/drivers/edac/amd64_edac.c

https://bitbucket.org/cresqo/cm7-p500-kernel · C · 3087 lines · 1869 code · 520 blank · 698 comment · 305 complexity · 2296b64b5b7caa0e645455e3e5c17301 MD5 · raw file

Large files are truncated click here to view the full file

  1. #include "amd64_edac.h"
  2. #include <asm/k8.h>
  3. static struct edac_pci_ctl_info *amd64_ctl_pci;
  4. static int report_gart_errors;
  5. module_param(report_gart_errors, int, 0644);
  6. /*
  7. * Set by command line parameter. If BIOS has enabled the ECC, this override is
  8. * cleared to prevent re-enabling the hardware by this driver.
  9. */
  10. static int ecc_enable_override;
  11. module_param(ecc_enable_override, int, 0644);
  12. static struct msr __percpu *msrs;
  13. /* Lookup table for all possible MC control instances */
  14. struct amd64_pvt;
  15. static struct mem_ctl_info *mci_lookup[EDAC_MAX_NUMNODES];
  16. static struct amd64_pvt *pvt_lookup[EDAC_MAX_NUMNODES];
  17. /*
  18. * Address to DRAM bank mapping: see F2x80 for K8 and F2x[1,0]80 for Fam10 and
  19. * later.
  20. */
  21. static int ddr2_dbam_revCG[] = {
  22. [0] = 32,
  23. [1] = 64,
  24. [2] = 128,
  25. [3] = 256,
  26. [4] = 512,
  27. [5] = 1024,
  28. [6] = 2048,
  29. };
  30. static int ddr2_dbam_revD[] = {
  31. [0] = 32,
  32. [1] = 64,
  33. [2 ... 3] = 128,
  34. [4] = 256,
  35. [5] = 512,
  36. [6] = 256,
  37. [7] = 512,
  38. [8 ... 9] = 1024,
  39. [10] = 2048,
  40. };
  41. static int ddr2_dbam[] = { [0] = 128,
  42. [1] = 256,
  43. [2 ... 4] = 512,
  44. [5 ... 6] = 1024,
  45. [7 ... 8] = 2048,
  46. [9 ... 10] = 4096,
  47. [11] = 8192,
  48. };
  49. static int ddr3_dbam[] = { [0] = -1,
  50. [1] = 256,
  51. [2] = 512,
  52. [3 ... 4] = -1,
  53. [5 ... 6] = 1024,
  54. [7 ... 8] = 2048,
  55. [9 ... 10] = 4096,
  56. [11] = 8192,
  57. };
  58. /*
  59. * Valid scrub rates for the K8 hardware memory scrubber. We map the scrubbing
  60. * bandwidth to a valid bit pattern. The 'set' operation finds the 'matching-
  61. * or higher value'.
  62. *
  63. *FIXME: Produce a better mapping/linearisation.
  64. */
  65. struct scrubrate scrubrates[] = {
  66. { 0x01, 1600000000UL},
  67. { 0x02, 800000000UL},
  68. { 0x03, 400000000UL},
  69. { 0x04, 200000000UL},
  70. { 0x05, 100000000UL},
  71. { 0x06, 50000000UL},
  72. { 0x07, 25000000UL},
  73. { 0x08, 12284069UL},
  74. { 0x09, 6274509UL},
  75. { 0x0A, 3121951UL},
  76. { 0x0B, 1560975UL},
  77. { 0x0C, 781440UL},
  78. { 0x0D, 390720UL},
  79. { 0x0E, 195300UL},
  80. { 0x0F, 97650UL},
  81. { 0x10, 48854UL},
  82. { 0x11, 24427UL},
  83. { 0x12, 12213UL},
  84. { 0x13, 6101UL},
  85. { 0x14, 3051UL},
  86. { 0x15, 1523UL},
  87. { 0x16, 761UL},
  88. { 0x00, 0UL}, /* scrubbing off */
  89. };
  90. /*
  91. * Memory scrubber control interface. For K8, memory scrubbing is handled by
  92. * hardware and can involve L2 cache, dcache as well as the main memory. With
  93. * F10, this is extended to L3 cache scrubbing on CPU models sporting that
  94. * functionality.
  95. *
  96. * This causes the "units" for the scrubbing speed to vary from 64 byte blocks
  97. * (dram) over to cache lines. This is nasty, so we will use bandwidth in
  98. * bytes/sec for the setting.
  99. *
  100. * Currently, we only do dram scrubbing. If the scrubbing is done in software on
  101. * other archs, we might not have access to the caches directly.
  102. */
  103. /*
  104. * scan the scrub rate mapping table for a close or matching bandwidth value to
  105. * issue. If requested is too big, then use last maximum value found.
  106. */
  107. static int amd64_search_set_scrub_rate(struct pci_dev *ctl, u32 new_bw,
  108. u32 min_scrubrate)
  109. {
  110. u32 scrubval;
  111. int i;
  112. /*
  113. * map the configured rate (new_bw) to a value specific to the AMD64
  114. * memory controller and apply to register. Search for the first
  115. * bandwidth entry that is greater or equal than the setting requested
  116. * and program that. If at last entry, turn off DRAM scrubbing.
  117. */
  118. for (i = 0; i < ARRAY_SIZE(scrubrates); i++) {
  119. /*
  120. * skip scrub rates which aren't recommended
  121. * (see F10 BKDG, F3x58)
  122. */
  123. if (scrubrates[i].scrubval < min_scrubrate)
  124. continue;
  125. if (scrubrates[i].bandwidth <= new_bw)
  126. break;
  127. /*
  128. * if no suitable bandwidth found, turn off DRAM scrubbing
  129. * entirely by falling back to the last element in the
  130. * scrubrates array.
  131. */
  132. }
  133. scrubval = scrubrates[i].scrubval;
  134. if (scrubval)
  135. edac_printk(KERN_DEBUG, EDAC_MC,
  136. "Setting scrub rate bandwidth: %u\n",
  137. scrubrates[i].bandwidth);
  138. else
  139. edac_printk(KERN_DEBUG, EDAC_MC, "Turning scrubbing off.\n");
  140. pci_write_bits32(ctl, K8_SCRCTRL, scrubval, 0x001F);
  141. return 0;
  142. }
  143. static int amd64_set_scrub_rate(struct mem_ctl_info *mci, u32 *bandwidth)
  144. {
  145. struct amd64_pvt *pvt = mci->pvt_info;
  146. u32 min_scrubrate = 0x0;
  147. switch (boot_cpu_data.x86) {
  148. case 0xf:
  149. min_scrubrate = K8_MIN_SCRUB_RATE_BITS;
  150. break;
  151. case 0x10:
  152. min_scrubrate = F10_MIN_SCRUB_RATE_BITS;
  153. break;
  154. case 0x11:
  155. min_scrubrate = F11_MIN_SCRUB_RATE_BITS;
  156. break;
  157. default:
  158. amd64_printk(KERN_ERR, "Unsupported family!\n");
  159. return -EINVAL;
  160. }
  161. return amd64_search_set_scrub_rate(pvt->misc_f3_ctl, *bandwidth,
  162. min_scrubrate);
  163. }
  164. static int amd64_get_scrub_rate(struct mem_ctl_info *mci, u32 *bw)
  165. {
  166. struct amd64_pvt *pvt = mci->pvt_info;
  167. u32 scrubval = 0;
  168. int status = -1, i;
  169. amd64_read_pci_cfg(pvt->misc_f3_ctl, K8_SCRCTRL, &scrubval);
  170. scrubval = scrubval & 0x001F;
  171. edac_printk(KERN_DEBUG, EDAC_MC,
  172. "pci-read, sdram scrub control value: %d \n", scrubval);
  173. for (i = 0; i < ARRAY_SIZE(scrubrates); i++) {
  174. if (scrubrates[i].scrubval == scrubval) {
  175. *bw = scrubrates[i].bandwidth;
  176. status = 0;
  177. break;
  178. }
  179. }
  180. return status;
  181. }
  182. /* Map from a CSROW entry to the mask entry that operates on it */
  183. static inline u32 amd64_map_to_dcs_mask(struct amd64_pvt *pvt, int csrow)
  184. {
  185. if (boot_cpu_data.x86 == 0xf && pvt->ext_model < K8_REV_F)
  186. return csrow;
  187. else
  188. return csrow >> 1;
  189. }
  190. /* return the 'base' address the i'th CS entry of the 'dct' DRAM controller */
  191. static u32 amd64_get_dct_base(struct amd64_pvt *pvt, int dct, int csrow)
  192. {
  193. if (dct == 0)
  194. return pvt->dcsb0[csrow];
  195. else
  196. return pvt->dcsb1[csrow];
  197. }
  198. /*
  199. * Return the 'mask' address the i'th CS entry. This function is needed because
  200. * there number of DCSM registers on Rev E and prior vs Rev F and later is
  201. * different.
