power5-pmu.c 17 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629
  1. // SPDX-License-Identifier: GPL-2.0-or-later
  2. /*
  3. * Performance counter support for POWER5 (not POWER5++) processors.
  4. *
  5. * Copyright 2009 Paul Mackerras, IBM Corporation.
  6. */
  7. #include <linux/kernel.h>
  8. #include <linux/perf_event.h>
  9. #include <linux/string.h>
  10. #include <asm/reg.h>
  11. #include <asm/cputable.h>
  12. #include "internal.h"
  13. /*
  14. * Bits in event code for POWER5 (not POWER5++)
  15. */
  16. #define PM_PMC_SH 20 /* PMC number (1-based) for direct events */
  17. #define PM_PMC_MSK 0xf
  18. #define PM_PMC_MSKS (PM_PMC_MSK << PM_PMC_SH)
  19. #define PM_UNIT_SH 16 /* TTMMUX number and setting - unit select */
  20. #define PM_UNIT_MSK 0xf
  21. #define PM_BYTE_SH 12 /* Byte number of event bus to use */
  22. #define PM_BYTE_MSK 7
  23. #define PM_GRS_SH 8 /* Storage subsystem mux select */
  24. #define PM_GRS_MSK 7
  25. #define PM_BUSEVENT_MSK 0x80 /* Set if event uses event bus */
  26. #define PM_PMCSEL_MSK 0x7f
  27. /* Values in PM_UNIT field */
  28. #define PM_FPU 0
  29. #define PM_ISU0 1
  30. #define PM_IFU 2
  31. #define PM_ISU1 3
  32. #define PM_IDU 4
  33. #define PM_ISU0_ALT 6
  34. #define PM_GRS 7
  35. #define PM_LSU0 8
  36. #define PM_LSU1 0xc
  37. #define PM_LASTUNIT 0xc
  38. /*
  39. * Bits in MMCR1 for POWER5
  40. */
  41. #define MMCR1_TTM0SEL_SH 62
  42. #define MMCR1_TTM1SEL_SH 60
  43. #define MMCR1_TTM2SEL_SH 58
  44. #define MMCR1_TTM3SEL_SH 56
  45. #define MMCR1_TTMSEL_MSK 3
  46. #define MMCR1_TD_CP_DBG0SEL_SH 54
  47. #define MMCR1_TD_CP_DBG1SEL_SH 52
  48. #define MMCR1_TD_CP_DBG2SEL_SH 50
  49. #define MMCR1_TD_CP_DBG3SEL_SH 48
  50. #define MMCR1_GRS_L2SEL_SH 46
  51. #define MMCR1_GRS_L2SEL_MSK 3
  52. #define MMCR1_GRS_L3SEL_SH 44
  53. #define MMCR1_GRS_L3SEL_MSK 3
  54. #define MMCR1_GRS_MCSEL_SH 41
  55. #define MMCR1_GRS_MCSEL_MSK 7
  56. #define MMCR1_GRS_FABSEL_SH 39
  57. #define MMCR1_GRS_FABSEL_MSK 3
  58. #define MMCR1_PMC1_ADDER_SEL_SH 35
  59. #define MMCR1_PMC2_ADDER_SEL_SH 34
  60. #define MMCR1_PMC3_ADDER_SEL_SH 33
  61. #define MMCR1_PMC4_ADDER_SEL_SH 32
  62. #define MMCR1_PMC1SEL_SH 25
  63. #define MMCR1_PMC2SEL_SH 17
  64. #define MMCR1_PMC3SEL_SH 9
  65. #define MMCR1_PMC4SEL_SH 1
  66. #define MMCR1_PMCSEL_SH(n) (MMCR1_PMC1SEL_SH - (n) * 8)
  67. #define MMCR1_PMCSEL_MSK 0x7f
  68. /*
  69. * Layout of constraint bits:
  70. * 6666555555555544444444443333333333222222222211111111110000000000
  71. * 3210987654321098765432109876543210987654321098765432109876543210
  72. * <><>[ ><><>< ><> [ >[ >[ >< >< >< >< ><><><><><><>
  73. * T0T1 NC G0G1G2 G3 UC PS1PS2 B0 B1 B2 B3 P6P5P4P3P2P1
  74. *
  75. * T0 - TTM0 constraint
  76. * 54-55: TTM0SEL value (0=FPU, 2=IFU, 3=ISU1) 0xc0_0000_0000_0000
  77. *
  78. * T1 - TTM1 constraint
  79. * 52-53: TTM1SEL value (0=IDU, 3=GRS) 0x30_0000_0000_0000
  80. *
  81. * NC - number of counters
  82. * 51: NC error 0x0008_0000_0000_0000
  83. * 48-50: number of events needing PMC1-4 0x0007_0000_0000_0000
  84. *
  85. * G0..G3 - GRS mux constraints
  86. * 46-47: GRS_L2SEL value
