power5-pmu.c 16 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627
  1. /*
  2. * Performance counter support for POWER5 (not POWER5++) processors.
  3. *
  4. * Copyright 2009 Paul Mackerras, IBM Corporation.
  5. *
  6. * This program is free software; you can redistribute it and/or
  7. * modify it under the terms of the GNU General Public License
  8. * as published by the Free Software Foundation; either version
  9. * 2 of the License, or (at your option) any later version.
  10. */
  11. #include <linux/kernel.h>
  12. #include <linux/perf_counter.h>
  13. #include <linux/string.h>
  14. #include <asm/reg.h>
  15. #include <asm/cputable.h>
  16. /*
  17. * Bits in event code for POWER5 (not POWER5++)
  18. */
  19. #define PM_PMC_SH 20 /* PMC number (1-based) for direct events */
  20. #define PM_PMC_MSK 0xf
  21. #define PM_PMC_MSKS (PM_PMC_MSK << PM_PMC_SH)
  22. #define PM_UNIT_SH 16 /* TTMMUX number and setting - unit select */
  23. #define PM_UNIT_MSK 0xf
  24. #define PM_BYTE_SH 12 /* Byte number of event bus to use */
  25. #define PM_BYTE_MSK 7
  26. #define PM_GRS_SH 8 /* Storage subsystem mux select */
  27. #define PM_GRS_MSK 7
  28. #define PM_BUSEVENT_MSK 0x80 /* Set if event uses event bus */
  29. #define PM_PMCSEL_MSK 0x7f
  30. /* Values in PM_UNIT field */
  31. #define PM_FPU 0
  32. #define PM_ISU0 1
  33. #define PM_IFU 2
  34. #define PM_ISU1 3
  35. #define PM_IDU 4
  36. #define PM_ISU0_ALT 6
  37. #define PM_GRS 7
  38. #define PM_LSU0 8
  39. #define PM_LSU1 0xc
  40. #define PM_LASTUNIT 0xc
  41. /*
  42. * Bits in MMCR1 for POWER5
  43. */
  44. #define MMCR1_TTM0SEL_SH 62
  45. #define MMCR1_TTM1SEL_SH 60
  46. #define MMCR1_TTM2SEL_SH 58
  47. #define MMCR1_TTM3SEL_SH 56
  48. #define MMCR1_TTMSEL_MSK 3
  49. #define MMCR1_TD_CP_DBG0SEL_SH 54
  50. #define MMCR1_TD_CP_DBG1SEL_SH 52
  51. #define MMCR1_TD_CP_DBG2SEL_SH 50
  52. #define MMCR1_TD_CP_DBG3SEL_SH 48
  53. #define MMCR1_GRS_L2SEL_SH 46
  54. #define MMCR1_GRS_L2SEL_MSK 3
  55. #define MMCR1_GRS_L3SEL_SH 44
  56. #define MMCR1_GRS_L3SEL_MSK 3
  57. #define MMCR1_GRS_MCSEL_SH 41
  58. #define MMCR1_GRS_MCSEL_MSK 7
  59. #define MMCR1_GRS_FABSEL_SH 39
  60. #define MMCR1_GRS_FABSEL_MSK 3
  61. #define MMCR1_PMC1_ADDER_SEL_SH 35
  62. #define MMCR1_PMC2_ADDER_SEL_SH 34
  63. #define MMCR1_PMC3_ADDER_SEL_SH 33
  64. #define MMCR1_PMC4_ADDER_SEL_SH 32
  65. #define MMCR1_PMC1SEL_SH 25
  66. #define MMCR1_PMC2SEL_SH 17
  67. #define MMCR1_PMC3SEL_SH 9
  68. #define MMCR1_PMC4SEL_SH 1
  69. #define MMCR1_PMCSEL_SH(n) (MMCR1_PMC1SEL_SH - (n) * 8)
  70. #define MMCR1_PMCSEL_MSK 0x7f
  71. /*
  72. * Bits in MMCRA
  73. */
  74. /*
  75. * Layout of constraint bits:
  76. * 6666555555555544444444443333333333222222222211111111110000000000
  77. * 3210987654321098765432109876543210987654321098765432109876543210
