amd.c 20 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841
  1. #include <linux/export.h>
  2. #include <linux/init.h>
  3. #include <linux/bitops.h>
  4. #include <linux/elf.h>
  5. #include <linux/mm.h>
  6. #include <linux/io.h>
  7. #include <linux/sched.h>
  8. #include <asm/processor.h>
  9. #include <asm/apic.h>
  10. #include <asm/cpu.h>
  11. #include <asm/pci-direct.h>
  12. #ifdef CONFIG_X86_64
  13. # include <asm/numa_64.h>
  14. # include <asm/mmconfig.h>
  15. # include <asm/cacheflush.h>
  16. #endif
  17. #include "cpu.h"
  18. #ifdef CONFIG_X86_32
  19. /*
  20. * B step AMD K6 before B 9730xxxx have hardware bugs that can cause
  21. * misexecution of code under Linux. Owners of such processors should
  22. * contact AMD for precise details and a CPU swap.
  23. *
  24. * See http://www.multimania.com/poulot/k6bug.html
  25. * and section 2.6.2 of "AMD-K6 Processor Revision Guide - Model 6"
  26. * (Publication # 21266 Issue Date: August 1998)
  27. *
  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..
  31. */
  32. extern void vide(void);
  33. __asm__(".align 4\nvide: ret");
  34. static void __cpuinit init_amd_k5(struct cpuinfo_x86 *c)
  35. {
  36. /*
  37. * General Systems BIOSen alias the cpu frequency registers
  38. * of the Elan at 0x000df000. Unfortuantly, one of the Linux
  39. * drivers subsequently pokes it, and changes the CPU speed.
  40. * Workaround : Remove the unneeded alias.
  41. */
  42. #define CBAR (0xfffc) /* Configuration Base Address (32-bit) */
  43. #define CBAR_ENB (0x80000000)
  44. #define CBAR_KEY (0X000000CB)
  45. if (c->x86_model == 9 || c->x86_model == 10) {
  46. if (inl(CBAR) & CBAR_ENB)
  47. outl(0 | CBAR_KEY, CBAR);
  48. }
  49. }
  50. static void __cpuinit init_amd_k6(struct cpuinfo_x86 *c)
  51. {
  52. u32 l, h;
  53. int mbytes = num_physpages >> (20-PAGE_SHIFT);
  54. if (c->x86_model < 6) {
  55. /* Based on AMD doc 20734R - June 2000 */
  56. if (c->x86_model == 0) {
  57. clear_cpu_cap(c, X86_FEATURE_APIC);
  58. set_cpu_cap(c, X86_FEATURE_PGE);
  59. }
  60. return;
  61. }
  62. if (c->x86_model == 6 && c->x86_mask == 1) {
  63. const int K6_BUG_LOOP = 1000000;
  64. int n;
  65. void (*f_vide)(void);
  66. unsigned long d, d2;
  67. printk(KERN_INFO "AMD K6 stepping B detected - ");
  68. /*
  69. * It looks like AMD fixed the 2.6.2 bug and improved indirect
  70. * calls at the same time.