  202. */
  203. static u32 amd64_get_dct_mask(struct amd64_pvt *pvt, int dct, int csrow)
  204. {
  205. if (dct == 0)
  206. return pvt->dcsm0[amd64_map_to_dcs_mask(pvt, csrow)];
  207. else
  208. return pvt->dcsm1[amd64_map_to_dcs_mask(pvt, csrow)];
  209. }
  210. /*
  211. * In *base and *limit, pass back the full 40-bit base and limit physical
  212. * addresses for the node given by node_id. This information is obtained from
  213. * DRAM Base (section 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers. The
  214. * base and limit addresses are of type SysAddr, as defined at the start of
  215. * section 3.4.4 (p. 70). They are the lowest and highest physical addresses
  216. * in the address range they represent.
  217. */
  218. static void amd64_get_base_and_limit(struct amd64_pvt *pvt, int node_id,
  219. u64 *base, u64 *limit)
  220. {
  221. *base = pvt->dram_base[node_id];
  222. *limit = pvt->dram_limit[node_id];
  223. }
  224. /*
  225. * Return 1 if the SysAddr given by sys_addr matches the base/limit associated
  226. * with node_id
  227. */
  228. static int amd64_base_limit_match(struct amd64_pvt *pvt,
  229. u64 sys_addr, int node_id)
  230. {
  231. u64 base, limit, addr;
  232. amd64_get_base_and_limit(pvt, node_id, &base, &limit);
  233. /* The K8 treats this as a 40-bit value. However, bits 63-40 will be
  234. * all ones if the most significant implemented address bit is 1.
  235. * Here we discard bits 63-40. See section 3.4.2 of AMD publication
  236. * 24592: AMD x86-64 Architecture Programmer's Manual Volume 1
  237. * Application Programming.
  238. */
  239. addr = sys_addr & 0x000000ffffffffffull;
  240. return (addr >= base) && (addr <= limit);
  241. }
  242. /*
  243. * Attempt to map a SysAddr to a node. On success, return a pointer to the
  244. * mem_ctl_info structure for the node that the SysAddr maps to.
  245. *
  246. * On failure, return NULL.
  247. */
  248. static struct mem_ctl_info *find_mc_by_sys_addr(struct mem_ctl_info *mci,
  249. u64 sys_addr)
  250. {
  251. struct amd64_pvt *pvt;
  252. int node_id;
  253. u32 intlv_en, bits;
  254. /*
  255. * Here we use the DRAM Base (section 3.4.4.1) and DRAM Limit (section
  256. * 3.4.4.2) registers to map the SysAddr to a node ID.
  257. */
  258. pvt = mci->pvt_info;
  259. /*
  260. * The value of this field should be the same for all DRAM Base
  261. * registers. Therefore we arbitrarily choose to read it from the
  262. * register for node 0.
  263. */
  264. intlv_en = pvt->dram_IntlvEn[0];
  265. if (intlv_en == 0) {
  266. for (node_id = 0; node_id < DRAM_REG_COUNT; node_id++) {
  267. if (amd64_base_limit_match(pvt, sys_addr, node_id))
  268. goto found;
  269. }
  270. goto err_no_match;
  271. }
  272. if (unlikely((intlv_en != 0x01) &&
  273. (intlv_en != 0x03) &&
  274. (intlv_en != 0x07))) {
  275. amd64_printk(KERN_WARNING, "junk value of 0x%x extracted from "
  276. "IntlvEn field of DRAM Base Register for node 0: "
  277. "this probably indicates a BIOS bug.\n", intlv_en);
  278. return NULL;
  279. }
  280. bits = (((u32) sys_addr) >> 12) & intlv_en;
  281. for (node_id = 0; ; ) {
  282. if ((pvt->dram_IntlvSel[node_id] & intlv_en) == bits)
  283. break; /* intlv_sel field matches */
  284. if (++node_id >= DRAM_REG_COUNT)
  285. goto err_no_match;
  286. }
  287. /* sanity test for sys_addr */
  288. if (unlikely(!amd64_base_limit_match(pvt, sys_addr, node_id))) {
  289. amd64_printk(KERN_WARNING,
  290. "%s(): sys_addr 0x%llx falls outside base/limit "
  291. "address range for node %d with node interleaving "
  292. "enabled.\n",
  293. __func__, sys_addr, node_id);
  294. return NULL;
  295. }
  296. found:
  297. return edac_mc_find(node_id);
  298. err_no_match:
  299. debugf2("sys_addr 0x%lx doesn't match any node\n",
  300. (unsigned long)sys_addr);
  301. return NULL;
  302. }
  303. /*
  304. * Extract the DRAM CS base address from selected csrow register.
  305. */
  306. static u64 base_from_dct_base(struct amd64_pvt *pvt, int csrow)
  307. {
  308. return ((u64) (amd64_get_dct_base(pvt, 0, csrow) & pvt->dcsb_base)) <<
  309. pvt->dcs_shift;
  310. }
  311. /*
  312. * Extract the mask from the dcsb0[csrow] entry in a CPU revision-specific way.
  313. */
  314. static u64 mask_from_dct_mask(struct amd64_pvt *pvt, int csrow)
  315. {
  316. u64 dcsm_bits, other_bits;
  317. u64 mask;
  318. /* Extract bits from DRAM CS Mask. */
  319. dcsm_bits = amd64_get_dct_mask(pvt, 0, csrow) & pvt->dcsm_mask;
  320. other_bits = pvt->dcsm_mask;
  321. other_bits = ~(other_bits << pvt->dcs_shift);
  322. /*
  323. * The extracted bits from DCSM belong in the spaces represented by
  324. * the cleared bits in other_bits.
  325. */
  326. mask = (dcsm_bits << pvt->dcs_shift) | other_bits;
  327. return mask;
  328. }
  329. /*
  330. * @input_addr is an InputAddr associated with the node given by mci. Return the
  331. * csrow that input_addr maps to, or -1 on failure (no csrow claims input_addr).
  332. */
  333. static int input_addr_to_csrow(struct mem_ctl_info *mci, u64 input_addr)
  334. {
  335. struct amd64_pvt *pvt;
  336. int csrow;
  337. u64 base, mask;
  338. pvt = mci->pvt_info;
  339. /*
  340. * Here we use the DRAM CS Base and DRAM CS Mask registers. For each CS
  341. * base/mask register pair, test the condition shown near the start of
  342. * section 3.5.4 (p. 84, BKDG #26094, K8, revA-E).
  343. */
  344. for (csrow = 0; csrow < pvt->cs_count; csrow++) {
  345. /* This DRAM chip select is disabled on this node */
  346. if ((pvt->dcsb0[csrow] & K8_DCSB_CS_ENABLE) == 0)
  347. continue;
  348. base = base_from_dct_base(pvt, csrow);
  349. mask = ~mask_from_dct_mask(pvt, csrow);
  350. if ((input_addr & mask) == (base & mask)) {
  351. debugf2("InputAddr 0x%lx matches csrow %d (node %d)\n",
  352. (unsigned long)input_addr, csrow,
  353. pvt->mc_node_id);
  354. return csrow;
  355. }
  356. }
  357. debugf2("no matching csrow for InputAddr 0x%lx (MC node %d)\n",
  358. (unsigned long)input_addr, pvt->mc_node_id);
  359. return -1;
  360. }
  361. /*
  362. * Return the base value defined by the DRAM Base register for the node
  363. * represented by mci. This function returns the full 40-bit value despite the
  364. * fact that the register only stores bits 39-24 of the value. See section
  365. * 3.4.4.1 (BKDG #26094, K8, revA-E)
  366. */
  367. static inline u64 get_dram_base(struct mem_ctl_info *mci)
  368. {
  369. struct amd64_pvt *pvt = mci->pvt_info;
  370. return pvt->dram_base[pvt->mc_node_id];
  371. }
  372. /*
  373. * Obtain info from the DRAM Hole Address Register (section 3.4.8, pub #26094)
  374. * for the node represented by mci. Info is passed back in *hole_base,
  375. * *hole_offset, and *hole_size. Function returns 0 if info is valid or 1 if
  376. * info is invalid. Info may be invalid for either of the following reasons:
  377. *
  378. * - The revision of the node is not E or greater. In this case, the DRAM Hole
  379. * Address Register does not exist.
  380. *
  381. * - The DramHoleValid bit is cleared in the DRAM Hole Address Register,
  382. * indicating that its contents are not valid.
  383. *
  384. * The values passed back in *hole_base, *hole_offset, and *hole_size are
  385. * complete 32-bit values despite the fact that the bitfields in the DHAR
  386. * only represent bits 31-24 of the base and offset values.
  387. */
  388. int amd64_get_dram_hole_info(struct mem_ctl_info *mci, u64 *hole_base,
  389. u64 *hole_offset, u64 *hole_size)
  390. {
  391. struct amd64_pvt *pvt = mci->pvt_info;
  392. u64 base;
  393. /* only revE and later have the DRAM Hole Address Register */
  394. if (boot_cpu_data.x86 == 0xf && pvt->ext_model < K8_REV_E) {
  395. debugf1(" revision %d for node %d does not support DHAR\n",
  396. pvt->ext_model, pvt->mc_node_id);
  397. return 1;
  398. }
  399. /* only valid for Fam10h */
  400. if (boot_cpu_data.x86 == 0x10 &&
  401. (pvt->dhar & F10_DRAM_MEM_HOIST_VALID) == 0) {
  402. debugf1(" Dram Memory Hoisting is DISABLED on this system\n");
  403. return 1;
  404. }
  405. if ((pvt->dhar & DHAR_VALID) == 0) {
  406. debugf1(" Dram Memory Hoisting is DISABLED on this node %d\n",
  407. pvt->mc_node_id);
  408. return 1;
  409. }
  410. /* This node has Memory Hoisting */
  411. /* +------------------+--------------------+--------------------+-----
  412. * | memory | DRAM hole | relocated |
  413. * | [0, (x - 1)] | [x, 0xffffffff] | addresses from |
  414. * | | | DRAM hole |
  415. * | | | [0x100000000, |
  416. * | | | (0x100000000+ |
  417. * | | | (0xffffffff-x))] |
  418. * +------------------+--------------------+--------------------+-----
  419. *
  420. * Above is a diagram of physical memory showing the DRAM hole and the
  421. * relocated addresses from the DRAM hole. As shown, the DRAM hole
  422. * starts at address x (the base address) and extends through address
  423. * 0xffffffff. The DRAM Hole Address Register (DHAR) relocates the
  424. * addresses in the hole so that they start at 0x100000000.