  78. * <><>[ ><><>< ><> [ >[ >[ >< >< >< >< ><><><><><><>
  79. * T0T1 NC G0G1G2 G3 UC PS1PS2 B0 B1 B2 B3 P6P5P4P3P2P1
  80. *
  81. * T0 - TTM0 constraint
  82. * 54-55: TTM0SEL value (0=FPU, 2=IFU, 3=ISU1) 0xc0_0000_0000_0000
  83. *
  84. * T1 - TTM1 constraint
  85. * 52-53: TTM1SEL value (0=IDU, 3=GRS) 0x30_0000_0000_0000
  86. *
  87. * NC - number of counters
  88. * 51: NC error 0x0008_0000_0000_0000
  89. * 48-50: number of events needing PMC1-4 0x0007_0000_0000_0000
  90. *
  91. * G0..G3 - GRS mux constraints
  92. * 46-47: GRS_L2SEL value
  93. * 44-45: GRS_L3SEL value
  94. * 41-44: GRS_MCSEL value
  95. * 39-40: GRS_FABSEL value
  96. * Note that these match up with their bit positions in MMCR1
  97. *
  98. * UC - unit constraint: can't have all three of FPU|IFU|ISU1, ISU0, IDU|GRS
  99. * 37: UC3 error 0x20_0000_0000
  100. * 36: FPU|IFU|ISU1 events needed 0x10_0000_0000
  101. * 35: ISU0 events needed 0x08_0000_0000
  102. * 34: IDU|GRS events needed 0x04_0000_0000
  103. *
  104. * PS1
  105. * 33: PS1 error 0x2_0000_0000
  106. * 31-32: count of events needing PMC1/2 0x1_8000_0000
  107. *
  108. * PS2
  109. * 30: PS2 error 0x4000_0000
  110. * 28-29: count of events needing PMC3/4 0x3000_0000
  111. *
  112. * B0
  113. * 24-27: Byte 0 event source 0x0f00_0000
  114. * Encoding as for the event code
  115. *
  116. * B1, B2, B3
  117. * 20-23, 16-19, 12-15: Byte 1, 2, 3 event sources
  118. *
  119. * P1..P6
  120. * 0-11: Count of events needing PMC1..PMC6
  121. */
  122. static const int grsel_shift[8] = {
  123. MMCR1_GRS_L2SEL_SH, MMCR1_GRS_L2SEL_SH, MMCR1_GRS_L2SEL_SH,
  124. MMCR1_GRS_L3SEL_SH, MMCR1_GRS_L3SEL_SH, MMCR1_GRS_L3SEL_SH,
  125. MMCR1_GRS_MCSEL_SH, MMCR1_GRS_FABSEL_SH
  126. };
  127. /* Masks and values for using events from the various units */
  128. static unsigned long unit_cons[PM_LASTUNIT+1][2] = {
  129. [PM_FPU] = { 0xc0002000000000ul, 0x00001000000000ul },
  130. [PM_ISU0] = { 0x00002000000000ul, 0x00000800000000ul },
  131. [PM_ISU1] = { 0xc0002000000000ul, 0xc0001000000000ul },
  132. [PM_IFU] = { 0xc0002000000000ul, 0x80001000000000ul },
  133. [PM_IDU] = { 0x30002000000000ul, 0x00000400000000ul },
  134. [PM_GRS] = { 0x30002000000000ul, 0x30000400000000ul },
  135. };
  136. static int power5_get_constraint(u64 event, unsigned long *maskp,
  137. unsigned long *valp)
  138. {
  139. int pmc, byte, unit, sh;
  140. int bit, fmask;
  141. unsigned long mask = 0, value = 0;
  142. int grp = -1;
  143. pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
  144. if (pmc) {
  145. if (pmc > 6)
  146. return -1;
  147. sh = (pmc - 1) * 2;
  148. mask |= 2 << sh;
  149. value |= 1 << sh;
  150. if (pmc <= 4)
  151. grp = (pmc - 1) >> 1;
  152. else if (event != 0x500009 && event != 0x600005)
  153. return -1;
  154. }
  155. if (event & PM_BUSEVENT_MSK) {