  87. * 44-45: GRS_L3SEL value
  88. * 41-44: GRS_MCSEL value
  89. * 39-40: GRS_FABSEL value
  90. * Note that these match up with their bit positions in MMCR1
  91. *
  92. * UC - unit constraint: can't have all three of FPU|IFU|ISU1, ISU0, IDU|GRS
  93. * 37: UC3 error 0x20_0000_0000
  94. * 36: FPU|IFU|ISU1 events needed 0x10_0000_0000
  95. * 35: ISU0 events needed 0x08_0000_0000
  96. * 34: IDU|GRS events needed 0x04_0000_0000
  97. *
  98. * PS1
  99. * 33: PS1 error 0x2_0000_0000
  100. * 31-32: count of events needing PMC1/2 0x1_8000_0000
  101. *
  102. * PS2
  103. * 30: PS2 error 0x4000_0000
  104. * 28-29: count of events needing PMC3/4 0x3000_0000
  105. *
  106. * B0
  107. * 24-27: Byte 0 event source 0x0f00_0000
  108. * Encoding as for the event code
  109. *
  110. * B1, B2, B3
  111. * 20-23, 16-19, 12-15: Byte 1, 2, 3 event sources
  112. *
  113. * P1..P6
  114. * 0-11: Count of events needing PMC1..PMC6
  115. */
  116. static const int grsel_shift[8] = {
  117. MMCR1_GRS_L2SEL_SH, MMCR1_GRS_L2SEL_SH, MMCR1_GRS_L2SEL_SH,
  118. MMCR1_GRS_L3SEL_SH, MMCR1_GRS_L3SEL_SH, MMCR1_GRS_L3SEL_SH,
  119. MMCR1_GRS_MCSEL_SH, MMCR1_GRS_FABSEL_SH
  120. };
  121. /* Masks and values for using events from the various units */
  122. static unsigned long unit_cons[PM_LASTUNIT+1][2] = {
  123. [PM_FPU] = { 0xc0002000000000ul, 0x00001000000000ul },
  124. [PM_ISU0] = { 0x00002000000000ul, 0x00000800000000ul },
  125. [PM_ISU1] = { 0xc0002000000000ul, 0xc0001000000000ul },
  126. [PM_IFU] = { 0xc0002000000000ul, 0x80001000000000ul },
  127. [PM_IDU] = { 0x30002000000000ul, 0x00000400000000ul },
  128. [PM_GRS] = { 0x30002000000000ul, 0x30000400000000ul },
  129. };
  130. static int power5_get_constraint(u64 event, unsigned long *maskp,
  131. unsigned long *valp, u64 event_config1 __maybe_unused)
  132. {
  133. int pmc, byte, unit, sh;
  134. int bit, fmask;
  135. unsigned long mask = 0, value = 0;
  136. int grp = -1;
  137. pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
  138. if (pmc) {
  139. if (pmc > 6)
  140. return -1;
  141. sh = (pmc - 1) * 2;
  142. mask |= 2 << sh;
  143. value |= 1 << sh;
  144. if (pmc <= 4)
  145. grp = (pmc - 1) >> 1;
  146. else if (event != 0x500009 && event != 0x600005)
  147. return -1;
  148. }
  149. if (event & PM_BUSEVENT_MSK) {
  150. unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
  151. if (unit > PM_LASTUNIT)
  152. return -1;
  153. if (unit == PM_ISU0_ALT)
  154. unit = PM_ISU0;
  155. mask |= unit_cons[unit][0];
  156. value |= unit_cons[unit][1];
  157. byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
  158. if (byte >= 4) {
  159. if (unit != PM_LSU1)
  160. return -1;
  161. /* Map LSU1 low word (bytes 4-7) to unit LSU1+1 */
  162. ++unit;
  163. byte &= 3;
  164. }
  165. if (unit == PM_GRS) {
  166. bit = event & 7;
  167. fmask = (bit == 6)? 7: 3;
  168. sh = grsel_shift[bit];
  169. mask |= (unsigned long)fmask << sh;
  170. value |= (unsigned long)((event >> PM_GRS_SH) & fmask)
  171. << sh;
  172. }
  173. /*
  174. * Bus events on bytes 0 and 2 can be counted
  175. * on PMC1/2; bytes 1 and 3 on PMC3/4.
