arch/x86/kernel/cpu/amd_64.c

   1 #include <linux/init.h>
   2 #include <linux/mm.h>
   3
   4 #include <asm/numa_64.h>
   5 #include <asm/mmconfig.h>
   6 #include <asm/cacheflush.h>
   7
   8 #include <mach_apic.h>
   9 #include "cpu.h"
  10
  11 extern int __cpuinit get_model_name(struct cpuinfo_x86 *c);
  12 extern void __cpuinit display_cacheinfo(struct cpuinfo_x86 *c);
  13
  14 int force_mwait __cpuinitdata;
  15
  16 #ifdef CONFIG_NUMA
  17 static int __cpuinit nearby_node(int apicid)
  18 {
  19         int i, node;
  20
  21         for (i = apicid - 1; i >= 0; i--) {
  22                 node = apicid_to_node[i];
  23                 if (node != NUMA_NO_NODE && node_online(node))
  24                         return node;
  25         }
  26         for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
  27                 node = apicid_to_node[i];
  28                 if (node != NUMA_NO_NODE && node_online(node))
  29                         return node;
  30         }
  31         return first_node(node_online_map); /* Shouldn't happen */
  32 }
  33 #endif
  34
  35 /*
  36  * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
  37  * Assumes number of cores is a power of two.
  38  */
  39 static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
  40 {
  41 #ifdef CONFIG_SMP
  42         unsigned bits;
  43 #ifdef CONFIG_NUMA
  44         int cpu = smp_processor_id();
  45         int node = 0;
  46         unsigned apicid = hard_smp_processor_id();
  47 #endif
  48         bits = c->x86_coreid_bits;
  49
  50         /* Low order bits define the core id (index of core in socket) */
  51         c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
  52         /* Convert the initial APIC ID into the socket ID */
  53         c->phys_proc_id = c->initial_apicid >> bits;
  54
  55 #ifdef CONFIG_NUMA
  56         node = c->phys_proc_id;
  57         if (apicid_to_node[apicid] != NUMA_NO_NODE)
  58                 node = apicid_to_node[apicid];
  59         if (!node_online(node)) {
  60                 /* Two possibilities here:
  61                    - The CPU is missing memory and no node was created.
  62                    In that case try picking one from a nearby CPU
  63                    - The APIC IDs differ from the HyperTransport node IDs
  64                    which the K8 northbridge parsing fills in.
  65                    Assume they are all increased by a constant offset,
  66                    but in the same order as the HT nodeids.
  67                    If that doesn't result in a usable node fall back to the
  68                    path for the previous case.  */
  69
  70                 int ht_nodeid = c->initial_apicid;
  71
  72                 if (ht_nodeid >= 0 &&
  73                     apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
  74                         node = apicid_to_node[ht_nodeid];
  75                 /* Pick a nearby node */
  76                 if (!node_online(node))
  77                         node = nearby_node(apicid);
  78         }
  79         numa_set_node(cpu, node);
  80
  81         printk(KERN_INFO "CPU %d/%x -> Node %d\n", cpu, apicid, node);
  82 #endif
  83 #endif
  84 }
  85
  86 static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
  87 {
  88 #ifdef CONFIG_SMP
  89         unsigned bits, ecx;
  90
  91         /* Multi core CPU? */
  92         if (c->extended_cpuid_level < 0x80000008)
  93                 return;
  94
  95         ecx = cpuid_ecx(0x80000008);
  96
  97         c->x86_max_cores = (ecx & 0xff) + 1;
  98
  99         /* CPU telling us the core id bits shift? */
 100         bits = (ecx >> 12) & 0xF;
 101
 102         /* Otherwise recompute */
 103         if (bits == 0) {
 104                 while ((1 << bits) < c->x86_max_cores)
 105                         bits++;
 106         }
 107
 108         c->x86_coreid_bits = bits;
 109
 110 #endif
 111 }
 112
 113 /* AMD systems with C1E don't have a working lAPIC timer. Check for that. */
 114 static __cpuinit int amd_apic_timer_broken(struct cpuinfo_x86 *c)
 115 {
 116         u32 lo, hi;
 117
 118         if (c->x86 < 0x0F)
 119                 return 0;
 120
 121         /* Family 0x0f models < rev F do not have this MSR */