  156. unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
  157. if (unit > PM_LASTUNIT)
  158. return -1;
  159. if (unit == PM_ISU0_ALT)
  160. unit = PM_ISU0;
  161. mask |= unit_cons[unit][0];
  162. value |= unit_cons[unit][1];
  163. byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
  164. if (byte >= 4) {
  165. if (unit != PM_LSU1)
  166. return -1;
  167. /* Map LSU1 low word (bytes 4-7) to unit LSU1+1 */
  168. ++unit;
  169. byte &= 3;
  170. }
  171. if (unit == PM_GRS) {
  172. bit = event & 7;
  173. fmask = (bit == 6)? 7: 3;
  174. sh = grsel_shift[bit];
  175. mask |= (unsigned long)fmask << sh;
  176. value |= (unsigned long)((event >> PM_GRS_SH) & fmask)
  177. << sh;
  178. }
  179. /*
  180. * Bus events on bytes 0 and 2 can be counted
  181. * on PMC1/2; bytes 1 and 3 on PMC3/4.
  182. */
  183. if (!pmc)
  184. grp = byte & 1;
  185. /* Set byte lane select field */
  186. mask |= 0xfUL << (24 - 4 * byte);
  187. value |= (unsigned long)unit << (24 - 4 * byte);
  188. }
  189. if (grp == 0) {
  190. /* increment PMC1/2 field */
  191. mask |= 0x200000000ul;
  192. value |= 0x080000000ul;
  193. } else if (grp == 1) {
  194. /* increment PMC3/4 field */
  195. mask |= 0x40000000ul;
  196. value |= 0x10000000ul;
  197. }
  198. if (pmc < 5) {
  199. /* need a counter from PMC1-4 set */
  200. mask |= 0x8000000000000ul;
  201. value |= 0x1000000000000ul;
  202. }
  203. *maskp = mask;
  204. *valp = value;
  205. return 0;
  206. }
  207. #define MAX_ALT 3 /* at most 3 alternatives for any event */
  208. static const unsigned int event_alternatives[][MAX_ALT] = {
  209. { 0x120e4, 0x400002 }, /* PM_GRP_DISP_REJECT */
  210. { 0x410c7, 0x441084 }, /* PM_THRD_L2MISS_BOTH_CYC */
  211. { 0x100005, 0x600005 }, /* PM_RUN_CYC */
  212. { 0x100009, 0x200009, 0x500009 }, /* PM_INST_CMPL */
  213. { 0x300009, 0x400009 }, /* PM_INST_DISP */
  214. };
  215. /*
  216. * Scan the alternatives table for a match and return the
  217. * index into the alternatives table if found, else -1.
  218. */
  219. static int find_alternative(u64 event)
  220. {
  221. int i, j;
  222. for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
  223. if (event < event_alternatives[i][0])
  224. break;
  225. for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
  226. if (event == event_alternatives[i][j])
  227. return i;
  228. }
  229. return -1;
  230. }
  231. static const unsigned char bytedecode_alternatives[4][4] = {
  232. /* PMC 1 */ { 0x21, 0x23, 0x25, 0x27 },
  233. /* PMC 2 */ { 0x07, 0x17, 0x0e, 0x1e },
  234. /* PMC 3 */ { 0x20, 0x22, 0x24, 0x26 },
  235. /* PMC 4 */ { 0x07, 0x17, 0x0e, 0x1e }
  236. };
  237. /*
  238. * Some direct events for decodes of event bus byte 3 have alternative
  239. * PMCSEL values on other counters. This returns the alternative
  240. * event code for those that do, or -1 otherwise.