  176. */
  177. if (!pmc)
  178. grp = byte & 1;
  179. /* Set byte lane select field */
  180. mask |= 0xfUL << (24 - 4 * byte);
  181. value |= (unsigned long)unit << (24 - 4 * byte);
  182. }
  183. if (grp == 0) {
  184. /* increment PMC1/2 field */
  185. mask |= 0x200000000ul;
  186. value |= 0x080000000ul;
  187. } else if (grp == 1) {
  188. /* increment PMC3/4 field */
  189. mask |= 0x40000000ul;
  190. value |= 0x10000000ul;
  191. }
  192. if (pmc < 5) {
  193. /* need a counter from PMC1-4 set */
  194. mask |= 0x8000000000000ul;
  195. value |= 0x1000000000000ul;
  196. }
  197. *maskp = mask;
  198. *valp = value;
  199. return 0;
  200. }
  201. #define MAX_ALT 3 /* at most 3 alternatives for any event */
  202. static const unsigned int event_alternatives[][MAX_ALT] = {
  203. { 0x120e4, 0x400002 }, /* PM_GRP_DISP_REJECT */
  204. { 0x410c7, 0x441084 }, /* PM_THRD_L2MISS_BOTH_CYC */
  205. { 0x100005, 0x600005 }, /* PM_RUN_CYC */
  206. { 0x100009, 0x200009, 0x500009 }, /* PM_INST_CMPL */
  207. { 0x300009, 0x400009 }, /* PM_INST_DISP */
  208. };
  209. /*
  210. * Scan the alternatives table for a match and return the
  211. * index into the alternatives table if found, else -1.
  212. */
  213. static int find_alternative(u64 event)
  214. {
  215. int i, j;
  216. for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
  217. if (event < event_alternatives[i][0])
  218. break;
  219. for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
  220. if (event == event_alternatives[i][j])
  221. return i;
  222. }
  223. return -1;
  224. }
  225. static const unsigned char bytedecode_alternatives[4][4] = {
  226. /* PMC 1 */ { 0x21, 0x23, 0x25, 0x27 },
  227. /* PMC 2 */ { 0x07, 0x17, 0x0e, 0x1e },
  228. /* PMC 3 */ { 0x20, 0x22, 0x24, 0x26 },
  229. /* PMC 4 */ { 0x07, 0x17, 0x0e, 0x1e }
  230. };
  231. /*
  232. * Some direct events for decodes of event bus byte 3 have alternative
  233. * PMCSEL values on other counters. This returns the alternative
  234. * event code for those that do, or -1 otherwise.
