1 #include <linux/init.h>
2 #include <linux/bitops.h>
6 #include <asm/processor.h>
9 #include <asm/pci-direct.h>
12 # include <asm/numa_64.h>
13 # include <asm/mmconfig.h>
14 # include <asm/cacheflush.h>
21 * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
22 * misexecution of code under Linux. Owners of such processors should
23 * contact AMD for precise details and a CPU swap.
25 * See http://www.multimania.com/poulot/k6bug.html
26 * http://www.amd.com/K6/k6docs/revgd.html
28 * The following test is erm.. interesting. AMD neglected to up
29 * the chip setting when fixing the bug but they also tweaked some
30 * performance at the same time..
33 extern void vide(void);
34 __asm__(".align 4\nvide: ret");
36 static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
39 * General Systems BIOSen alias the cpu frequency registers
40 * of the Elan at 0x000df000. Unfortuantly, one of the Linux
41 * drivers subsequently pokes it, and changes the CPU speed.
42 * Workaround : Remove the unneeded alias.
44 #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
45 #define CBAR_ENB (0x80000000)
46 #define CBAR_KEY (0X000000CB)
47 if (c->x86_model == 9 || c->x86_model == 10) {
48 if (inl (CBAR) & CBAR_ENB)
49 outl (0 | CBAR_KEY, CBAR);
54 static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
57 int mbytes = num_physpages >> (20-PAGE_SHIFT);
59 if (c->x86_model < 6) {
60 /* Based on AMD doc 20734R - June 2000 */
61 if (c->x86_model == 0) {
62 clear_cpu_cap(c, X86_FEATURE_APIC);
63 set_cpu_cap(c, X86_FEATURE_PGE);
68 if (c->x86_model == 6 && c->x86_mask == 1) {
69 const int K6_BUG_LOOP = 1000000;
74 printk(KERN_INFO "AMD K6 stepping B detected - ");
77 * It looks like AMD fixed the 2.6.2 bug and improved indirect
78 * calls at the same time.
89 if (d > 20*K6_BUG_LOOP)
90 printk("system stability may be impaired when more than 32 MB are used.\n");
92 printk("probably OK (after B9730xxxx).\n");
93 printk(KERN_INFO "Please see http://membres.lycos.fr/poulot/k6bug.html\n");
96 /* K6 with old style WHCR */
97 if (c->x86_model < 8 ||
98 (c->x86_model == 8 && c->x86_mask < 8)) {
99 /* We can only write allocate on the low 508Mb */
103 rdmsr(MSR_K6_WHCR, l, h);
104 if ((l&0x0000FFFF) == 0) {
106 l = (1<<0)|((mbytes/4)<<1);
107 local_irq_save(flags);
109 wrmsr(MSR_K6_WHCR, l, h);
110 local_irq_restore(flags);
111 printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
117 if ((c->x86_model == 8 && c->x86_mask > 7) ||
118 c->x86_model == 9 || c->x86_model == 13) {
119 /* The more serious chips .. */
124 rdmsr(MSR_K6_WHCR, l, h);
125 if ((l&0xFFFF0000) == 0) {
127 l = ((mbytes>>2)<<22)|(1<<16);
128 local_irq_save(flags);
130 wrmsr(MSR_K6_WHCR, l, h);
131 local_irq_restore(flags);
132 printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
139 if (c->x86_model == 10) {
140 /* AMD Geode LX is model 10 */
141 /* placeholder for any needed mods */
146 static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
149 /* calling is from identify_secondary_cpu() ? */
150 if (c->cpu_index == boot_cpu_id)
154 * Certain Athlons might work (for various values of 'work') in SMP
155 * but they are not certified as MP capable.
157 /* Athlon 660/661 is valid. */
158 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
162 /* Duron 670 is valid */
163 if ((c->x86_model == 7) && (c->x86_mask == 0))
167 * Athlon 662, Duron 671, and Athlon >model 7 have capability
168 * bit. It's worth noting that the A5 stepping (662) of some
169 * Athlon XP's have the MP bit set.