  425. */
  426. base = dhar_base(pvt->dhar);
  427. *hole_base = base;
  428. *hole_size = (0x1ull << 32) - base;
  429. if (boot_cpu_data.x86 > 0xf)
  430. *hole_offset = f10_dhar_offset(pvt->dhar);
  431. else
  432. *hole_offset = k8_dhar_offset(pvt->dhar);
  433. debugf1(" DHAR info for node %d base 0x%lx offset 0x%lx size 0x%lx\n",
  434. pvt->mc_node_id, (unsigned long)*hole_base,
  435. (unsigned long)*hole_offset, (unsigned long)*hole_size);
  436. return 0;
  437. }
  438. EXPORT_SYMBOL_GPL(amd64_get_dram_hole_info);
  439. /*
  440. * Return the DramAddr that the SysAddr given by @sys_addr maps to. It is
  441. * assumed that sys_addr maps to the node given by mci.
  442. *
  443. * The first part of section 3.4.4 (p. 70) shows how the DRAM Base (section
  444. * 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers are used to translate a
  445. * SysAddr to a DramAddr. If the DRAM Hole Address Register (DHAR) is enabled,
  446. * then it is also involved in translating a SysAddr to a DramAddr. Sections
  447. * 3.4.8 and 3.5.8.2 describe the DHAR and how it is used for memory hoisting.
  448. * These parts of the documentation are unclear. I interpret them as follows:
  449. *
  450. * When node n receives a SysAddr, it processes the SysAddr as follows:
  451. *
  452. * 1. It extracts the DRAMBase and DRAMLimit values from the DRAM Base and DRAM
  453. * Limit registers for node n. If the SysAddr is not within the range
  454. * specified by the base and limit values, then node n ignores the Sysaddr
  455. * (since it does not map to node n). Otherwise continue to step 2 below.
  456. *
  457. * 2. If the DramHoleValid bit of the DHAR for node n is clear, the DHAR is
  458. * disabled so skip to step 3 below. Otherwise see if the SysAddr is within
  459. * the range of relocated addresses (starting at 0x100000000) from the DRAM
  460. * hole. If not, skip to step 3 below. Else get the value of the
  461. * DramHoleOffset field from the DHAR. To obtain the DramAddr, subtract the
  462. * offset defined by this value from the SysAddr.
  463. *
  464. * 3. Obtain the base address for node n from the DRAMBase field of the DRAM
  465. * Base register for node n. To obtain the DramAddr, subtract the base
  466. * address from the SysAddr, as shown near the start of section 3.4.4 (p.70).
  467. */
  468. static u64 sys_addr_to_dram_addr(struct mem_ctl_info *mci, u64 sys_addr)
  469. {
  470. u64 dram_base, hole_base, hole_offset, hole_size, dram_addr;
  471. int ret = 0;
  472. dram_base = get_dram_base(mci);
  473. ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
  474. &hole_size);
  475. if (!ret) {
  476. if ((sys_addr >= (1ull << 32)) &&
  477. (sys_addr < ((1ull << 32) + hole_size))) {
  478. /* use DHAR to translate SysAddr to DramAddr */
  479. dram_addr = sys_addr - hole_offset;
  480. debugf2("using DHAR to translate SysAddr 0x%lx to "
  481. "DramAddr 0x%lx\n",
  482. (unsigned long)sys_addr,
  483. (unsigned long)dram_addr);
  484. return dram_addr;
  485. }
  486. }
  487. /*
  488. * Translate the SysAddr to a DramAddr as shown near the start of
  489. * section 3.4.4 (p. 70). Although sys_addr is a 64-bit value, the k8
  490. * only deals with 40-bit values. Therefore we discard bits 63-40 of
  491. * sys_addr below. If bit 39 of sys_addr is 1 then the bits we
  492. * discard are all 1s. Otherwise the bits we discard are all 0s. See
  493. * section 3.4.2 of AMD publication 24592: AMD x86-64 Architecture
  494. * Programmer's Manual Volume 1 Application Programming.
  495. */
  496. dram_addr = (sys_addr & 0xffffffffffull) - dram_base;
  497. debugf2("using DRAM Base register to translate SysAddr 0x%lx to "
  498. "DramAddr 0x%lx\n", (unsigned long)sys_addr,
  499. (unsigned long)dram_addr);
  500. return dram_addr;
  501. }
  502. /*
  503. * @intlv_en is the value of the IntlvEn field from a DRAM Base register
  504. * (section 3.4.4.1). Return the number of bits from a SysAddr that are used
  505. * for node interleaving.
  506. */
  507. static int num_node_interleave_bits(unsigned intlv_en)
  508. {
  509. static const int intlv_shift_table[] = { 0, 1, 0, 2, 0, 0, 0, 3 };
  510. int n;
  511. BUG_ON(intlv_en > 7);
  512. n = intlv_shift_table[intlv_en];
  513. return n;
  514. }
  515. /* Translate the DramAddr given by @dram_addr to an InputAddr. */
  516. static u64 dram_addr_to_input_addr(struct mem_ctl_info *mci, u64 dram_addr)
  517. {
  518. struct amd64_pvt *pvt;
  519. int intlv_shift;
  520. u64 input_addr;
  521. pvt = mci->pvt_info;
  522. /*
  523. * See the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
  524. * concerning translating a DramAddr to an InputAddr.
  525. */
  526. intlv_shift = num_node_interleave_bits(pvt->dram_IntlvEn[0]);
  527. input_addr = ((dram_addr >> intlv_shift) & 0xffffff000ull) +
  528. (dram_addr & 0xfff);
  529. debugf2(" Intlv Shift=%d DramAddr=0x%lx maps to InputAddr=0x%lx\n",
  530. intlv_shift, (unsigned long)dram_addr,
  531. (unsigned long)input_addr);
  532. return input_addr;
  533. }
  534. /*
  535. * Translate the SysAddr represented by @sys_addr to an InputAddr. It is
  536. * assumed that @sys_addr maps to the node given by mci.
  537. */
  538. static u64 sys_addr_to_input_addr(struct mem_ctl_info *mci, u64 sys_addr)
  539. {
  540. u64 input_addr;
  541. input_addr =
  542. dram_addr_to_input_addr(mci, sys_addr_to_dram_addr(mci, sys_addr));
  543. debugf2("SysAdddr 0x%lx translates to InputAddr 0x%lx\n",
  544. (unsigned long)sys_addr, (unsigned long)input_addr);
  545. return input_addr;
  546. }
  547. /*
  548. * @input_addr is an InputAddr associated with the node represented by mci.
  549. * Translate @input_addr to a DramAddr and return the result.
  550. */
  551. static u64 input_addr_to_dram_addr(struct mem_ctl_info *mci, u64 input_addr)
  552. {
  553. struct amd64_pvt *pvt;
  554. int node_id, intlv_shift;
  555. u64 bits, dram_addr;
  556. u32 intlv_sel;
  557. /*
  558. * Near the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
  559. * shows how to translate a DramAddr to an InputAddr. Here we reverse
  560. * this procedure. When translating from a DramAddr to an InputAddr, the
  561. * bits used for node interleaving are discarded. Here we recover these
  562. * bits from the IntlvSel field of the DRAM Limit register (section
  563. * 3.4.4.2) for the node that input_addr is associated with.
  564. */
  565. pvt = mci->pvt_info;
  566. node_id = pvt->mc_node_id;
  567. BUG_ON((node_id < 0) || (node_id > 7));
  568. intlv_shift = num_node_interleave_bits(pvt->dram_IntlvEn[0]);
  569. if (intlv_shift == 0) {
  570. debugf1(" InputAddr 0x%lx translates to DramAddr of "
  571. "same value\n", (unsigned long)input_addr);
  572. return input_addr;
  573. }
  574. bits = ((input_addr & 0xffffff000ull) << intlv_shift) +
  575. (input_addr & 0xfff);
  576. intlv_sel = pvt->dram_IntlvSel[node_id] & ((1 << intlv_shift) - 1);
  577. dram_addr = bits + (intlv_sel << 12);
  578. debugf1("InputAddr 0x%lx translates to DramAddr 0x%lx "
  579. "(%d node interleave bits)\n", (unsigned long)input_addr,
  580. (unsigned long)dram_addr, intlv_shift);
  581. return dram_addr;
  582. }
  583. /*
  584. * @dram_addr is a DramAddr that maps to the node represented by mci. Convert
  585. * @dram_addr to a SysAddr.