  71. */
  72. n = K6_BUG_LOOP;
  73. f_vide = vide;
  74. rdtscl(d);
  75. while (n--)
  76. f_vide();
  77. rdtscl(d2);
  78. d = d2-d;
  79. if (d > 20*K6_BUG_LOOP)
  80. printk(KERN_CONT
  81. "system stability may be impaired when more than 32 MB are used.\n");
  82. else
  83. printk(KERN_CONT "probably OK (after B9730xxxx).\n");
  84. }
  85. /* K6 with old style WHCR */
  86. if (c->x86_model < 8 ||
  87. (c->x86_model == 8 && c->x86_mask < 8)) {
  88. /* We can only write allocate on the low 508Mb */
  89. if (mbytes > 508)
  90. mbytes = 508;
  91. rdmsr(MSR_K6_WHCR, l, h);
  92. if ((l&0x0000FFFF) == 0) {
  93. unsigned long flags;
  94. l = (1<<0)|((mbytes/4)<<1);
  95. local_irq_save(flags);
  96. wbinvd();
  97. wrmsr(MSR_K6_WHCR, l, h);
  98. local_irq_restore(flags);
  99. printk(KERN_INFO "Enabling old style K6 write allocation for %d Mb\n",
  100. mbytes);
  101. }
  102. return;
  103. }
  104. if ((c->x86_model == 8 && c->x86_mask > 7) ||
  105. c->x86_model == 9 || c->x86_model == 13) {
  106. /* The more serious chips .. */
  107. if (mbytes > 4092)
  108. mbytes = 4092;
  109. rdmsr(MSR_K6_WHCR, l, h);
  110. if ((l&0xFFFF0000) == 0) {
  111. unsigned long flags;
  112. l = ((mbytes>>2)<<22)|(1<<16);
  113. local_irq_save(flags);
  114. wbinvd();
  115. wrmsr(MSR_K6_WHCR, l, h);
  116. local_irq_restore(flags);
  117. printk(KERN_INFO "Enabling new style K6 write allocation for %d Mb\n",
  118. mbytes);
  119. }
  120. return;
  121. }
  122. if (c->x86_model == 10) {
  123. /* AMD Geode LX is model 10 */
  124. /* placeholder for any needed mods */
  125. return;
  126. }
  127. }
  128. static void __cpuinit amd_k7_smp_check(struct cpuinfo_x86 *c)
  129. {
  130. /* calling is from identify_secondary_cpu() ? */
  131. if (!c->cpu_index)
  132. return;
  133. /*
  134. * Certain Athlons might work (for various values of 'work') in SMP
  135. * but they are not certified as MP capable.
  136. */
  137. /* Athlon 660/661 is valid. */
  138. if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
  139. (c->x86_mask == 1)))
  140. goto valid_k7;
  141. /* Duron 670 is valid */
  142. if ((c->x86_model == 7) && (c->x86_mask == 0))
  143. goto valid_k7;
  144. /*
  145. * Athlon 662, Duron 671, and Athlon >model 7 have capability
  146. * bit. It's worth noting that the A5 stepping (662) of some
  147. * Athlon XP's have the MP bit set.
  148. * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
  149. * more.
  150. */
  151. if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
  152. ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
  153. (c->x86_model > 7))
  154. if (cpu_has_mp)
  155. goto valid_k7;
  156. /* If we get here, not a certified SMP capable AMD system. */
  157. /*
  158. * Don't taint if we are running SMP kernel on a single non-MP
  159. * approved Athlon
  160. */
  161. WARN_ONCE(1, "WARNING: This combination of AMD"
  162. " processors is not suitable for SMP.\n");
  163. if (!test_taint(TAINT_UNSAFE_SMP))
  164. add_taint(TAINT_UNSAFE_SMP);
  165. valid_k7:
  166. ;
  167. }
  168. static void __cpuinit init_amd_k7(struct cpuinfo_x86 *c)
  169. {
  170. u32 l, h;
  171. /*
  172. * Bit 15 of Athlon specific MSR 15, needs to be 0
  173. * to enable SSE on Palomino/Morgan/Barton CPU's.
  174. * If the BIOS didn't enable it already, enable it here.