  241. */
  242. static s64 find_alternative_bdecode(u64 event)
  243. {
  244. int pmc, altpmc, pp, j;
  245. pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
  246. if (pmc == 0 || pmc > 4)
  247. return -1;
  248. altpmc = 5 - pmc; /* 1 <-> 4, 2 <-> 3 */
  249. pp = event & PM_PMCSEL_MSK;
  250. for (j = 0; j < 4; ++j) {
  251. if (bytedecode_alternatives[pmc - 1][j] == pp) {
  252. return (event & ~(PM_PMC_MSKS | PM_PMCSEL_MSK)) |
  253. (altpmc << PM_PMC_SH) |
  254. bytedecode_alternatives[altpmc - 1][j];
  255. }
  256. }
  257. return -1;
  258. }
  259. static int power5_get_alternatives(u64 event, unsigned int flags, u64 alt[])
  260. {
  261. int i, j, nalt = 1;
  262. s64 ae;
  263. alt[0] = event;
  264. nalt = 1;
  265. i = find_alternative(event);
  266. if (i >= 0) {
  267. for (j = 0; j < MAX_ALT; ++j) {
  268. ae = event_alternatives[i][j];
  269. if (ae && ae != event)
  270. alt[nalt++] = ae;
  271. }
  272. } else {
  273. ae = find_alternative_bdecode(event);
  274. if (ae > 0)
  275. alt[nalt++] = ae;
  276. }
  277. return nalt;
  278. }
  279. /*
  280. * Map of which direct events on which PMCs are marked instruction events.
  281. * Indexed by PMCSEL value, bit i (LE) set if PMC i is a marked event.
  282. * Bit 0 is set if it is marked for all PMCs.
  283. * The 0x80 bit indicates a byte decode PMCSEL value.
  284. */
  285. static unsigned char direct_event_is_marked[0x28] = {
  286. 0, /* 00 */
  287. 0x1f, /* 01 PM_IOPS_CMPL */
  288. 0x2, /* 02 PM_MRK_GRP_DISP */
  289. 0xe, /* 03 PM_MRK_ST_CMPL, PM_MRK_ST_GPS, PM_MRK_ST_CMPL_INT */
  290. 0, /* 04 */
  291. 0x1c, /* 05 PM_MRK_BRU_FIN, PM_MRK_INST_FIN, PM_MRK_CRU_FIN */
  292. 0x80, /* 06 */
  293. 0x80, /* 07 */
  294. 0, 0, 0,/* 08 - 0a */
  295. 0x18, /* 0b PM_THRESH_TIMEO, PM_MRK_GRP_TIMEO */
  296. 0, /* 0c */
  297. 0x80, /* 0d */
  298. 0x80, /* 0e */
  299. 0, /* 0f */
  300. 0, /* 10 */
  301. 0x14, /* 11 PM_MRK_GRP_BR_REDIR, PM_MRK_GRP_IC_MISS */
  302. 0, /* 12 */
  303. 0x10, /* 13 PM_MRK_GRP_CMPL */
  304. 0x1f, /* 14 PM_GRP_MRK, PM_MRK_{FXU,FPU,LSU}_FIN */
  305. 0x2, /* 15 PM_MRK_GRP_ISSUED */
  306. 0x80, /* 16 */
  307. 0x80, /* 17 */
  308. 0, 0, 0, 0, 0,
  309. 0x80, /* 1d */
  310. 0x80, /* 1e */
  311. 0, /* 1f */
  312. 0x80, /* 20 */
  313. 0x80, /* 21 */
  314. 0x80, /* 22 */
  315. 0x80, /* 23 */
  316. 0x80, /* 24 */
  317. 0x80, /* 25 */
  318. 0x80, /* 26 */
  319. 0x80, /* 27 */
  320. };
  321. /*
  322. * Returns 1 if event counts things relating to marked instructions
  323. * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
  324. */
  325. static int power5_marked_instr_event(u64 event)
  326. {
  327. int pmc, psel;
  328. int bit, byte, unit;
  329. u32 mask;
  330. pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
  331. psel = event & PM_PMCSEL_MSK;