  586. */
  587. static u64 dram_addr_to_sys_addr(struct mem_ctl_info *mci, u64 dram_addr)
  588. {
  589. struct amd64_pvt *pvt = mci->pvt_info;
  590. u64 hole_base, hole_offset, hole_size, base, limit, sys_addr;
  591. int ret = 0;
  592. ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
  593. &hole_size);
  594. if (!ret) {
  595. if ((dram_addr >= hole_base) &&
  596. (dram_addr < (hole_base + hole_size))) {
  597. sys_addr = dram_addr + hole_offset;
  598. debugf1("using DHAR to translate DramAddr 0x%lx to "
  599. "SysAddr 0x%lx\n", (unsigned long)dram_addr,
  600. (unsigned long)sys_addr);
  601. return sys_addr;
  602. }
  603. }
  604. amd64_get_base_and_limit(pvt, pvt->mc_node_id, &base, &limit);
  605. sys_addr = dram_addr + base;
  606. /*
  607. * The sys_addr we have computed up to this point is a 40-bit value
  608. * because the k8 deals with 40-bit values. However, the value we are
  609. * supposed to return is a full 64-bit physical address. The AMD
  610. * x86-64 architecture specifies that the most significant implemented
  611. * address bit through bit 63 of a physical address must be either all
  612. * 0s or all 1s. Therefore we sign-extend the 40-bit sys_addr to a
  613. * 64-bit value below. See section 3.4.2 of AMD publication 24592:
  614. * AMD x86-64 Architecture Programmer's Manual Volume 1 Application
  615. * Programming.
  616. */
  617. sys_addr |= ~((sys_addr & (1ull << 39)) - 1);
  618. debugf1(" Node %d, DramAddr 0x%lx to SysAddr 0x%lx\n",
  619. pvt->mc_node_id, (unsigned long)dram_addr,
  620. (unsigned long)sys_addr);
  621. return sys_addr;
  622. }
  623. /*
  624. * @input_addr is an InputAddr associated with the node given by mci. Translate
  625. * @input_addr to a SysAddr.
  626. */
  627. static inline u64 input_addr_to_sys_addr(struct mem_ctl_info *mci,
  628. u64 input_addr)
  629. {
  630. return dram_addr_to_sys_addr(mci,
  631. input_addr_to_dram_addr(mci, input_addr));
  632. }
  633. /*
  634. * Find the minimum and maximum InputAddr values that map to the given @csrow.
  635. * Pass back these values in *input_addr_min and *input_addr_max.
  636. */
  637. static void find_csrow_limits(struct mem_ctl_info *mci, int csrow,
  638. u64 *input_addr_min, u64 *input_addr_max)
  639. {
  640. struct amd64_pvt *pvt;
  641. u64 base, mask;
  642. pvt = mci->pvt_info;
  643. BUG_ON((csrow < 0) || (csrow >= pvt->cs_count));
  644. base = base_from_dct_base(pvt, csrow);
  645. mask = mask_from_dct_mask(pvt, csrow);
  646. *input_addr_min = base & ~mask;
  647. *input_addr_max = base | mask | pvt->dcs_mask_notused;
  648. }
  649. /* Map the Error address to a PAGE and PAGE OFFSET. */
  650. static inline void error_address_to_page_and_offset(u64 error_address,
  651. u32 *page, u32 *offset)
  652. {
  653. *page = (u32) (error_address >> PAGE_SHIFT);
  654. *offset = ((u32) error_address) & ~PAGE_MASK;
  655. }
  656. /*
  657. * @sys_addr is an error address (a SysAddr) extracted from the MCA NB Address
  658. * Low (section 3.6.4.5) and MCA NB Address High (section 3.6.4.6) registers
  659. * of a node that detected an ECC memory error. mci represents the node that
  660. * the error address maps to (possibly different from the node that detected
  661. * the error). Return the number of the csrow that sys_addr maps to, or -1 on
  662. * error.
  663. */
  664. static int sys_addr_to_csrow(struct mem_ctl_info *mci, u64 sys_addr)
  665. {
  666. int csrow;
  667. csrow = input_addr_to_csrow(mci, sys_addr_to_input_addr(mci, sys_addr));
  668. if (csrow == -1)
  669. amd64_mc_printk(mci, KERN_ERR,
  670. "Failed to translate InputAddr to csrow for "
  671. "address 0x%lx\n", (unsigned long)sys_addr);
  672. return csrow;
  673. }
  674. static int get_channel_from_ecc_syndrome(struct mem_ctl_info *, u16);
  675. static void amd64_cpu_display_info(struct amd64_pvt *pvt)
  676. {
  677. if (boot_cpu_data.x86 == 0x11)
  678. edac_printk(KERN_DEBUG, EDAC_MC, "F11h CPU detected\n");
  679. else if (boot_cpu_data.x86 == 0x10)
  680. edac_printk(KERN_DEBUG, EDAC_MC, "F10h CPU detected\n");
  681. else if (boot_cpu_data.x86 == 0xf)
  682. edac_printk(KERN_DEBUG, EDAC_MC, "%s detected\n",
  683. (pvt->ext_model >= K8_REV_F) ?
  684. "Rev F or later" : "Rev E or earlier");
  685. else
  686. /* we'll hardly ever ever get here */
  687. edac_printk(KERN_ERR, EDAC_MC, "Unknown cpu!\n");
  688. }
  689. /*
  690. * Determine if the DIMMs have ECC enabled. ECC is enabled ONLY if all the DIMMs
  691. * are ECC capable.
  692. */
  693. static enum edac_type amd64_determine_edac_cap(struct amd64_pvt *pvt)
  694. {
  695. int bit;
  696. enum dev_type edac_cap = EDAC_FLAG_NONE;
  697. bit = (boot_cpu_data.x86 > 0xf || pvt->ext_model >= K8_REV_F)
  698. ? 19
  699. : 17;
  700. if (pvt->dclr0 & BIT(bit))
  701. edac_cap = EDAC_FLAG_SECDED;
  702. return edac_cap;
  703. }
  704. static void amd64_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt);
  705. static void amd64_dump_dramcfg_low(u32 dclr, int chan)
  706. {
  707. debugf1("F2x%d90 (DRAM Cfg Low): 0x%08x\n", chan, dclr);
  708. debugf1(" DIMM type: %sbuffered; all DIMMs support ECC: %s\n",
  709. (dclr & BIT(16)) ? "un" : "",
  710. (dclr & BIT(19)) ? "yes" : "no");
  711. debugf1(" PAR/ERR parity: %s\n",
  712. (dclr & BIT(8)) ? "enabled" : "disabled");
  713. debugf1(" DCT 128bit mode width: %s\n",
  714. (dclr & BIT(11)) ? "128b" : "64b");
  715. debugf1(" x4 logical DIMMs present: L0: %s L1: %s L2: %s L3: %s\n",
  716. (dclr & BIT(12)) ? "yes" : "no",
  717. (dclr & BIT(13)) ? "yes" : "no",
  718. (dclr & BIT(14)) ? "yes" : "no",
  719. (dclr & BIT(15)) ? "yes" : "no");
  720. }
  721. /* Display and decode various NB registers for debug purposes. */
  722. static void amd64_dump_misc_regs(struct amd64_pvt *pvt)
  723. {
  724. int ganged;
  725. debugf1("F3xE8 (NB Cap): 0x%08x\n", pvt->nbcap);
  726. debugf1(" NB two channel DRAM capable: %s\n",
  727. (pvt->nbcap & K8_NBCAP_DCT_DUAL) ? "yes" : "no");
  728. debugf1(" ECC capable: %s, ChipKill ECC capable: %s\n",
  729. (pvt->nbcap & K8_NBCAP_SECDED) ? "yes" : "no",
  730. (pvt->nbcap & K8_NBCAP_CHIPKILL) ? "yes" : "no");
  731. amd64_dump_dramcfg_low(pvt->dclr0, 0);
  732. debugf1("F3xB0 (Online Spare): 0x%08x\n", pvt->online_spare);
  733. debugf1("F1xF0 (DRAM Hole Address): 0x%08x, base: 0x%08x, "
  734. "offset: 0x%08x\n",
  735. pvt->dhar,
  736. dhar_base(pvt->dhar),
  737. (boot_cpu_data.x86 == 0xf) ? k8_dhar_offset(pvt->dhar)
  738. : f10_dhar_offset(pvt->dhar));
  739. debugf1(" DramHoleValid: %s\n",
  740. (pvt->dhar & DHAR_VALID) ? "yes" : "no");
  741. /* everything below this point is Fam10h and above */
  742. if (boot_cpu_data.x86 == 0xf) {
  743. amd64_debug_display_dimm_sizes(0, pvt);
  744. return;
  745. }
  746. /* Only if NOT ganged does dclr1 have valid info */
  747. if (!dct_ganging_enabled(pvt))
  748. amd64_dump_dramcfg_low(pvt->dclr1, 1);
  749. /*
  750. * Determine if ganged and then dump memory sizes for first controller,
  751. * and if NOT ganged dump info for 2nd controller.