  235. */
  236. static s64 find_alternative_bdecode(u64 event)
  237. {
  238. int pmc, altpmc, pp, j;
  239. pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
  240. if (pmc == 0 || pmc > 4)
  241. return -1;
  242. altpmc = 5 - pmc; /* 1 <-> 4, 2 <-> 3 */
  243. pp = event & PM_PMCSEL_MSK;
  244. for (j = 0; j < 4; ++j) {
  245. if (bytedecode_alternatives[pmc - 1][j] == pp) {
  246. return (event & ~(PM_PMC_MSKS | PM_PMCSEL_MSK)) |
  247. (altpmc << PM_PMC_SH) |
  248. bytedecode_alternatives[altpmc - 1][j];
  249. }
  250. }
  251. return -1;
  252. }
  253. static int power5_get_alternatives(u64 event, unsigned int flags, u64 alt[])
  254. {
  255. int i, j, nalt = 1;
  256. s64 ae;
  257. alt[0] = event;
  258. nalt = 1;
  259. i = find_alternative(event);
  260. if (i >= 0) {
  261. for (j = 0; j < MAX_ALT; ++j) {
  262. ae = event_alternatives[i][j];
  263. if (ae && ae != event)
  264. alt[nalt++] = ae;
  265. }
  266. } else {
  267. ae = find_alternative_bdecode(event);
  268. if (ae > 0)
  269. alt[nalt++] = ae;
  270. }
  271. return nalt;
  272. }
  273. /*
  274. * Map of which direct events on which PMCs are marked instruction events.
  275. * Indexed by PMCSEL value, bit i (LE) set if PMC i is a marked event.
  276. * Bit 0 is set if it is marked for all PMCs.
  277. * The 0x80 bit indicates a byte decode PMCSEL value.
  278. */
  279. static unsigned char direct_event_is_marked[0x28] = {
  280. 0, /* 00 */
  281. 0x1f, /* 01 PM_IOPS_CMPL */
  282. 0x2, /* 02 PM_MRK_GRP_DISP */
  283. 0xe, /* 03 PM_MRK_ST_CMPL, PM_MRK_ST_GPS, PM_MRK_ST_CMPL_INT */
  284. 0, /* 04 */
  285. 0x1c, /* 05 PM_MRK_BRU_FIN, PM_MRK_INST_FIN, PM_MRK_CRU_FIN */
  286. 0x80, /* 06 */
  287. 0x80, /* 07 */
  288. 0, 0, 0,/* 08 - 0a */
  289. 0x18, /* 0b PM_THRESH_TIMEO, PM_MRK_GRP_TIMEO */
  290. 0, /* 0c */
  291. 0x80, /* 0d */
  292. 0x80, /* 0e */
  293. 0, /* 0f */
  294. 0, /* 10 */
  295. 0x14, /* 11 PM_MRK_GRP_BR_REDIR, PM_MRK_GRP_IC_MISS */
  296. 0, /* 12 */
  297. 0x10, /* 13 PM_MRK_GRP_CMPL */
  298. 0x1f, /* 14 PM_GRP_MRK, PM_MRK_{FXU,FPU,LSU}_FIN */
  299. 0x2, /* 15 PM_MRK_GRP_ISSUED */
  300. 0x80, /* 16 */
  301. 0x80, /* 17 */
  302. 0, 0, 0, 0, 0,
  303. 0x80, /* 1d */
  304. 0x80, /* 1e */
  305. 0, /* 1f */
  306. 0x80, /* 20 */
  307. 0x80, /* 21 */
  308. 0x80, /* 22 */
  309. 0x80, /* 23 */
  310. 0x80, /* 24 */
  311. 0x80, /* 25 */
  312. 0x80, /* 26 */
  313. 0x80, /* 27 */
  314. };
  315. /*
  316. * Returns 1 if event counts things relating to marked instructions
  317. * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
  318. */
  319. static int power5_marked_instr_event(u64 event)
  320. {
  321. int pmc, psel;
  322. int bit, byte, unit;
  323. u32 mask;
  324. pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
  325. psel = event & PM_PMCSEL_MSK;
  326. if (pmc >= 5)
  327. return 0;
  328. bit = -1;