170 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
173 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
174 ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
179 /* If we get here, not a certified SMP capable AMD system. */
182 * Don't taint if we are running SMP kernel on a single non-MP
185 WARN_ONCE(1, "WARNING: This combination of AMD"
186 "processors is not suitable for SMP.\n");
187 if (!test_taint(TAINT_UNSAFE_SMP))
188 add_taint(TAINT_UNSAFE_SMP);
195 static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
200 * Bit 15 of Athlon specific MSR 15, needs to be 0
201 * to enable SSE on Palomino/Morgan/Barton CPU's.
202 * If the BIOS didn't enable it already, enable it here.
204 if (c->x86_model >= 6 && c->x86_model <= 10) {
205 if (!cpu_has(c, X86_FEATURE_XMM)) {
206 printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
207 rdmsr(MSR_K7_HWCR, l, h);
209 wrmsr(MSR_K7_HWCR, l, h);
210 set_cpu_cap(c, X86_FEATURE_XMM);
215 * It's been determined by AMD that Athlons since model 8 stepping 1
216 * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
217 * As per AMD technical note 27212 0.2
219 if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
220 rdmsr(MSR_K7_CLK_CTL, l, h);
221 if ((l & 0xfff00000) != 0x20000000) {
222 printk ("CPU: CLK_CTL MSR was %x. Reprogramming to %x\n", l,
223 ((l & 0x000fffff)|0x20000000));
224 wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
228 set_cpu_cap(c, X86_FEATURE_K7);
234 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
235 static int __cpuinit nearby_node(int apicid)
239 for (i = apicid - 1; i >= 0; i--) {
240 node = apicid_to_node[i];
241 if (node != NUMA_NO_NODE && node_online(node))
244 for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
245 node = apicid_to_node[i];
246 if (node != NUMA_NO_NODE && node_online(node))
249 return first_node(node_online_map); /* Shouldn't happen */
254 * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
255 * Assumes number of cores is a power of two.
257 static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
262 bits = c->x86_coreid_bits;
264 /* Low order bits define the core id (index of core in socket) */
265 c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
266 /* Convert the initial APIC ID into the socket ID */
267 c->phys_proc_id = c->initial_apicid >> bits;
271 static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
273 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_64)
274 int cpu = smp_processor_id();
276 unsigned apicid = cpu_has_apic ? hard_smp_processor_id() : c->apicid;
278 node = c->phys_proc_id;
279 if (apicid_to_node[apicid] != NUMA_NO_NODE)
280 node = apicid_to_node[apicid];
281 if (!node_online(node)) {
282 /* Two possibilities here:
283 - The CPU is missing memory and no node was created.
284 In that case try picking one from a nearby CPU
285 - The APIC IDs differ from the HyperTransport node IDs
286 which the K8 northbridge parsing fills in.
287 Assume they are all increased by a constant offset,
288 but in the same order as the HT nodeids.