  752. */
  753. ganged = dct_ganging_enabled(pvt);
  754. amd64_debug_display_dimm_sizes(0, pvt);
  755. if (!ganged)
  756. amd64_debug_display_dimm_sizes(1, pvt);
  757. }
  758. /* Read in both of DBAM registers */
  759. static void amd64_read_dbam_reg(struct amd64_pvt *pvt)
  760. {
  761. amd64_read_pci_cfg(pvt->dram_f2_ctl, DBAM0, &pvt->dbam0);
  762. if (boot_cpu_data.x86 >= 0x10)
  763. amd64_read_pci_cfg(pvt->dram_f2_ctl, DBAM1, &pvt->dbam1);
  764. }
  765. /*
  766. * NOTE: CPU Revision Dependent code: Rev E and Rev F
  767. *
  768. * Set the DCSB and DCSM mask values depending on the CPU revision value. Also
  769. * set the shift factor for the DCSB and DCSM values.
  770. *
  771. * ->dcs_mask_notused, RevE:
  772. *
  773. * To find the max InputAddr for the csrow, start with the base address and set
  774. * all bits that are "don't care" bits in the test at the start of section
  775. * 3.5.4 (p. 84).
  776. *
  777. * The "don't care" bits are all set bits in the mask and all bits in the gaps
  778. * between bit ranges [35:25] and [19:13]. The value REV_E_DCS_NOTUSED_BITS
  779. * represents bits [24:20] and [12:0], which are all bits in the above-mentioned
  780. * gaps.
  781. *
  782. * ->dcs_mask_notused, RevF and later:
  783. *
  784. * To find the max InputAddr for the csrow, start with the base address and set
  785. * all bits that are "don't care" bits in the test at the start of NPT section
  786. * 4.5.4 (p. 87).
  787. *
  788. * The "don't care" bits are all set bits in the mask and all bits in the gaps
  789. * between bit ranges [36:27] and [21:13].
  790. *
  791. * The value REV_F_F1Xh_DCS_NOTUSED_BITS represents bits [26:22] and [12:0],
  792. * which are all bits in the above-mentioned gaps.
  793. */
  794. static void amd64_set_dct_base_and_mask(struct amd64_pvt *pvt)
  795. {
  796. if (boot_cpu_data.x86 == 0xf && pvt->ext_model < K8_REV_F) {
  797. pvt->dcsb_base = REV_E_DCSB_BASE_BITS;
  798. pvt->dcsm_mask = REV_E_DCSM_MASK_BITS;
  799. pvt->dcs_mask_notused = REV_E_DCS_NOTUSED_BITS;
  800. pvt->dcs_shift = REV_E_DCS_SHIFT;
  801. pvt->cs_count = 8;
  802. pvt->num_dcsm = 8;
  803. } else {
  804. pvt->dcsb_base = REV_F_F1Xh_DCSB_BASE_BITS;
  805. pvt->dcsm_mask = REV_F_F1Xh_DCSM_MASK_BITS;
  806. pvt->dcs_mask_notused = REV_F_F1Xh_DCS_NOTUSED_BITS;
  807. pvt->dcs_shift = REV_F_F1Xh_DCS_SHIFT;
  808. if (boot_cpu_data.x86 == 0x11) {
  809. pvt->cs_count = 4;
  810. pvt->num_dcsm = 2;
  811. } else {
  812. pvt->cs_count = 8;
  813. pvt->num_dcsm = 4;
  814. }
  815. }
  816. }
  817. /*
  818. * Function 2 Offset F10_DCSB0; read in the DCS Base and DCS Mask hw registers
  819. */
  820. static void amd64_read_dct_base_mask(struct amd64_pvt *pvt)
  821. {
  822. int cs, reg;
  823. amd64_set_dct_base_and_mask(pvt);
  824. for (cs = 0; cs < pvt->cs_count; cs++) {
  825. reg = K8_DCSB0 + (cs * 4);
  826. if (!amd64_read_pci_cfg(pvt->dram_f2_ctl, reg, &pvt->dcsb0[cs]))
  827. debugf0(" DCSB0[%d]=0x%08x reg: F2x%x\n",
  828. cs, pvt->dcsb0[cs], reg);
  829. /* If DCT are NOT ganged, then read in DCT1's base */
  830. if (boot_cpu_data.x86 >= 0x10 && !dct_ganging_enabled(pvt)) {
  831. reg = F10_DCSB1 + (cs * 4);
  832. if (!amd64_read_pci_cfg(pvt->dram_f2_ctl, reg,
  833. &pvt->dcsb1[cs]))
  834. debugf0(" DCSB1[%d]=0x%08x reg: F2x%x\n",
  835. cs, pvt->dcsb1[cs], reg);
  836. } else {
  837. pvt->dcsb1[cs] = 0;
  838. }
  839. }
  840. for (cs = 0; cs < pvt->num_dcsm; cs++) {
  841. reg = K8_DCSM0 + (cs * 4);
  842. if (!amd64_read_pci_cfg(pvt->dram_f2_ctl, reg, &pvt->dcsm0[cs]))
  843. debugf0(" DCSM0[%d]=0x%08x reg: F2x%x\n",
  844. cs, pvt->dcsm0[cs], reg);
  845. /* If DCT are NOT ganged, then read in DCT1's mask */
  846. if (boot_cpu_data.x86 >= 0x10 && !dct_ganging_enabled(pvt)) {
  847. reg = F10_DCSM1 + (cs * 4);
  848. if (!amd64_read_pci_cfg(pvt->dram_f2_ctl, reg,
  849. &pvt->dcsm1[cs]))
  850. debugf0(" DCSM1[%d]=0x%08x reg: F2x%x\n",
  851. cs, pvt->dcsm1[cs], reg);
  852. } else {
  853. pvt->dcsm1[cs] = 0;
  854. }
  855. }
  856. }
  857. static enum mem_type amd64_determine_memory_type(struct amd64_pvt *pvt)
  858. {
  859. enum mem_type type;
  860. if (boot_cpu_data.x86 >= 0x10 || pvt->ext_model >= K8_REV_F) {
  861. if (pvt->dchr0 & DDR3_MODE)
  862. type = (pvt->dclr0 & BIT(16)) ? MEM_DDR3 : MEM_RDDR3;
  863. else
  864. type = (pvt->dclr0 & BIT(16)) ? MEM_DDR2 : MEM_RDDR2;
  865. } else {
  866. type = (pvt->dclr0 & BIT(18)) ? MEM_DDR : MEM_RDDR;
  867. }
  868. debugf1(" Memory type is: %s\n", edac_mem_types[type]);
  869. return type;
  870. }
  871. /*
  872. * Read the DRAM Configuration Low register. It differs between CG, D & E revs
  873. * and the later RevF memory controllers (DDR vs DDR2)
  874. *
  875. * Return:
  876. * number of memory channels in operation
  877. * Pass back:
  878. * contents of the DCL0_LOW register
  879. */
  880. static int k8_early_channel_count(struct amd64_pvt *pvt)
  881. {
  882. int flag, err = 0;
  883. err = amd64_read_pci_cfg(pvt->dram_f2_ctl, F10_DCLR_0, &pvt->dclr0);
  884. if (err)
  885. return err;
  886. if ((boot_cpu_data.x86_model >> 4) >= K8_REV_F) {
  887. /* RevF (NPT) and later */
  888. flag = pvt->dclr0 & F10_WIDTH_128;
  889. } else {
  890. /* RevE and earlier */
  891. flag = pvt->dclr0 & REVE_WIDTH_128;
  892. }
  893. /* not used */
  894. pvt->dclr1 = 0;
  895. return (flag) ? 2 : 1;
  896. }
  897. /* extract the ERROR ADDRESS for the K8 CPUs */
  898. static u64 k8_get_error_address(struct mem_ctl_info *mci,
  899. struct err_regs *info)
  900. {
  901. return (((u64) (info->nbeah & 0xff)) << 32) +
  902. (info->nbeal & ~0x03);
  903. }
  904. /*
  905. * Read the Base and Limit registers for K8 based Memory controllers; extract
  906. * fields from the 'raw' reg into separate data fields
  907. *
  908. * Isolates: BASE, LIMIT, IntlvEn, IntlvSel, RW_EN
  909. */
  910. static void k8_read_dram_base_limit(struct amd64_pvt *pvt, int dram)
  911. {
  912. u32 low;
  913. u32 off = dram << 3; /* 8 bytes between DRAM entries */
  914. amd64_read_pci_cfg(pvt->addr_f1_ctl, K8_DRAM_BASE_LOW + off, &low);
  915. /* Extract parts into separate data entries */
  916. pvt->dram_base[dram] = ((u64) low & 0xFFFF0000) << 8;
  917. pvt->dram_IntlvEn[dram] = (low >> 8) & 0x7;
  918. pvt->dram_rw_en[dram] = (low & 0x3);
  919. amd64_read_pci_cfg(pvt->addr_f1_ctl, K8_DRAM_LIMIT_LOW + off, &low);
  920. /*
  921. * Extract parts into separate data entries. Limit is the HIGHEST memory
  922. * location of the region, so lower 24 bits need to be all ones
  923. */
  924. pvt->dram_limit[dram] = (((u64) low & 0xFFFF0000) << 8) | 0x00FFFFFF;
  925. pvt->dram_IntlvSel[dram] = (low >> 8) & 0x7;
  926. pvt->dram_DstNode[dram] = (low & 0x7);
  927. }
  928. static void k8_map_sysaddr_to_csrow(struct mem_ctl_info *mci,
  929. struct err_regs *info,
  930. u64 sys_addr)
  931. {
  932. struct mem_ctl_info *src_mci;
  933. unsigned short syndrome;
  934. int channel, csrow;
  935. u32 page, offset;
  936. /* Extract the syndrome parts and form a 16-bit syndrome */
  937. syndrome = HIGH_SYNDROME(info->nbsl) << 8;
  938. syndrome |= LOW_SYNDROME(info->nbsh);
  939. /* CHIPKILL enabled */
  940. if (info->nbcfg & K8_NBCFG_CHIPKILL) {
  941. channel = get_channel_from_ecc_syndrome(mci, syndrome);
  942. if (channel < 0) {
  943. /*
  944. * Syndrome didn't map, so we don't know which of the
  945. * 2 DIMMs is in error. So we need to ID 'both' of them
  946. * as suspect.