  175. */
  176. if (c->x86_model >= 6 && c->x86_model <= 10) {
  177. if (!cpu_has(c, X86_FEATURE_XMM)) {
  178. printk(KERN_INFO "Enabling disabled K7/SSE Support.\n");
  179. rdmsr(MSR_K7_HWCR, l, h);
  180. l &= ~0x00008000;
  181. wrmsr(MSR_K7_HWCR, l, h);
  182. set_cpu_cap(c, X86_FEATURE_XMM);
  183. }
  184. }
  185. /*
  186. * It's been determined by AMD that Athlons since model 8 stepping 1
  187. * are more robust with CLK_CTL set to 200xxxxx instead of 600xxxxx
  188. * As per AMD technical note 27212 0.2
  189. */
  190. if ((c->x86_model == 8 && c->x86_mask >= 1) || (c->x86_model > 8)) {
  191. rdmsr(MSR_K7_CLK_CTL, l, h);
  192. if ((l & 0xfff00000) != 0x20000000) {
  193. printk(KERN_INFO
  194. "CPU: CLK_CTL MSR was %x. Reprogramming to %x\n",
  195. l, ((l & 0x000fffff)|0x20000000));
  196. wrmsr(MSR_K7_CLK_CTL, (l & 0x000fffff)|0x20000000, h);
  197. }
  198. }
  199. set_cpu_cap(c, X86_FEATURE_K7);
  200. amd_k7_smp_check(c);
  201. }
  202. #endif
  203. #ifdef CONFIG_NUMA
  204. /*
  205. * To workaround broken NUMA config. Read the comment in
  206. * srat_detect_node().
  207. */
  208. static int __cpuinit nearby_node(int apicid)
  209. {
  210. int i, node;
  211. for (i = apicid - 1; i >= 0; i--) {
  212. node = __apicid_to_node[i];
  213. if (node != NUMA_NO_NODE && node_online(node))
  214. return node;
  215. }
  216. for (i = apicid + 1; i < MAX_LOCAL_APIC; i++) {
  217. node = __apicid_to_node[i];
  218. if (node != NUMA_NO_NODE && node_online(node))
  219. return node;
  220. }
  221. return first_node(node_online_map); /* Shouldn't happen */
  222. }
  223. #endif
  224. /*
  225. * Fixup core topology information for
  226. * (1) AMD multi-node processors
  227. * Assumption: Number of cores in each internal node is the same.
  228. * (2) AMD processors supporting compute units
  229. */
  230. #ifdef CONFIG_X86_HT
  231. static void __cpuinit amd_get_topology(struct cpuinfo_x86 *c)
  232. {
  233. u32 nodes, cores_per_cu = 1;
  234. u8 node_id;
  235. int cpu = smp_processor_id();
  236. /* get information required for multi-node processors */
  237. if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
  238. u32 eax, ebx, ecx, edx;
  239. cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
  240. nodes = ((ecx >> 8) & 7) + 1;
  241. node_id = ecx & 7;
  242. /* get compute unit information */
  243. smp_num_siblings = ((ebx >> 8) & 3) + 1;
  244. c->compute_unit_id = ebx & 0xff;
  245. cores_per_cu += ((ebx >> 8) & 3);
  246. } else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
  247. u64 value;
  248. rdmsrl(MSR_FAM10H_NODE_ID, value);
  249. nodes = ((value >> 3) & 7) + 1;
  250. node_id = value & 7;
  251. } else
  252. return;
  253. /* fixup multi-node processor information */
  254. if (nodes > 1) {
  255. u32 cores_per_node;
  256. u32 cus_per_node;
  257. set_cpu_cap(c, X86_FEATURE_AMD_DCM);
  258. cores_per_node = c->x86_max_cores / nodes;
  259. cus_per_node = cores_per_node / cores_per_cu;
  260. /* store NodeID, use llc_shared_map to store sibling info */
  261. per_cpu(cpu_llc_id, cpu) = node_id;
  262. /* core id has to be in the [0 .. cores_per_node - 1] range */
  263. c->cpu_core_id %= cores_per_node;
  264. c->compute_unit_id %= cus_per_node;
  265. }
  266. }
  267. #endif
  268. /*
  269. * On a AMD dual core setup the lower bits of the APIC id distingush the cores.
  270. * Assumes number of cores is a power of two.