  329. if (psel < sizeof(direct_event_is_marked)) {
  330. if (direct_event_is_marked[psel] & (1 << pmc))
  331. return 1;
  332. if (direct_event_is_marked[psel] & 0x80)
  333. bit = 4;
  334. else if (psel == 0x08)
  335. bit = pmc - 1;
  336. else if (psel == 0x10)
  337. bit = 4 - pmc;
  338. else if (psel == 0x1b && (pmc == 1 || pmc == 3))
  339. bit = 4;
  340. } else if ((psel & 0x58) == 0x40)
  341. bit = psel & 7;
  342. if (!(event & PM_BUSEVENT_MSK))
  343. return 0;
  344. byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
  345. unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
  346. if (unit == PM_LSU0) {
  347. /* byte 1 bits 0-7, byte 2 bits 0,2-4,6 */
  348. mask = 0x5dff00;
  349. } else if (unit == PM_LSU1 && byte >= 4) {
  350. byte -= 4;
  351. /* byte 4 bits 1,3,5,7, byte 5 bits 6-7, byte 7 bits 0-4,6 */
  352. mask = 0x5f00c0aa;
  353. } else
  354. return 0;
  355. return (mask >> (byte * 8 + bit)) & 1;
  356. }
  357. static int power5_compute_mmcr(u64 event[], int n_ev,
  358. unsigned int hwc[], struct mmcr_regs *mmcr,
  359. struct perf_event *pevents[],
  360. u32 flags __maybe_unused)
  361. {
  362. unsigned long mmcr1 = 0;
  363. unsigned long mmcra = MMCRA_SDAR_DCACHE_MISS | MMCRA_SDAR_ERAT_MISS;
  364. unsigned int pmc, unit, byte, psel;
  365. unsigned int ttm, grp;
  366. int i, isbus, bit, grsel;
  367. unsigned int pmc_inuse = 0;
  368. unsigned int pmc_grp_use[2];
  369. unsigned char busbyte[4];
  370. unsigned char unituse[16];
  371. int ttmuse;
  372. if (n_ev > 6)
  373. return -1;
  374. /* First pass to count resource use */
  375. pmc_grp_use[0] = pmc_grp_use[1] = 0;
  376. memset(busbyte, 0, sizeof(busbyte));
  377. memset(unituse, 0, sizeof(unituse));
  378. for (i = 0; i < n_ev; ++i) {
  379. pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
  380. if (pmc) {
  381. if (pmc > 6)
  382. return -1;
  383. if (pmc_inuse & (1 << (pmc - 1)))
  384. return -1;
  385. pmc_inuse |= 1 << (pmc - 1);
  386. /* count 1/2 vs 3/4 use */
  387. if (pmc <= 4)
  388. ++pmc_grp_use[(pmc - 1) >> 1];
  389. }
  390. if (event[i] & PM_BUSEVENT_MSK) {
  391. unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
  392. byte = (event[i] >> PM_BYTE_SH) & PM_BYTE_MSK;
  393. if (unit > PM_LASTUNIT)
  394. return -1;
  395. if (unit == PM_ISU0_ALT)
  396. unit = PM_ISU0;
  397. if (byte >= 4) {
  398. if (unit != PM_LSU1)
  399. return -1;
  400. ++unit;
  401. byte &= 3;
  402. }
  403. if (!pmc)
  404. ++pmc_grp_use[byte & 1];
  405. if (busbyte[byte] && busbyte[byte] != unit)
  406. return -1;
  407. busbyte[byte] = unit;
  408. unituse[unit] = 1;
  409. }
  410. }
  411. if (pmc_grp_use[0] > 2 || pmc_grp_use[1] > 2)
  412. return -1;
  413. /*
  414. * Assign resources and set multiplexer selects.
  415. *
  416. * PM_ISU0 can go either on TTM0 or TTM1, but that's the only
  417. * choice we have to deal with.