  332. if (pmc >= 5)
  333. return 0;
  334. bit = -1;
  335. if (psel < sizeof(direct_event_is_marked)) {
  336. if (direct_event_is_marked[psel] & (1 << pmc))
  337. return 1;
  338. if (direct_event_is_marked[psel] & 0x80)
  339. bit = 4;
  340. else if (psel == 0x08)
  341. bit = pmc - 1;
  342. else if (psel == 0x10)
  343. bit = 4 - pmc;
  344. else if (psel == 0x1b && (pmc == 1 || pmc == 3))
  345. bit = 4;
  346. } else if ((psel & 0x58) == 0x40)
  347. bit = psel & 7;
  348. if (!(event & PM_BUSEVENT_MSK))
  349. return 0;
  350. byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
  351. unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
  352. if (unit == PM_LSU0) {
  353. /* byte 1 bits 0-7, byte 2 bits 0,2-4,6 */
  354. mask = 0x5dff00;
  355. } else if (unit == PM_LSU1 && byte >= 4) {
  356. byte -= 4;
  357. /* byte 4 bits 1,3,5,7, byte 5 bits 6-7, byte 7 bits 0-4,6 */
  358. mask = 0x5f00c0aa;
  359. } else
  360. return 0;
  361. return (mask >> (byte * 8 + bit)) & 1;
  362. }
  363. static int power5_compute_mmcr(u64 event[], int n_ev,
  364. unsigned int hwc[], unsigned long mmcr[])
  365. {
  366. unsigned long mmcr1 = 0;
  367. unsigned long mmcra = 0;
  368. unsigned int pmc, unit, byte, psel;
  369. unsigned int ttm, grp;
  370. int i, isbus, bit, grsel;
  371. unsigned int pmc_inuse = 0;
  372. unsigned int pmc_grp_use[2];
  373. unsigned char busbyte[4];
  374. unsigned char unituse[16];
  375. int ttmuse;
  376. if (n_ev > 6)
  377. return -1;
  378. /* First pass to count resource use */
  379. pmc_grp_use[0] = pmc_grp_use[1] = 0;
  380. memset(busbyte, 0, sizeof(busbyte));
  381. memset(unituse, 0, sizeof(unituse));
  382. for (i = 0; i < n_ev; ++i) {
  383. pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
  384. if (pmc) {
  385. if (pmc > 6)
  386. return -1;
  387. if (pmc_inuse & (1 << (pmc - 1)))
  388. return -1;
  389. pmc_inuse |= 1 << (pmc - 1);
  390. /* count 1/2 vs 3/4 use */
  391. if (pmc <= 4)
  392. ++pmc_grp_use[(pmc - 1) >> 1];
  393. }
  394. if (event[i] & PM_BUSEVENT_MSK) {
  395. unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
  396. byte = (event[i] >> PM_BYTE_SH) & PM_BYTE_MSK;
  397. if (unit > PM_LASTUNIT)
  398. return -1;
  399. if (unit == PM_ISU0_ALT)
  400. unit = PM_ISU0;
  401. if (byte >= 4) {
  402. if (unit != PM_LSU1)
  403. return -1;
  404. ++unit;
  405. byte &= 3;
  406. }
  407. if (!pmc)
  408. ++pmc_grp_use[byte & 1];
  409. if (busbyte[byte] && busbyte[byte] != unit)
  410. return -1;
  411. busbyte[byte] = unit;
  412. unituse[unit] = 1;
  413. }
  414. }
  415. if (pmc_grp_use[0] > 2 || pmc_grp_use[1] > 2)
  416. return -1;
  417. /*
  418. * Assign resources and set multiplexer selects.
  419. *
  420. * PM_ISU0 can go either on TTM0 or TTM1, but that's the only
  421. * choice we have to deal with.