  271. */
  272. static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
  273. {
  274. #ifdef CONFIG_X86_HT
  275. unsigned bits;
  276. int cpu = smp_processor_id();
  277. bits = c->x86_coreid_bits;
  278. /* Low order bits define the core id (index of core in socket) */
  279. c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
  280. /* Convert the initial APIC ID into the socket ID */
  281. c->phys_proc_id = c->initial_apicid >> bits;
  282. /* use socket ID also for last level cache */
  283. per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
  284. amd_get_topology(c);
  285. #endif
  286. }
  287. int amd_get_nb_id(int cpu)
  288. {
  289. int id = 0;
  290. #ifdef CONFIG_SMP
  291. id = per_cpu(cpu_llc_id, cpu);
  292. #endif
  293. return id;
  294. }
  295. EXPORT_SYMBOL_GPL(amd_get_nb_id);
  296. static void __cpuinit srat_detect_node(struct cpuinfo_x86 *c)
  297. {
  298. #ifdef CONFIG_NUMA
  299. int cpu = smp_processor_id();
  300. int node;
  301. unsigned apicid = c->apicid;
  302. node = numa_cpu_node(cpu);
  303. if (node == NUMA_NO_NODE)
  304. node = per_cpu(cpu_llc_id, cpu);
  305. /*
  306. * On multi-fabric platform (e.g. Numascale NumaChip) a
  307. * platform-specific handler needs to be called to fixup some
  308. * IDs of the CPU.
  309. */
  310. if (x86_cpuinit.fixup_cpu_id)
  311. x86_cpuinit.fixup_cpu_id(c, node);
  312. if (!node_online(node)) {
  313. /*
  314. * Two possibilities here:
  315. *
  316. * - The CPU is missing memory and no node was created. In
  317. * that case try picking one from a nearby CPU.
  318. *
  319. * - The APIC IDs differ from the HyperTransport node IDs
  320. * which the K8 northbridge parsing fills in. Assume
  321. * they are all increased by a constant offset, but in
  322. * the same order as the HT nodeids. If that doesn't
  323. * result in a usable node fall back to the path for the
  324. * previous case.
  325. *
  326. * This workaround operates directly on the mapping between
  327. * APIC ID and NUMA node, assuming certain relationship
  328. * between APIC ID, HT node ID and NUMA topology. As going
  329. * through CPU mapping may alter the outcome, directly
  330. * access __apicid_to_node[].
  331. */
  332. int ht_nodeid = c->initial_apicid;
  333. if (ht_nodeid >= 0 &&
  334. __apicid_to_node[ht_nodeid] != NUMA_NO_NODE)
  335. node = __apicid_to_node[ht_nodeid];
  336. /* Pick a nearby node */
  337. if (!node_online(node))
  338. node = nearby_node(apicid);
  339. }
  340. numa_set_node(cpu, node);
  341. #endif
  342. }
  343. static void __cpuinit early_init_amd_mc(struct cpuinfo_x86 *c)
  344. {
  345. #ifdef CONFIG_X86_HT
  346. unsigned bits, ecx;
  347. /* Multi core CPU? */
  348. if (c->extended_cpuid_level < 0x80000008)
  349. return;
  350. ecx = cpuid_ecx(0x80000008);
  351. c->x86_max_cores = (ecx & 0xff) + 1;
  352. /* CPU telling us the core id bits shift? */
  353. bits = (ecx >> 12) & 0xF;
  354. /* Otherwise recompute */
  355. if (bits == 0) {
  356. while ((1 << bits) < c->x86_max_cores)
  357. bits++;
  358. }
  359. c->x86_coreid_bits = bits;
  360. #endif
  361. /* F16h erratum 793, CVE-2013-6885 */
  362. if (c->x86 == 0x16 && c->x86_model <= 0xf) {
  363. u64 val;