  418. */
  419. if (unituse[PM_ISU0] &
  420. (unituse[PM_FPU] | unituse[PM_IFU] | unituse[PM_ISU1])) {
  421. unituse[PM_ISU0_ALT] = 1; /* move ISU to TTM1 */
  422. unituse[PM_ISU0] = 0;
  423. }
  424. /* Set TTM[01]SEL fields. */
  425. ttmuse = 0;
  426. for (i = PM_FPU; i <= PM_ISU1; ++i) {
  427. if (!unituse[i])
  428. continue;
  429. if (ttmuse++)
  430. return -1;
  431. mmcr1 |= (unsigned long)i << MMCR1_TTM0SEL_SH;
  432. }
  433. ttmuse = 0;
  434. for (; i <= PM_GRS; ++i) {
  435. if (!unituse[i])
  436. continue;
  437. if (ttmuse++)
  438. return -1;
  439. mmcr1 |= (unsigned long)(i & 3) << MMCR1_TTM1SEL_SH;
  440. }
  441. if (ttmuse > 1)
  442. return -1;
  443. /* Set byte lane select fields, TTM[23]SEL and GRS_*SEL. */
  444. for (byte = 0; byte < 4; ++byte) {
  445. unit = busbyte[byte];
  446. if (!unit)
  447. continue;
  448. if (unit == PM_ISU0 && unituse[PM_ISU0_ALT]) {
  449. /* get ISU0 through TTM1 rather than TTM0 */
  450. unit = PM_ISU0_ALT;
  451. } else if (unit == PM_LSU1 + 1) {
  452. /* select lower word of LSU1 for this byte */
  453. mmcr1 |= 1ul << (MMCR1_TTM3SEL_SH + 3 - byte);
  454. }
  455. ttm = unit >> 2;
  456. mmcr1 |= (unsigned long)ttm
  457. << (MMCR1_TD_CP_DBG0SEL_SH - 2 * byte);
  458. }
  459. /* Second pass: assign PMCs, set PMCxSEL and PMCx_ADDER_SEL fields */
  460. for (i = 0; i < n_ev; ++i) {
  461. pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
  462. unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
  463. byte = (event[i] >> PM_BYTE_SH) & PM_BYTE_MSK;
  464. psel = event[i] & PM_PMCSEL_MSK;
  465. isbus = event[i] & PM_BUSEVENT_MSK;
  466. if (!pmc) {
  467. /* Bus event or any-PMC direct event */
  468. for (pmc = 0; pmc < 4; ++pmc) {
  469. if (pmc_inuse & (1 << pmc))
  470. continue;
  471. grp = (pmc >> 1) & 1;
  472. if (isbus) {
  473. if (grp == (byte & 1))
  474. break;
  475. } else if (pmc_grp_use[grp] < 2) {
  476. ++pmc_grp_use[grp];
  477. break;
  478. }
  479. }
  480. pmc_inuse |= 1 << pmc;
  481. } else if (pmc <= 4) {
  482. /* Direct event */
  483. --pmc;
  484. if ((psel == 8 || psel == 0x10) && isbus && (byte & 2))
  485. /* add events on higher-numbered bus */
  486. mmcr1 |= 1ul << (MMCR1_PMC1_ADDER_SEL_SH - pmc);
  487. } else {
  488. /* Instructions or run cycles on PMC5/6 */
  489. --pmc;
  490. }
  491. if (isbus && unit == PM_GRS) {
  492. bit = psel & 7;
  493. grsel = (event[i] >> PM_GRS_SH) & PM_GRS_MSK;
  494. mmcr1 |= (unsigned long)grsel << grsel_shift[bit];
  495. }
  496. if (power5_marked_instr_event(event[i]))
  497. mmcra |= MMCRA_SAMPLE_ENABLE;
  498. if (pmc <= 3)
  499. mmcr1 |= psel << MMCR1_PMCSEL_SH(pmc);
  500. hwc[i] = pmc;
  501. }
  502. /* Return MMCRx values */
  503. mmcr->mmcr0 = 0;
  504. if (pmc_inuse & 1)
  505. mmcr->mmcr0 = MMCR0_PMC1CE;
  506. if (pmc_inuse & 0x3e)
  507. mmcr->mmcr0 |= MMCR0_PMCjCE;
  508. mmcr->mmcr1 = mmcr1;
  509. mmcr->mmcra = mmcra;
  510. return 0;
  511. }
  512. static void power5_disable_pmc(unsigned int pmc, struct mmcr_regs *mmcr)
  513. {
  514. if (pmc <= 3)
  515. mmcr->mmcr1 &= ~(0x7fUL << MMCR1_PMCSEL_SH(pmc));
  516. }
  517. static int power5_generic_events[] = {
  518. [PERF_COUNT_HW_CPU_CYCLES] = 0xf,
  519. [PERF_COUNT_HW_INSTRUCTIONS] = 0x100009,
  520. [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4c1090, /* LD_REF_L1 */
  521. [PERF_COUNT_HW_CACHE_MISSES] = 0x3c1088, /* LD_MISS_L1 */
  522. [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x230e4, /* BR_ISSUED */
  523. [PERF_COUNT_HW_BRANCH_MISSES] = 0x230e5, /* BR_MPRED_CR */
  524. };
  525. #define C(x) PERF_COUNT_HW_CACHE_##x
  526. /*
  527. * Table of generalized cache-related events.