  422. */
  423. if (unituse[PM_ISU0] &
  424. (unituse[PM_FPU] | unituse[PM_IFU] | unituse[PM_ISU1])) {
  425. unituse[PM_ISU0_ALT] = 1; /* move ISU to TTM1 */
  426. unituse[PM_ISU0] = 0;
  427. }
  428. /* Set TTM[01]SEL fields. */
  429. ttmuse = 0;
  430. for (i = PM_FPU; i <= PM_ISU1; ++i) {
  431. if (!unituse[i])
  432. continue;
  433. if (ttmuse++)
  434. return -1;
  435. mmcr1 |= (unsigned long)i << MMCR1_TTM0SEL_SH;
  436. }
  437. ttmuse = 0;
  438. for (; i <= PM_GRS; ++i) {
  439. if (!unituse[i])
  440. continue;
  441. if (ttmuse++)
  442. return -1;
  443. mmcr1 |= (unsigned long)(i & 3) << MMCR1_TTM1SEL_SH;
  444. }
  445. if (ttmuse > 1)
  446. return -1;
  447. /* Set byte lane select fields, TTM[23]SEL and GRS_*SEL. */
  448. for (byte = 0; byte < 4; ++byte) {
  449. unit = busbyte[byte];
  450. if (!unit)
  451. continue;
  452. if (unit == PM_ISU0 && unituse[PM_ISU0_ALT]) {
  453. /* get ISU0 through TTM1 rather than TTM0 */
  454. unit = PM_ISU0_ALT;
  455. } else if (unit == PM_LSU1 + 1) {
  456. /* select lower word of LSU1 for this byte */
  457. mmcr1 |= 1ul << (MMCR1_TTM3SEL_SH + 3 - byte);
  458. }
  459. ttm = unit >> 2;
  460. mmcr1 |= (unsigned long)ttm
  461. << (MMCR1_TD_CP_DBG0SEL_SH - 2 * byte);
  462. }
  463. /* Second pass: assign PMCs, set PMCxSEL and PMCx_ADDER_SEL fields */
  464. for (i = 0; i < n_ev; ++i) {
  465. pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
  466. unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
  467. byte = (event[i] >> PM_BYTE_SH) & PM_BYTE_MSK;
  468. psel = event[i] & PM_PMCSEL_MSK;
  469. isbus = event[i] & PM_BUSEVENT_MSK;
  470. if (!pmc) {
  471. /* Bus event or any-PMC direct event */
  472. for (pmc = 0; pmc < 4; ++pmc) {
  473. if (pmc_inuse & (1 << pmc))
  474. continue;
  475. grp = (pmc >> 1) & 1;
  476. if (isbus) {
  477. if (grp == (byte & 1))
  478. break;
  479. } else if (pmc_grp_use[grp] < 2) {
  480. ++pmc_grp_use[grp];
  481. break;
  482. }
  483. }
  484. pmc_inuse |= 1 << pmc;
  485. } else if (pmc <= 4) {
  486. /* Direct event */
  487. --pmc;
  488. if ((psel == 8 || psel == 0x10) && isbus && (byte & 2))
  489. /* add events on higher-numbered bus */
  490. mmcr1 |= 1ul << (MMCR1_PMC1_ADDER_SEL_SH - pmc);
  491. } else {
  492. /* Instructions or run cycles on PMC5/6 */
  493. --pmc;
  494. }
  495. if (isbus && unit == PM_GRS) {
  496. bit = psel & 7;
  497. grsel = (event[i] >> PM_GRS_SH) & PM_GRS_MSK;
  498. mmcr1 |= (unsigned long)grsel << grsel_shift[bit];
  499. }
  500. if (power5_marked_instr_event(event[i]))
  501. mmcra |= MMCRA_SAMPLE_ENABLE;
  502. if (pmc <= 3)
  503. mmcr1 |= psel << MMCR1_PMCSEL_SH(pmc);
  504. hwc[i] = pmc;
  505. }
  506. /* Return MMCRx values */
  507. mmcr[0] = 0;
  508. if (pmc_inuse & 1)
  509. mmcr[0] = MMCR0_PMC1CE;
  510. if (pmc_inuse & 0x3e)
  511. mmcr[0] |= MMCR0_PMCjCE;
  512. mmcr[1] = mmcr1;
  513. mmcr[2] = mmcra;
  514. return 0;
  515. }
  516. static void power5_disable_pmc(unsigned int pmc, unsigned long mmcr[])
  517. {
  518. if (pmc <= 3)
  519. mmcr[1] &= ~(0x7fUL << MMCR1_PMCSEL_SH(pmc));
  520. }
  521. static int power5_generic_events[] = {
  522. [PERF_COUNT_HW_CPU_CYCLES] = 0xf,
  523. [PERF_COUNT_HW_INSTRUCTIONS] = 0x100009,
  524. [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4c1090, /* LD_REF_L1 */
  525. [PERF_COUNT_HW_CACHE_MISSES] = 0x3c1088, /* LD_MISS_L1 */
  526. [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x230e4, /* BR_ISSUED */
  527. [PERF_COUNT_HW_BRANCH_MISSES] = 0x230e5, /* BR_MPRED_CR */
  528. };
  529. #define C(x) PERF_COUNT_HW_CACHE_##x
  530. /*
  531. * Table of generalized cache-related events.