  947. */
  948. amd64_mc_printk(mci, KERN_WARNING,
  949. "unknown syndrome 0x%x - possible error "
  950. "reporting race\n", syndrome);
  951. edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
  952. return;
  953. }
  954. } else {
  955. /*
  956. * non-chipkill ecc mode
  957. *
  958. * The k8 documentation is unclear about how to determine the
  959. * channel number when using non-chipkill memory. This method
  960. * was obtained from email communication with someone at AMD.
  961. * (Wish the email was placed in this comment - norsk)
  962. */
  963. channel = ((sys_addr & BIT(3)) != 0);
  964. }
  965. /*
  966. * Find out which node the error address belongs to. This may be
  967. * different from the node that detected the error.
  968. */
  969. src_mci = find_mc_by_sys_addr(mci, sys_addr);
  970. if (!src_mci) {
  971. amd64_mc_printk(mci, KERN_ERR,
  972. "failed to map error address 0x%lx to a node\n",
  973. (unsigned long)sys_addr);
  974. edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
  975. return;
  976. }
  977. /* Now map the sys_addr to a CSROW */
  978. csrow = sys_addr_to_csrow(src_mci, sys_addr);
  979. if (csrow < 0) {
  980. edac_mc_handle_ce_no_info(src_mci, EDAC_MOD_STR);
  981. } else {
  982. error_address_to_page_and_offset(sys_addr, &page, &offset);
  983. edac_mc_handle_ce(src_mci, page, offset, syndrome, csrow,
  984. channel, EDAC_MOD_STR);
  985. }
  986. }
  987. static int k8_dbam_to_chip_select(struct amd64_pvt *pvt, int cs_mode)
  988. {
  989. int *dbam_map;
  990. if (pvt->ext_model >= K8_REV_F)
  991. dbam_map = ddr2_dbam;
  992. else if (pvt->ext_model >= K8_REV_D)
  993. dbam_map = ddr2_dbam_revD;
  994. else
  995. dbam_map = ddr2_dbam_revCG;
  996. return dbam_map[cs_mode];
  997. }
  998. /*
  999. * Get the number of DCT channels in use.
  1000. *
  1001. * Return:
  1002. * number of Memory Channels in operation
  1003. * Pass back:
  1004. * contents of the DCL0_LOW register
  1005. */
  1006. static int f10_early_channel_count(struct amd64_pvt *pvt)
  1007. {
  1008. int dbams[] = { DBAM0, DBAM1 };
  1009. int i, j, channels = 0;
  1010. u32 dbam;
  1011. /* If we are in 128 bit mode, then we are using 2 channels */
  1012. if (pvt->dclr0 & F10_WIDTH_128) {
  1013. channels = 2;
  1014. return channels;
  1015. }
  1016. /*
  1017. * Need to check if in unganged mode: In such, there are 2 channels,
  1018. * but they are not in 128 bit mode and thus the above 'dclr0' status
  1019. * bit will be OFF.
  1020. *
  1021. * Need to check DCT0[0] and DCT1[0] to see if only one of them has
  1022. * their CSEnable bit on. If so, then SINGLE DIMM case.
  1023. */
  1024. debugf0("Data width is not 128 bits - need more decoding\n");
  1025. /*
  1026. * Check DRAM Bank Address Mapping values for each DIMM to see if there
  1027. * is more than just one DIMM present in unganged mode. Need to check
  1028. * both controllers since DIMMs can be placed in either one.
  1029. */
  1030. for (i = 0; i < ARRAY_SIZE(dbams); i++) {
  1031. if (amd64_read_pci_cfg(pvt->dram_f2_ctl, dbams[i], &dbam))
  1032. goto err_reg;
  1033. for (j = 0; j < 4; j++) {
  1034. if (DBAM_DIMM(j, dbam) > 0) {
  1035. channels++;
  1036. break;
  1037. }
  1038. }
  1039. }
  1040. if (channels > 2)
  1041. channels = 2;
  1042. debugf0("MCT channel count: %d\n", channels);
  1043. return channels;
  1044. err_reg:
  1045. return -1;
  1046. }
  1047. static int f10_dbam_to_chip_select(struct amd64_pvt *pvt, int cs_mode)
  1048. {
  1049. int *dbam_map;
  1050. if (pvt->dchr0 & DDR3_MODE || pvt->dchr1 & DDR3_MODE)
  1051. dbam_map = ddr3_dbam;
  1052. else
  1053. dbam_map = ddr2_dbam;
  1054. return dbam_map[cs_mode];
  1055. }
  1056. /* Enable extended configuration access via 0xCF8 feature */
  1057. static void amd64_setup(struct amd64_pvt *pvt)
  1058. {
  1059. u32 reg;
  1060. amd64_read_pci_cfg(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, &reg);
  1061. pvt->flags.cf8_extcfg = !!(reg & F10_NB_CFG_LOW_ENABLE_EXT_CFG);
  1062. reg |= F10_NB_CFG_LOW_ENABLE_EXT_CFG;
  1063. pci_write_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, reg);
  1064. }
  1065. /* Restore the extended configuration access via 0xCF8 feature */
  1066. static void amd64_teardown(struct amd64_pvt *pvt)
  1067. {
  1068. u32 reg;
  1069. amd64_read_pci_cfg(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, &reg);
  1070. reg &= ~F10_NB_CFG_LOW_ENABLE_EXT_CFG;
  1071. if (pvt->flags.cf8_extcfg)
  1072. reg |= F10_NB_CFG_LOW_ENABLE_EXT_CFG;
  1073. pci_write_config_dword(pvt->misc_f3_ctl, F10_NB_CFG_HIGH, reg);
  1074. }
  1075. static u64 f10_get_error_address(struct mem_ctl_info *mci,
  1076. struct err_regs *info)
  1077. {
  1078. return (((u64) (info->nbeah & 0xffff)) << 32) +
  1079. (info->nbeal & ~0x01);
  1080. }
  1081. /*
  1082. * Read the Base and Limit registers for F10 based Memory controllers. Extract
  1083. * fields from the 'raw' reg into separate data fields.
  1084. *
  1085. * Isolates: BASE, LIMIT, IntlvEn, IntlvSel, RW_EN.
  1086. */
  1087. static void f10_read_dram_base_limit(struct amd64_pvt *pvt, int dram)
  1088. {
  1089. u32 high_offset, low_offset, high_base, low_base, high_limit, low_limit;
  1090. low_offset = K8_DRAM_BASE_LOW + (dram << 3);
  1091. high_offset = F10_DRAM_BASE_HIGH + (dram << 3);
  1092. /* read the 'raw' DRAM BASE Address register */
  1093. amd64_read_pci_cfg(pvt->addr_f1_ctl, low_offset, &low_base);
  1094. /* Read from the ECS data register */
  1095. amd64_read_pci_cfg(pvt->addr_f1_ctl, high_offset, &high_base);
  1096. /* Extract parts into separate data entries */
  1097. pvt->dram_rw_en[dram] = (low_base & 0x3);
  1098. if (pvt->dram_rw_en[dram] == 0)
  1099. return;
  1100. pvt->dram_IntlvEn[dram] = (low_base >> 8) & 0x7;
  1101. pvt->dram_base[dram] = (((u64)high_base & 0x000000FF) << 40) |
  1102. (((u64)low_base & 0xFFFF0000) << 8);
  1103. low_offset = K8_DRAM_LIMIT_LOW + (dram << 3);
  1104. high_offset = F10_DRAM_LIMIT_HIGH + (dram << 3);
  1105. /* read the 'raw' LIMIT registers */
  1106. amd64_read_pci_cfg(pvt->addr_f1_ctl, low_offset, &low_limit);
  1107. /* Read from the ECS data register for the HIGH portion */
  1108. amd64_read_pci_cfg(pvt->addr_f1_ctl, high_offset, &high_limit);
  1109. pvt->dram_DstNode[dram] = (low_limit & 0x7);
  1110. pvt->dram_IntlvSel[dram] = (low_limit >> 8) & 0x7;
  1111. /*
  1112. * Extract address values and form a LIMIT address. Limit is the HIGHEST
  1113. * memory location of the region, so low 24 bits need to be all ones.