 122         if (c->x86 == 0x0f && c->x86_model < 0x40)
 123                 return 0;
 124
 125         rdmsr(MSR_K8_INT_PENDING_MSG, lo, hi);
 126         if (lo & K8_INTP_C1E_ACTIVE_MASK) {
 127                 if (smp_processor_id() != boot_cpu_physical_apicid)
 128                         printk(KERN_INFO "AMD C1E detected late. "
 129                                "Force timer broadcast.\n");
 130                 return 1;
 131         }
 132         return 0;
 133 }
 134
 135 void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
 136 {
 137         early_init_amd_mc(c);
 138
 139         /* c->x86_power is 8000_0007 edx. Bit 8 is constant TSC */
 140         if (c->x86_power & (1<<8))
 141                 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
 142 }
 143
 144 void __cpuinit init_amd(struct cpuinfo_x86 *c)
 145 {
 146         unsigned level;
 147
 148 #ifdef CONFIG_SMP
 149         unsigned long value;
 150
 151         /*
 152          * Disable TLB flush filter by setting HWCR.FFDIS on K8
 153          * bit 6 of msr C001_0015
 154          *
 155          * Errata 63 for SH-B3 steppings
 156          * Errata 122 for all steppings (F+ have it disabled by default)
 157          */
 158         if (c->x86 == 15) {
 159                 rdmsrl(MSR_K8_HWCR, value);
 160                 value |= 1 << 6;
 161                 wrmsrl(MSR_K8_HWCR, value);
 162         }
 163 #endif
 164
 165         /* Bit 31 in normal CPUID used for nonstandard 3DNow ID;
 166            3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway */
 167         clear_cpu_cap(c, 0*32+31);
 168
 169         /* On C+ stepping K8 rep microcode works well for copy/memset */
 170         level = cpuid_eax(1);
 171         if (c->x86 == 15 && ((level >= 0x0f48 && level < 0x0f50) ||
 172                              level >= 0x0f58))
 173                 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
 174         if (c->x86 == 0x10 || c->x86 == 0x11)
 175                 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
 176
 177         /* Enable workaround for FXSAVE leak */
 178         if (c->x86 >= 6)
 179                 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
 180
 181         level = get_model_name(c);
 182         if (!level) {
 183                 switch (c->x86) {
 184                 case 15:
 185                         /* Should distinguish Models here, but this is only
 186                            a fallback anyways. */
 187                         strcpy(c->x86_model_id, "Hammer");
 188                         break;
 189                 }
 190         }
 191         display_cacheinfo(c);
 192
 193         /* Multi core CPU? */
 194         if (c->extended_cpuid_level >= 0x80000008)
 195                 amd_detect_cmp(c);
 196
 197         if (c->extended_cpuid_level >= 0x80000006 &&
 198                 (cpuid_edx(0x80000006) & 0xf000))
 199                 num_cache_leaves = 4;
 200         else
 201                 num_cache_leaves = 3;
 202
 203         if (c->x86 == 0xf || c->x86 == 0x10 || c->x86 == 0x11)
 204                 set_cpu_cap(c, X86_FEATURE_K8);
 205
 206         /* MFENCE stops RDTSC speculation */
 207         set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
 208
 209         if (c->x86 == 0x10)
 210                 fam10h_check_enable_mmcfg();
 211
 212         if (c->x86 == 0x10)
 213                 amd_enable_pci_ext_cfg(c);
 214
 215         if (amd_apic_timer_broken(c))
 216                 disable_apic_timer = 1;
 217
 218         if (c == &boot_cpu_data && c->x86 >= 0xf && c->x86 <= 0x11) {
 219                 unsigned long long tseg;
 220
 221                 /*
 222                  * Split up direct mapping around the TSEG SMM area.
 223                  * Don't do it for gbpages because there seems very little
 224                  * benefit in doing so.
 225                  */
 226                 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg) &&
 227                     (tseg >> PMD_SHIFT) <
 228                         (max_pfn_mapped >> (PMD_SHIFT-PAGE_SHIFT)))
 229                         set_memory_4k((unsigned long)__va(tseg), 1);
 230         }
 231 }