  364. if (!rdmsrl_amd_safe(MSR_AMD64_LS_CFG, &val) &&
  365. !(val & BIT(15)))
  366. wrmsrl_amd_safe(MSR_AMD64_LS_CFG, val | BIT(15));
  367. }
  368. }
  369. static void __cpuinit bsp_init_amd(struct cpuinfo_x86 *c)
  370. {
  371. if (cpu_has(c, X86_FEATURE_CONSTANT_TSC)) {
  372. if (c->x86 > 0x10 ||
  373. (c->x86 == 0x10 && c->x86_model >= 0x2)) {
  374. u64 val;
  375. rdmsrl(MSR_K7_HWCR, val);
  376. if (!(val & BIT(24)))
  377. printk(KERN_WARNING FW_BUG "TSC doesn't count "
  378. "with P0 frequency!\n");
  379. }
  380. }
  381. if (c->x86 == 0x15) {
  382. unsigned long upperbit;
  383. u32 cpuid, assoc;
  384. cpuid = cpuid_edx(0x80000005);
  385. assoc = cpuid >> 16 & 0xff;
  386. upperbit = ((cpuid >> 24) << 10) / assoc;
  387. va_align.mask = (upperbit - 1) & PAGE_MASK;
  388. va_align.flags = ALIGN_VA_32 | ALIGN_VA_64;
  389. }
  390. }
  391. static void __cpuinit early_init_amd(struct cpuinfo_x86 *c)
  392. {
  393. early_init_amd_mc(c);
  394. /*
  395. * c->x86_power is 8000_0007 edx. Bit 8 is TSC runs at constant rate
  396. * with P/T states and does not stop in deep C-states
  397. */
  398. if (c->x86_power & (1 << 8)) {
  399. set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
  400. set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
  401. if (!check_tsc_unstable())
  402. sched_clock_stable = 1;
  403. }
  404. #ifdef CONFIG_X86_64
  405. set_cpu_cap(c, X86_FEATURE_SYSCALL32);
  406. #else
  407. /* Set MTRR capability flag if appropriate */
  408. if (c->x86 == 5)
  409. if (c->x86_model == 13 || c->x86_model == 9 ||
  410. (c->x86_model == 8 && c->x86_mask >= 8))
  411. set_cpu_cap(c, X86_FEATURE_K6_MTRR);
  412. #endif
  413. #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_PCI)
  414. /* check CPU config space for extended APIC ID */
  415. if (cpu_has_apic && c->x86 >= 0xf) {
  416. unsigned int val;
  417. val = read_pci_config(0, 24, 0, 0x68);
  418. if ((val & ((1 << 17) | (1 << 18))) == ((1 << 17) | (1 << 18)))
  419. set_cpu_cap(c, X86_FEATURE_EXTD_APICID);
  420. }
  421. #endif
  422. }
  423. static void __cpuinit init_amd(struct cpuinfo_x86 *c)
  424. {
  425. u32 dummy;
  426. #ifdef CONFIG_SMP
  427. unsigned long long value;
  428. /*
  429. * Disable TLB flush filter by setting HWCR.FFDIS on K8
  430. * bit 6 of msr C001_0015
  431. *
  432. * Errata 63 for SH-B3 steppings
  433. * Errata 122 for all steppings (F+ have it disabled by default)
  434. */
  435. if (c->x86 == 0xf) {
  436. rdmsrl(MSR_K7_HWCR, value);
  437. value |= 1 << 6;
  438. wrmsrl(MSR_K7_HWCR, value);
  439. }
  440. #endif
  441. early_init_amd(c);
  442. /*
  443. * Bit 31 in normal CPUID used for nonstandard 3DNow ID;
  444. * 3DNow is IDd by bit 31 in extended CPUID (1*32+31) anyway
  445. */
  446. clear_cpu_cap(c, 0*32+31);
  447. #ifdef CONFIG_X86_64
  448. /* On C+ stepping K8 rep microcode works well for copy/memset */
  449. if (c->x86 == 0xf) {
  450. u32 level;
  451. level = cpuid_eax(1);
  452. if ((level >= 0x0f48 && level < 0x0f50) || level >= 0x0f58)
  453. set_cpu_cap(c, X86_FEATURE_REP_GOOD);
  454. /*
  455. * Some BIOSes incorrectly force this feature, but only K8
  456. * revision D (model = 0x14) and later actually support it.
  457. * (AMD Erratum #110, docId: 25759).