  1114. */
  1115. pvt->dram_limit[dram] = (((u64)high_limit & 0x000000FF) << 40) |
  1116. (((u64) low_limit & 0xFFFF0000) << 8) |
  1117. 0x00FFFFFF;
  1118. }
  1119. static void f10_read_dram_ctl_register(struct amd64_pvt *pvt)
  1120. {
  1121. if (!amd64_read_pci_cfg(pvt->dram_f2_ctl, F10_DCTL_SEL_LOW,
  1122. &pvt->dram_ctl_select_low)) {
  1123. debugf0("F2x110 (DCTL Sel. Low): 0x%08x, "
  1124. "High range addresses at: 0x%x\n",
  1125. pvt->dram_ctl_select_low,
  1126. dct_sel_baseaddr(pvt));
  1127. debugf0(" DCT mode: %s, All DCTs on: %s\n",
  1128. (dct_ganging_enabled(pvt) ? "ganged" : "unganged"),
  1129. (dct_dram_enabled(pvt) ? "yes" : "no"));
  1130. if (!dct_ganging_enabled(pvt))
  1131. debugf0(" Address range split per DCT: %s\n",
  1132. (dct_high_range_enabled(pvt) ? "yes" : "no"));
  1133. debugf0(" DCT data interleave for ECC: %s, "
  1134. "DRAM cleared since last warm reset: %s\n",
  1135. (dct_data_intlv_enabled(pvt) ? "enabled" : "disabled"),
  1136. (dct_memory_cleared(pvt) ? "yes" : "no"));
  1137. debugf0(" DCT channel interleave: %s, "
  1138. "DCT interleave bits selector: 0x%x\n",
  1139. (dct_interleave_enabled(pvt) ? "enabled" : "disabled"),
  1140. dct_sel_interleave_addr(pvt));
  1141. }
  1142. amd64_read_pci_cfg(pvt->dram_f2_ctl, F10_DCTL_SEL_HIGH,
  1143. &pvt->dram_ctl_select_high);
  1144. }
  1145. /*
  1146. * determine channel based on the interleaving mode: F10h BKDG, 2.8.9 Memory
  1147. * Interleaving Modes.
  1148. */
  1149. static u32 f10_determine_channel(struct amd64_pvt *pvt, u64 sys_addr,
  1150. int hi_range_sel, u32 intlv_en)
  1151. {
  1152. u32 cs, temp, dct_sel_high = (pvt->dram_ctl_select_low >> 1) & 1;
  1153. if (dct_ganging_enabled(pvt))
  1154. cs = 0;
  1155. else if (hi_range_sel)
  1156. cs = dct_sel_high;
  1157. else if (dct_interleave_enabled(pvt)) {
  1158. /*
  1159. * see F2x110[DctSelIntLvAddr] - channel interleave mode
  1160. */
  1161. if (dct_sel_interleave_addr(pvt) == 0)
  1162. cs = sys_addr >> 6 & 1;
  1163. else if ((dct_sel_interleave_addr(pvt) >> 1) & 1) {
  1164. temp = hweight_long((u32) ((sys_addr >> 16) & 0x1F)) % 2;
  1165. if (dct_sel_interleave_addr(pvt) & 1)
  1166. cs = (sys_addr >> 9 & 1) ^ temp;
  1167. else
  1168. cs = (sys_addr >> 6 & 1) ^ temp;
  1169. } else if (intlv_en & 4)
  1170. cs = sys_addr >> 15 & 1;
  1171. else if (intlv_en & 2)
  1172. cs = sys_addr >> 14 & 1;
  1173. else if (intlv_en & 1)
  1174. cs = sys_addr >> 13 & 1;
  1175. else
  1176. cs = sys_addr >> 12 & 1;
  1177. } else if (dct_high_range_enabled(pvt) && !dct_ganging_enabled(pvt))
  1178. cs = ~dct_sel_high & 1;
  1179. else
  1180. cs = 0;
  1181. return cs;
  1182. }
  1183. static inline u32 f10_map_intlv_en_to_shift(u32 intlv_en)
  1184. {
  1185. if (intlv_en == 1)
  1186. return 1;
  1187. else if (intlv_en == 3)
  1188. return 2;
  1189. else if (intlv_en == 7)
  1190. return 3;
  1191. return 0;
  1192. }
  1193. /* See F10h BKDG, 2.8.10.2 DctSelBaseOffset Programming */
  1194. static inline u64 f10_get_base_addr_offset(u64 sys_addr, int hi_range_sel,
  1195. u32 dct_sel_base_addr,
  1196. u64 dct_sel_base_off,
  1197. u32 hole_valid, u32 hole_off,
  1198. u64 dram_base)
  1199. {
  1200. u64 chan_off;
  1201. if (hi_range_sel) {
  1202. if (!(dct_sel_base_addr & 0xFFFF0000) &&
  1203. hole_valid && (sys_addr >= 0x100000000ULL))
  1204. chan_off = hole_off << 16;
  1205. else
  1206. chan_off = dct_sel_base_off;
  1207. } else {
  1208. if (hole_valid && (sys_addr >= 0x100000000ULL))
  1209. chan_off = hole_off << 16;
  1210. else
  1211. chan_off = dram_base & 0xFFFFF8000000ULL;
  1212. }
  1213. return (sys_addr & 0x0000FFFFFFFFFFC0ULL) -
  1214. (chan_off & 0x0000FFFFFF800000ULL);
  1215. }
  1216. /* Hack for the time being - Can we get this from BIOS?? */
  1217. #define CH0SPARE_RANK 0
  1218. #define CH1SPARE_RANK 1
  1219. /*
  1220. * checks if the csrow passed in is marked as SPARED, if so returns the new
  1221. * spare row
  1222. */
  1223. static inline int f10_process_possible_spare(int csrow,
  1224. u32 cs, struct amd64_pvt *pvt)
  1225. {
  1226. u32 swap_done;
  1227. u32 bad_dram_cs;
  1228. /* Depending on channel, isolate respective SPARING info */
  1229. if (cs) {
  1230. swap_done = F10_ONLINE_SPARE_SWAPDONE1(pvt->online_spare);
  1231. bad_dram_cs = F10_ONLINE_SPARE_BADDRAM_CS1(pvt->online_spare);
  1232. if (swap_done && (csrow == bad_dram_cs))
  1233. csrow = CH1SPARE_RANK;
  1234. } else {
  1235. swap_done = F10_ONLINE_SPARE_SWAPDONE0(pvt->online_spare);
  1236. bad_dram_cs = F10_ONLINE_SPARE_BADDRAM_CS0(pvt->online_spare);
  1237. if (swap_done && (csrow == bad_dram_cs))
  1238. csrow = CH0SPARE_RANK;
  1239. }
  1240. return csrow;
  1241. }
  1242. /*
  1243. * Iterate over the DRAM DCT "base" and "mask" registers looking for a
  1244. * SystemAddr match on the specified 'ChannelSelect' and 'NodeID'
  1245. *
  1246. * Return:
  1247. * -EINVAL: NOT FOUND
  1248. * 0..csrow = Chip-Select Row
  1249. */
  1250. static int f10_lookup_addr_in_dct(u32 in_addr, u32 nid, u32 cs)
  1251. {
  1252. struct mem_ctl_info *mci;
  1253. struct amd64_pvt *pvt;
  1254. u32 cs_base, cs_mask;
  1255. int cs_found = -EINVAL;
  1256. int csrow;
  1257. mci = mci_lookup[nid];
  1258. if (!mci)
  1259. return cs_found;
  1260. pvt = mci->pvt_info;
  1261. debugf1("InputAddr=0x%x channelselect=%d\n", in_addr, cs);
  1262. for (csrow = 0; csrow < pvt->cs_count; csrow++) {
  1263. cs_base = amd64_get_dct_base(pvt, cs, csrow);
  1264. if (!(cs_base & K8_DCSB_CS_ENABLE))
  1265. continue;
  1266. /*
  1267. * We have an ENABLED CSROW, Isolate just the MASK bits of the
  1268. * target: [28:19] and [13:5], which map to [36:27] and [21:13]
  1269. * of the actual address.