  528. * 0 means not supported, -1 means nonsensical, other values
  529. * are event codes.
  530. */
  531. static u64 power5_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
  532. [C(L1D)] = { /* RESULT_ACCESS RESULT_MISS */
  533. [C(OP_READ)] = { 0x4c1090, 0x3c1088 },
  534. [C(OP_WRITE)] = { 0x3c1090, 0xc10c3 },
  535. [C(OP_PREFETCH)] = { 0xc70e7, 0 },
  536. },
  537. [C(L1I)] = { /* RESULT_ACCESS RESULT_MISS */
  538. [C(OP_READ)] = { 0, 0 },
  539. [C(OP_WRITE)] = { -1, -1 },
  540. [C(OP_PREFETCH)] = { 0, 0 },
  541. },
  542. [C(LL)] = { /* RESULT_ACCESS RESULT_MISS */
  543. [C(OP_READ)] = { 0, 0x3c309b },
  544. [C(OP_WRITE)] = { 0, 0 },
  545. [C(OP_PREFETCH)] = { 0xc50c3, 0 },
  546. },
  547. [C(DTLB)] = { /* RESULT_ACCESS RESULT_MISS */
  548. [C(OP_READ)] = { 0x2c4090, 0x800c4 },
  549. [C(OP_WRITE)] = { -1, -1 },
  550. [C(OP_PREFETCH)] = { -1, -1 },
  551. },
  552. [C(ITLB)] = { /* RESULT_ACCESS RESULT_MISS */
  553. [C(OP_READ)] = { 0, 0x800c0 },
  554. [C(OP_WRITE)] = { -1, -1 },
  555. [C(OP_PREFETCH)] = { -1, -1 },
  556. },
  557. [C(BPU)] = { /* RESULT_ACCESS RESULT_MISS */
  558. [C(OP_READ)] = { 0x230e4, 0x230e5 },
  559. [C(OP_WRITE)] = { -1, -1 },
  560. [C(OP_PREFETCH)] = { -1, -1 },
  561. },
  562. [C(NODE)] = { /* RESULT_ACCESS RESULT_MISS */
  563. [C(OP_READ)] = { -1, -1 },
  564. [C(OP_WRITE)] = { -1, -1 },
  565. [C(OP_PREFETCH)] = { -1, -1 },
  566. },
  567. };
  568. static struct power_pmu power5_pmu = {
  569. .name = "POWER5",
  570. .n_counter = 6,
  571. .max_alternatives = MAX_ALT,
  572. .add_fields = 0x7000090000555ul,
  573. .test_adder = 0x3000490000000ul,
  574. .compute_mmcr = power5_compute_mmcr,
  575. .get_constraint = power5_get_constraint,
  576. .get_alternatives = power5_get_alternatives,
  577. .disable_pmc = power5_disable_pmc,
  578. .n_generic = ARRAY_SIZE(power5_generic_events),
  579. .generic_events = power5_generic_events,
  580. .cache_events = &power5_cache_events,
  581. .flags = PPMU_HAS_SSLOT,
  582. };
  583. int __init init_power5_pmu(void)
  584. {
  585. unsigned int pvr = mfspr(SPRN_PVR);
  586. if (PVR_VER(pvr) != PVR_POWER5)
  587. return -ENODEV;
  588. return register_power_pmu(&power5_pmu);
  589. }