  458. */
  459. if (c->x86_model < 0x14 && cpu_has(c, X86_FEATURE_LAHF_LM)) {
  460. u64 val;
  461. clear_cpu_cap(c, X86_FEATURE_LAHF_LM);
  462. if (!rdmsrl_amd_safe(0xc001100d, &val)) {
  463. val &= ~(1ULL << 32);
  464. wrmsrl_amd_safe(0xc001100d, val);
  465. }
  466. }
  467. }
  468. if (c->x86 >= 0x10)
  469. set_cpu_cap(c, X86_FEATURE_REP_GOOD);
  470. /* get apicid instead of initial apic id from cpuid */
  471. c->apicid = hard_smp_processor_id();
  472. #else
  473. /*
  474. * FIXME: We should handle the K5 here. Set up the write
  475. * range and also turn on MSR 83 bits 4 and 31 (write alloc,
  476. * no bus pipeline)
  477. */
  478. switch (c->x86) {
  479. case 4:
  480. init_amd_k5(c);
  481. break;
  482. case 5:
  483. init_amd_k6(c);
  484. break;
  485. case 6: /* An Athlon/Duron */
  486. init_amd_k7(c);
  487. break;
  488. }
  489. /* K6s reports MCEs but don't actually have all the MSRs */
  490. if (c->x86 < 6)
  491. clear_cpu_cap(c, X86_FEATURE_MCE);
  492. #endif
  493. /* Enable workaround for FXSAVE leak */
  494. if (c->x86 >= 6)
  495. set_cpu_cap(c, X86_FEATURE_FXSAVE_LEAK);
  496. if (!c->x86_model_id[0]) {
  497. switch (c->x86) {
  498. case 0xf:
  499. /* Should distinguish Models here, but this is only
  500. a fallback anyways. */
  501. strcpy(c->x86_model_id, "Hammer");
  502. break;
  503. }
  504. }
  505. /* re-enable TopologyExtensions if switched off by BIOS */
  506. if ((c->x86 == 0x15) &&
  507. (c->x86_model >= 0x10) && (c->x86_model <= 0x1f) &&
  508. !cpu_has(c, X86_FEATURE_TOPOEXT)) {
  509. u64 val;
  510. if (!rdmsrl_amd_safe(0xc0011005, &val)) {
  511. val |= 1ULL << 54;
  512. wrmsrl_amd_safe(0xc0011005, val);
  513. rdmsrl(0xc0011005, val);
  514. if (val & (1ULL << 54)) {
  515. set_cpu_cap(c, X86_FEATURE_TOPOEXT);
  516. printk(KERN_INFO FW_INFO "CPU: Re-enabling "
  517. "disabled Topology Extensions Support\n");
  518. }
  519. }
  520. }
  521. /*
  522. * The way access filter has a performance penalty on some workloads.
  523. * Disable it on the affected CPUs.
  524. */
  525. if ((c->x86 == 0x15) &&
  526. (c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
  527. u64 val;
  528. if (!rdmsrl_safe(0xc0011021, &val) && !(val & 0x1E)) {
  529. val |= 0x1E;
  530. checking_wrmsrl(0xc0011021, val);
  531. }
  532. }
  533. /*
  534. * The way access filter has a performance penalty on some workloads.
  535. * Disable it on the affected CPUs.
