpu));
 170         return 0;
 171 }
 172
 173 static int x2apic_cluster_probe(void)
 174 {
 175         u32 slots;
 176
 177         if (!x2apic_mode)
 178                 return 0;
 179
 180         slots = max_t(u32, L1_CACHE_BYTES/sizeof(u32), nr_cpu_ids);
 181         x86_cpu_to_logical_apicid = kcalloc(slots, sizeof(u32), GFP_KERNEL);
 182         if (!x86_cpu_to_logical_apicid)
 183                 return 0;
 184
 185         if (cpuhp_setup_state(CPUHP_X2APIC_PREPARE, "x86/x2apic:prepare",
 186                               x2apic_prepare_cpu, x2apic_dead_cpu) < 0) {
 187                 pr_err("Failed to register X2APIC_PREPARE\n");
 188                 kfree(x86_cpu_to_logical_apicid);
 189                 x86_cpu_to_logical_apicid = NULL;
 190                 return 0;
 191         }
 192         init_x2apic_ldr();
 193         return 1;
 194 }
 195
 196 static struct apic apic_x2apic_cluster __ro_after_init = {
 197
 198         .name                           = "cluster x2apic",
 199         .probe                          = x2apic_cluster_probe,
 200         .acpi_madt_oem_check            = x2apic_acpi_madt_oem_check,
 201         .apic_id_valid                  = x2apic_apic_id_valid,
 202         .apic_id_registered             = x2apic_apic_id_registered,
 203
 204         .delivery_mode                  = APIC_DELIVERY_MODE_FIXED,
 205         .dest_mode_logical              = true,
 206
 207         .disable_esr                    = 0,
 208
 209         .check_apicid_used              = NULL,
 210         .init_apic_ldr                  = init_x2apic_ldr,
 211         .ioapic_phys_id_map             = NULL,
 212         .setup_apic_routing             = NULL,
 213         .cpu_present_to_apicid          = default_cpu_present_to_apicid,
 214         .apicid_to_cpu_present          = NULL,
 215         .check_phys_apicid_present      = default_check_phys_apicid_present,
 216         .phys_pkg_id                    = x2apic_phys_pkg_id,
 217
 218         .get_apic_id                    = x2apic_get_apic_id,
 219         .set_apic_id                    = x2apic_set_apic_id,
 220
 221         .calc_dest_apicid               = x2apic_calc_apicid,
 222
 223         .send_IPI                       = x2apic_send_IPI,
 224         .send_IPI_mask                  = x2apic_send_IPI_mask,
 225         .send_IPI_mask_allbutself       = x2apic_send_IPI_mask_allbutself,
 226         .send_IPI_allbutself            = x2apic_send_IPI_allbutself,
 227         .send_IPI_all                   = x2apic_send_IPI_all,
 228         .send_IPI_self                  = x2apic_send_IPI_self,
 229
 230         .inquire_remote_apic            = NULL,
 231
 232         .read                           = native_apic_msr_read,
 233         .write                          = native_apic_msr_write,
 234         .eoi_write                      = native_apic_msr_eoi_write,
 235         .icr_read                       = native_x2apic_icr_read,
 236         .icr_write                      = native_x2apic_icr_write,
 237         .wait_icr_idle                  = native_x2apic_wait_icr_idle,
 238         .safe_wait_icr_idle             = native_safe_x2apic_wait_icr_idle,
 239 };
 240
 241 apic_driver(apic_x2apic_cluster);