  532. * 0 means not supported, -1 means nonsensical, other values
  533. * are event codes.
  534. */
  535. static int power5_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
  536. [C(L1D)] = { /* RESULT_ACCESS RESULT_MISS */
  537. [C(OP_READ)] = { 0x4c1090, 0x3c1088 },
  538. [C(OP_WRITE)] = { 0x3c1090, 0xc10c3 },
  539. [C(OP_PREFETCH)] = { 0xc70e7, 0 },
  540. },
  541. [C(L1I)] = { /* RESULT_ACCESS RESULT_MISS */
  542. [C(OP_READ)] = { 0, 0 },
  543. [C(OP_WRITE)] = { -1, -1 },
  544. [C(OP_PREFETCH)] = { 0, 0 },
  545. },
  546. [C(LL)] = { /* RESULT_ACCESS RESULT_MISS */
  547. [C(OP_READ)] = { 0, 0x3c309b },
  548. [C(OP_WRITE)] = { 0, 0 },
  549. [C(OP_PREFETCH)] = { 0xc50c3, 0 },
  550. },
  551. [C(DTLB)] = { /* RESULT_ACCESS RESULT_MISS */
  552. [C(OP_READ)] = { 0x2c4090, 0x800c4 },
  553. [C(OP_WRITE)] = { -1, -1 },
  554. [C(OP_PREFETCH)] = { -1, -1 },
  555. },
  556. [C(ITLB)] = { /* RESULT_ACCESS RESULT_MISS */
  557. [C(OP_READ)] = { 0, 0x800c0 },
  558. [C(OP_WRITE)] = { -1, -1 },
  559. [C(OP_PREFETCH)] = { -1, -1 },
  560. },
  561. [C(BPU)] = { /* RESULT_ACCESS RESULT_MISS */
  562. [C(OP_READ)] = { 0x230e4, 0x230e5 },
  563. [C(OP_WRITE)] = { -1, -1 },
  564. [C(OP_PREFETCH)] = { -1, -1 },
  565. },
  566. };
  567. static struct power_pmu power5_pmu = {
  568. .name = "POWER5",
  569. .n_counter = 6,
  570. .max_alternatives = MAX_ALT,
  571. .add_fields = 0x7000090000555ul,
  572. .test_adder = 0x3000490000000ul,
  573. .compute_mmcr = power5_compute_mmcr,
  574. .get_constraint = power5_get_constraint,
  575. .get_alternatives = power5_get_alternatives,
  576. .disable_pmc = power5_disable_pmc,
  577. .n_generic = ARRAY_SIZE(power5_generic_events),
  578. .generic_events = power5_generic_events,
  579. .cache_events = &power5_cache_events,
  580. };
  581. static int init_power5_pmu(void)
  582. {
  583. if (strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power5"))
  584. return -ENODEV;
  585. return register_power_pmu(&power5_pmu);
  586. }
  587. arch_initcall(init_power5_pmu);