289 If that doesn't result in a usable node fall back to the
290 path for the previous case. */
292 int ht_nodeid = c->initial_apicid;
294 if (ht_nodeid >= 0 &&
295 apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
296 node = apicid_to_node[ht_nodeid];
297 /* Pick a nearby node */
298 if (!node_online(node))
299 node = nearby_node(apicid);
301 numa_set_node(cpu, node);
303 printk(KERN_INFO "CPU %d/0x%x -> Node %d\n", cpu, apicid, node);
307 static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
312 /* Multi core CPU? */
313 if (c->extended_cpuid_level < 0x80000008)
316 ecx = cpuid_ecx(0x80000008);
318 c->x86_max_cores = (ecx & 0xff) + 1;
320 /* CPU telling us the core id bits shift? */
321 bits = (ecx >> 12) & 0xF;
323 /* Otherwise recompute */
325 while ((1 << bits) < c->x86_max_cores)
329 c->x86_coreid_bits = bits;
333 static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
335 early_init_amd_mc(c);
338 * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
339 * with P/T states and does not stop in deep C-states
341 if (c->x86_power & (1 << 8)) {
342 set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
343 set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
347 set_cpu_cap(c, X86_FEATURE_SYSCALL32);
349 /* Set MTRR capability flag if appropriate */
351 if (c->x86_model == 13 || c->x86_model == 9 ||
352 (c->x86_model == 8 && c->x86_mask >= 8))
353 set_cpu_cap(c, X86_FEATURE_K6_MTRR);
355 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
356 /* check CPU config space for extended APIC ID */
359 val = read_pci_config(0, 24, 0, 0x68);
360 if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
361 set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
366 static void __cpuinit init_amd(struct cpuinfo_x86 *c)
369 unsigned long long value;
372 * Disable TLB flush filter by setting HWCR.FFDIS on K8
373 * bit 6 of msr C001_0015
375 * Errata 63 for SH-B3 steppings
376 * Errata 122 for all steppings (F+ have it disabled by default)
379 rdmsrl(MSR_K7_HWCR, value);
381 wrmsrl(MSR_K7_HWCR, value);
388 * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
389 * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
391 clear_cpu_cap(c, 0*32+31);
394 /* On C+ stepping K8 rep microcode works well for copy/memset */
398 level = cpuid_eax(1);
399 if((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
400 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
402 if (c->x86 == 0x10 || c->x86 == 0x11)
403 set_cpu_cap(c, X86_FEATURE_REP_GOOD);
407 * FIXME: We should handle the K5 here. Set up the write
408 * range and also turn on MSR 83 bits 4 and 31 (write alloc,
419 case 6: /* An Athlon/Duron */
424 /* K6s reports MCEs but don't actually have all the MSRs */
426 clear_cpu_cap(c, X86_FEATURE_MCE);
429 /* Enable workaround for FXSAVE leak */
431 set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
433 if (!c->x86_model_id[0]) {
436 /* Should distinguish Models here, but this is only
437 a fallback anyways. */
438 strcpy(c->x86_model_id, "Hammer");
443 display_cacheinfo(c);
445 /* Multi core CPU? */
446 if (c->extended_cpuid_level >= 0x80000008) {
455 if (c->extended_cpuid_level >= 0x80000006) {
456 if ((c->x86 >= 0x0f) && (cpuid_edx(0x80000006) & 0xf000))
457 num_cache_leaves = 4;
459 num_cache_leaves = 3;
462 if (c->x86 >= 0xf && c->x86 <= 0x11)
463 set_cpu_cap(c, X86_FEATURE_K8);
466 /* MFENCE stops RDTSC speculation */
467 set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
471 if (c->x86 == 0x10) {
472 /* do this for boot cpu */
473 if (c == &boot_cpu_data)
474 check_enable_amd_mmconf_dmi();
476 fam10h_check_enable_mmcfg();
479 if (c == &boot_cpu_data && c->x86 >= 0xf && c->x86 <= 0x11) {
480 unsigned long long tseg;
483 * Split up direct mapping around the TSEG SMM area.
484 * Don't do it for gbpages because there seems very little
485 * benefit in doing so.
487 if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
488 printk(KERN_DEBUG "tseg: %010llx\n", tseg);
489 if ((tseg>>PMD_SHIFT) <
490 (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
492 (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
493 (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
494 set_memory_4k((unsigned long)__va(tseg), 1);
501 static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c, unsigned int size)
503 /* AMD errata T13 (order #21922) */
505 if (c->x86_model == 3 && c->x86_mask == 0) /* Duron Rev A0 */
507 if (c->x86_model == 4 &&
508 (c->x86_mask == 0 || c->x86_mask == 1)) /* Tbird rev A1/A2 */
515 static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
517 .c_ident = { "AuthenticAMD" },
520 { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
531 .c_size_cache = amd_size_cache,
533 .c_early_init = early_init_amd,
535 .c_x86_vendor = X86_VENDOR_AMD,
538 cpu_dev_register(amd_cpu_dev);