  1270. */
  1271. cs_base &= REV_F_F1Xh_DCSB_BASE_BITS;
  1272. /*
  1273. * Get the DCT Mask, and ENABLE the reserved bits: [18:16] and
  1274. * [4:0] to become ON. Then mask off bits [28:0] ([36:8])
  1275. */
  1276. cs_mask = amd64_get_dct_mask(pvt, cs, csrow);
  1277. debugf1(" CSROW=%d CSBase=0x%x RAW CSMask=0x%x\n",
  1278. csrow, cs_base, cs_mask);
  1279. cs_mask = (cs_mask | 0x0007C01F) & 0x1FFFFFFF;
  1280. debugf1(" Final CSMask=0x%x\n", cs_mask);
  1281. debugf1(" (InputAddr & ~CSMask)=0x%x "
  1282. "(CSBase & ~CSMask)=0x%x\n",
  1283. (in_addr & ~cs_mask), (cs_base & ~cs_mask));
  1284. if ((in_addr & ~cs_mask) == (cs_base & ~cs_mask)) {
  1285. cs_found = f10_process_possible_spare(csrow, cs, pvt);
  1286. debugf1(" MATCH csrow=%d\n", cs_found);
  1287. break;
  1288. }
  1289. }
  1290. return cs_found;
  1291. }
  1292. /* For a given @dram_range, check if @sys_addr falls within it. */
  1293. static int f10_match_to_this_node(struct amd64_pvt *pvt, int dram_range,
  1294. u64 sys_addr, int *nid, int *chan_sel)
  1295. {
  1296. int node_id, cs_found = -EINVAL, high_range = 0;
  1297. u32 intlv_en, intlv_sel, intlv_shift, hole_off;
  1298. u32 hole_valid, tmp, dct_sel_base, channel;
  1299. u64 dram_base, chan_addr, dct_sel_base_off;
  1300. dram_base = pvt->dram_base[dram_range];
  1301. intlv_en = pvt->dram_IntlvEn[dram_range];
  1302. node_id = pvt->dram_DstNode[dram_range];
  1303. intlv_sel = pvt->dram_IntlvSel[dram_range];
  1304. debugf1("(dram=%d) Base=0x%llx SystemAddr= 0x%llx Limit=0x%llx\n",
  1305. dram_range, dram_base, sys_addr, pvt->dram_limit[dram_range]);
  1306. /*
  1307. * This assumes that one node's DHAR is the same as all the other
  1308. * nodes' DHAR.
  1309. */
  1310. hole_off = (pvt->dhar & 0x0000FF80);
  1311. hole_valid = (pvt->dhar & 0x1);
  1312. dct_sel_base_off = (pvt->dram_ctl_select_high & 0xFFFFFC00) << 16;
  1313. debugf1(" HoleOffset=0x%x HoleValid=0x%x IntlvSel=0x%x\n",
  1314. hole_off, hole_valid, intlv_sel);
  1315. if (intlv_en ||
  1316. (intlv_sel != ((sys_addr >> 12) & intlv_en)))
  1317. return -EINVAL;
  1318. dct_sel_base = dct_sel_baseaddr(pvt);
  1319. /*
  1320. * check whether addresses >= DctSelBaseAddr[47:27] are to be used to
  1321. * select between DCT0 and DCT1.
  1322. */
  1323. if (dct_high_range_enabled(pvt) &&
  1324. !dct_ganging_enabled(pvt) &&
  1325. ((sys_addr >> 27) >= (dct_sel_base >> 11)))
  1326. high_range = 1;
  1327. channel = f10_determine_channel(pvt, sys_addr, high_range, intlv_en);
  1328. chan_addr = f10_get_base_addr_offset(sys_addr, high_range, dct_sel_base,
  1329. dct_sel_base_off, hole_valid,
  1330. hole_off, dram_base);
  1331. intlv_shift = f10_map_intlv_en_to_shift(intlv_en);
  1332. /* remove Node ID (in case of memory interleaving) */
  1333. tmp = chan_addr & 0xFC0;
  1334. chan_addr = ((chan_addr >> intlv_shift) & 0xFFFFFFFFF000ULL) | tmp;
  1335. /* remove channel interleave and hash */
  1336. if (dct_interleave_enabled(pvt) &&
  1337. !dct_high_range_enabled(pvt) &&
  1338. !dct_ganging_enabled(pvt)) {
  1339. if (dct_sel_interleave_addr(pvt) != 1)
  1340. chan_addr = (chan_addr >> 1) & 0xFFFFFFFFFFFFFFC0ULL;
  1341. else {
  1342. tmp = chan_addr & 0xFC0;
  1343. chan_addr = ((chan_addr & 0xFFFFFFFFFFFFC000ULL) >> 1)
  1344. | tmp;
  1345. }
  1346. }
  1347. debugf1(" (ChannelAddrLong=0x%llx) >> 8 becomes InputAddr=0x%x\n",
  1348. chan_addr, (u32)(chan_addr >> 8));
  1349. cs_found = f10_lookup_addr_in_dct(chan_addr >> 8, node_id, channel);
  1350. if (cs_found >= 0) {
  1351. *nid = node_id;
  1352. *chan_sel = channel;
  1353. }
  1354. return cs_found;
  1355. }
  1356. static int f10_translate_sysaddr_to_cs(struct amd64_pvt *pvt, u64 sys_addr,
  1357. int *node, int *chan_sel)
  1358. {
  1359. int dram_range, cs_found = -EINVAL;
  1360. u64 dram_base, dram_limit;
  1361. for (dram_range = 0; dram_range < DRAM_REG_COUNT; dram_range++) {
  1362. if (!pvt->dram_rw_en[dram_range])
  1363. continue;
  1364. dram_base = pvt->dram_base[dram_range];
  1365. dram_limit = pvt->dram_limit[dram_range];
  1366. if ((dram_base <= sys_addr) && (sys_addr <= dram_limit)) {
  1367. cs_found = f10_match_to_this_node(pvt, dram_range,
  1368. sys_addr, node,
  1369. chan_sel);
  1370. if (cs_found >= 0)
  1371. break;
  1372. }
  1373. }
  1374. return cs_found;
  1375. }
  1376. /*
  1377. * For reference see "2.8.5 Routing DRAM Requests" in F10 BKDG. This code maps
  1378. * a @sys_addr to NodeID, DCT (channel) and chip select (CSROW).
  1379. *
  1380. * The @sys_addr is usually an error address received from the hardware
  1381. * (MCX_ADDR).
  1382. */
  1383. static void f10_map_sysaddr_to_csrow(struct mem_ctl_info *mci,
  1384. struct err_regs *info,
  1385. u64 sys_addr)
  1386. {
  1387. struct amd64_pvt *pvt = mci->pvt_info;
  1388. u32 page, offset;
  1389. unsigned short syndrome;
  1390. int nid, csrow, chan = 0;
  1391. csrow = f10_translate_sysaddr_to_cs(pvt, sys_addr, &nid, &chan);
  1392. if (csrow < 0) {
  1393. edac_mc_handle_ce_no_info(mci, EDAC_MOD_STR);
  1394. return;
  1395. }
  1396. error_address_to_page_and_offset(sys_addr, &page, &offset);
  1397. syndrome = HIGH_SYNDROME(info->nbsl) << 8;
  1398. syndrome |= LOW_SYNDROME(info->nbsh);
  1399. /*
  1400. * We need the syndromes for channel detection only when we're
  1401. * ganged. Otherwise @chan should already contain the channel at
  1402. * this point.
  1403. */
  1404. if (dct_ganging_enabled(pvt) && (pvt->nbcfg & K8_NBCFG_CHIPKILL))
  1405. chan = get_channel_from_ecc_syndrome(mci, syndrome);
  1406. if (chan >= 0)
  1407. edac_mc_handle_ce(mci, page, offset, syndrome, csrow, chan,
  1408. EDAC_MOD_STR);
  1409. else
  1410. /*
  1411. * Channel unknown, report all channels on this CSROW as failed.
  1412. */
  1413. for (chan = 0; chan < mci->csrows[csrow].nr_channels; chan++)
  1414. edac_mc_handle_ce(mci, page, offset, syndrome,
  1415. csrow, chan, EDAC_MOD_STR);
  1416. }
  1417. /*
  1418. * debug routine to display the memory sizes of all logical DIMMs and its
  1419. * CSROWs as well
  1420. */
  1421. static void amd64_debug_display_dimm_sizes(int ctrl, struct amd64_pvt *pvt)
  1422. {
  1423. int dimm, size0, size1, factor = 0;
  1424. u32 dbam;
  1425. u32 *dcsb;
  1426. if (boot_cpu_data.x86 == 0xf) {
  1427. if (pvt->dclr0 & F10_WIDTH_128)
  1428. factor = 1;
  1429. /* K8 families < revF not supported yet */
  1430. if (pvt->ext_model < K8_REV_F)
  1431. return;
  1432. else
  1433. WARN_ON(ctrl != 0);
  1434. }
  1435. debugf1("F2x%d80 (DRAM Bank Address Mapping): 0x%08x\n",
  1436. ctrl, ctrl ? pvt->dbam1 : pvt->dbam0);
  1437. dbam = ctrl ? pvt->dbam1 : pvt->dbam0;
  1438. dcsb = ctrl ? pvt->dcsb1 : pvt->dcsb0;
  1439. edac_printk(KERN_DEBUG, EDAC_MC, "DCT%d chip selects:\n", ctrl);
  1440. /* Dump memory sizes for DIMM and its CSROWs */
  1441. for (dimm = 0; dimm < 4; dimm++) {
  1442. size0 = 0;
  1443. if (dcsb[dimm*2] & K8_DCSB_CS_ENABLE)
  1444. size0 = pvt->ops->dbam_to_cs(pvt, DBAM_DIMM(dimm, dbam));
  1445. size1 = 0;
  1446. if (dcsb[dimm*2 + 1] & K8_DCSB_CS_ENABLE)
  1447. size1 = pvt->ops->dbam_to_cs(pvt, DBAM_DIMM(dimm, dbam));
  1448. edac_printk(KERN_DEBUG, EDAC_MC, " %d: %5dMB %d: %5dMB\n",
  1449. dimm * 2, size0 << factor,
  1450. dimm * 2 + 1, size1 << factor);
  1451. }
  1452. }
  1453. /*
  1454. * There currently are 3 typ…