  536. */
  537. if ((c->x86 == 0x15) &&
  538. (c->x86_model >= 0x02) && (c->x86_model < 0x20)) {
  539. u64 val;
  540. if (!rdmsrl_safe(0xc0011021, &val) && !(val & 0x1E)) {
  541. val |= 0x1E;
  542. checking_wrmsrl(0xc0011021, val);
  543. }
  544. }
  545. cpu_detect_cache_sizes(c);
  546. /* Multi core CPU? */
  547. if (c->extended_cpuid_level >= 0x80000008) {
  548. amd_detect_cmp(c);
  549. srat_detect_node(c);
  550. }
  551. #ifdef CONFIG_X86_32
  552. detect_ht(c);
  553. #endif
  554. if (c->extended_cpuid_level >= 0x80000006) {
  555. if (cpuid_edx(0x80000006) & 0xf000)
  556. num_cache_leaves = 4;
  557. else
  558. num_cache_leaves = 3;
  559. }
  560. if (c->x86 >= 0xf)
  561. set_cpu_cap(c, X86_FEATURE_K8);
  562. if (cpu_has_xmm2) {
  563. /* MFENCE stops RDTSC speculation */
  564. set_cpu_cap(c, X86_FEATURE_MFENCE_RDTSC);
  565. }
  566. #ifdef CONFIG_X86_64
  567. if (c->x86 == 0x10) {
  568. /* do this for boot cpu */
  569. if (c == &boot_cpu_data)
  570. check_enable_amd_mmconf_dmi();
  571. fam10h_check_enable_mmcfg();
  572. }
  573. if (c == &boot_cpu_data && c->x86 >= 0xf) {
  574. unsigned long long tseg;
  575. /*
  576. * Split up direct mapping around the TSEG SMM area.
  577. * Don't do it for gbpages because there seems very little
  578. * benefit in doing so.
  579. */
  580. if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg)) {
  581. printk(KERN_DEBUG "tseg: %010llx\n", tseg);
  582. if ((tseg>>PMD_SHIFT) <
  583. (max_low_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) ||
  584. ((tseg>>PMD_SHIFT) <
  585. (max_pfn_mapped>>(PMD_SHIFT-PAGE_SHIFT)) &&
  586. (tseg>>PMD_SHIFT) >= (1ULL<<(32 - PMD_SHIFT))))
  587. set_memory_4k((unsigned long)__va(tseg), 1);
  588. }
  589. }
  590. #endif
  591. /*
  592. * Family 0x12 and above processors have APIC timer
  593. * running in deep C states.
  594. */
  595. if (c->x86 > 0x11)
  596. set_cpu_cap(c, X86_FEATURE_ARAT);
  597. /*
  598. * Disable GART TLB Walk Errors on Fam10h. We do this here
  599. * because this is always needed when GART is enabled, even in a
  600. * kernel which has no MCE support built in.
  601. */
  602. if (c->x86 == 0x10) {
  603. /*
  604. * BIOS should disable GartTlbWlk Errors themself. If
  605. * it doesn't do it here as suggested by the BKDG.
  606. *
  607. * Fixes: https://bugzilla.kernel.org/show_bug.cgi?id=33012
  608. */
  609. u64 mask;
  610. int err;
  611. err = rdmsrl_safe(MSR_AMD64_MCx_MASK(4), &mask);
  612. if (err == 0) {
  613. mask |= (1 << 10);
  614. checking_wrmsrl(MSR_AMD64_MCx_MASK(4), mask);
  615. }
  616. }
  617. rdmsr_safe(MSR_AMD64_PATCH_LEVEL, &c->microcode, &dummy);
  618. }
  619. #ifdef CONFIG_X86_32
  620. static unsigned int __cpuinit amd_size_cache(struct cpuinfo_x86 *c,
  621. unsigned int size)
  622. {
  623. /* AMD errata T13 (order #21922) */
  624. if ((c->x86 == 6)) {
  625. /* Duron Rev A0 */
  626. if (c->x86_model == 3 && c->x86_mask == 0)
  627. size = 64;
  628. /* Tbird rev A1/A2 */
  629. if (c->x86_model == 4 &&
  630. (c->x86_mask == 0 || c->x86_mask == 1))
  631. size = 256;
  632. }
  633. return size;
  634. }
  635. #endif
  636. static const struct cpu_dev __cpuinitconst amd_cpu_dev = {
  637. .c_vendor = "AMD",
  638. .c_ident = { "AuthenticAMD" },
  639. #ifdef CONFIG_X86_32
  640. .c_models = {
  641. { .vendor = X86_VENDOR_AMD, .family = 4, .model_names =
  642. {
  643. [3] = "486 DX/2",
  644. [7] = "486 DX/2-WB",
  645. [8] = "486 DX/4",
  646. [9] = "486 DX/4-WB",
  647. [14] = "Am5x86-WT",
  648. [15] = "Am5x86-WB"
  649. }
  650. },
  651. },
  652. .c_size_cache = amd_size_cache,
  653. #endif
  654. .c_early_init = early_init_amd,
  655. .c_bsp_init = bsp_init_amd,
  656. .c_init = init_amd,
  657. .c_x86_vendor = X86_VENDOR_AMD,
  658. };
  659. cpu_dev_register(amd_cpu_dev);
  660. /*
  661. * AMD errata checking
  662. *
  663. * Errata are defined as arrays of ints using the AMD_LEGACY_ERRATUM() or
  664. * AMD_OSVW_ERRATUM() macros. The latter is intended for newer errata that
  665. * have an OSVW id assigned, which it takes as first argument. Both take a
  666. * variable number of family-specific model-stepping ranges created by
  667. * AMD_MODEL_RANGE(). Each erratum also has to be declared as extern const
  668. * int[] in arch/x86/include/asm/processor.h.
  669. *
  670. * Example:
  671. *
  672. * const int amd_erratum_319[] =
  673. * AMD_LEGACY_ERRATUM(AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0x4, 0x2),
  674. * AMD_MODEL_RANGE(0x10, 0x8, 0x0, 0x8, 0x0),
  675. * AMD_MODEL_RANGE(0x10, 0x9, 0x0, 0x9, 0x0));
  676. */
  677. const int amd_erratum_400[] =
  678. AMD_OSVW_ERRATUM(1, AMD_MODEL_RANGE(0xf, 0x41, 0x2, 0xff, 0xf),
  679. AMD_MODEL_RANGE(0x10, 0x2, 0x1, 0xff, 0xf));
  680. EXPORT_SYMBOL_GPL(amd_erratum_400);
  681. const int amd_erratum_383[] =
  682. AMD_OSVW_ERRATUM(3, AMD_MODEL_RANGE(0x10, 0, 0, 0xff, 0xf));
  683. EXPORT_SYMBOL_GPL(amd_erratum_383);
  684. bool cpu_has_amd_erratum(const int *erratum)
  685. {
  686. struct cpuinfo_x86 *cpu = __this_cpu_ptr(&cpu_info);
  687. int osvw_id = *erratum++;
  688. u32 range;
  689. u32 ms;
  690. /*
  691. * If called early enough that current_cpu_data hasn't been initialized
  692. * yet, fall back to boot_cpu_data.
  693. */
  694. if (cpu->x86 == 0)
  695. cpu = &boot_cpu_data;
  696. if (cpu->x86_vendor != X86_VENDOR_AMD)
  697. return false;
  698. if (osvw_id >= 0 && osvw_id < 65536 &&
  699. cpu_has(cpu, X86_FEATURE_OSVW)) {
  700. u64 osvw_len;
  701. rdmsrl(MSR_AMD64_OSVW_ID_LENGTH, osvw_len);
  702. if (osvw_id < osvw_len) {
  703. u64 osvw_bits;
  704. rdmsrl(MSR_AMD64_OSVW_STATUS + (osvw_id >> 6),
  705. osvw_bits);
  706. return osvw_bits & (1ULL << (osvw_id & 0x3f));
  707. }
  708. }
  709. /* OSVW unavailable or ID unknown, match family-model-stepping range */
  710. ms = (cpu->x86_model << 4) | cpu->x86_mask;
  711. while ((range = *erratum++))
  712. if ((cpu->x86 == AMD_MODEL_RANGE_FAMILY(range)) &&
  713. (ms >= AMD_MODEL_RANGE_START(range)) &&
  714. (ms <= AMD_MODEL_RANGE_END(range)))
  715. return true;
  716. return false;
  717. }
  718. EXPORT_SYMBOL_GPL(cpu_has_amd_erratum);