1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Driver for FPGA Management Engine (FME) Global Performance Reporting 4 * 5 * Copyright 2019 Intel Corporation, Inc. 6 * 7 * Authors: 8 * Kang Luwei <luwei.kang@intel.com> 9 * Xiao Guangrong <guangrong.xiao@linux.intel.com> 10 * Wu Hao <hao.wu@intel.com> 11 * Xu Yilun <yilun.xu@intel.com> 12 * Joseph Grecco <joe.grecco@intel.com> 13 * Enno Luebbers <enno.luebbers@intel.com> 14 * Tim Whisonant <tim.whisonant@intel.com> 15 * Ananda Ravuri <ananda.ravuri@intel.com> 16 * Mitchel, Henry <henry.mitchel@intel.com> 17 */ 18 19 #include <linux/perf_event.h> 20 #include "dfl.h" 21 #include "dfl-fme.h" 22 23 /* 24 * Performance Counter Registers for Cache. 25 * 26 * Cache Events are listed below as CACHE_EVNT_*. 27 */ 28 #define CACHE_CTRL 0x8 29 #define CACHE_RESET_CNTR BIT_ULL(0) 30 #define CACHE_FREEZE_CNTR BIT_ULL(8) 31 #define CACHE_CTRL_EVNT GENMASK_ULL(19, 16) 32 #define CACHE_EVNT_RD_HIT 0x0 33 #define CACHE_EVNT_WR_HIT 0x1 34 #define CACHE_EVNT_RD_MISS 0x2 35 #define CACHE_EVNT_WR_MISS 0x3 36 #define CACHE_EVNT_RSVD 0x4 37 #define CACHE_EVNT_HOLD_REQ 0x5 38 #define CACHE_EVNT_DATA_WR_PORT_CONTEN 0x6 39 #define CACHE_EVNT_TAG_WR_PORT_CONTEN 0x7 40 #define CACHE_EVNT_TX_REQ_STALL 0x8 41 #define CACHE_EVNT_RX_REQ_STALL 0x9 42 #define CACHE_EVNT_EVICTIONS 0xa 43 #define CACHE_EVNT_MAX CACHE_EVNT_EVICTIONS 44 #define CACHE_CHANNEL_SEL BIT_ULL(20) 45 #define CACHE_CHANNEL_RD 0 46 #define CACHE_CHANNEL_WR 1 47 #define CACHE_CNTR0 0x10 48 #define CACHE_CNTR1 0x18 49 #define CACHE_CNTR_EVNT_CNTR GENMASK_ULL(47, 0) 50 #define CACHE_CNTR_EVNT GENMASK_ULL(63, 60) 51 52 /* 53 * Performance Counter Registers for Fabric. 54 * 55 * Fabric Events are listed below as FAB_EVNT_* 56 */ 57 #define FAB_CTRL 0x20 58 #define FAB_RESET_CNTR BIT_ULL(0) 59 #define FAB_FREEZE_CNTR BIT_ULL(8) 60 #define FAB_CTRL_EVNT GENMASK_ULL(19, 16) 61 #define FAB_EVNT_PCIE0_RD 0x0 62 #define FAB_EVNT_PCIE0_WR 0x1 63 #define FAB_EVNT_PCIE1_RD 0x2 64 #define FAB_EVNT_PCIE1_WR 0x3 65 #define FAB_EVNT_UPI_RD 0x4 66 #define FAB_EVNT_UPI_WR 0x5 67 #define FAB_EVNT_MMIO_RD 0x6 68 #define FAB_EVNT_MMIO_WR 0x7 69 #define FAB_EVNT_MAX FAB_EVNT_MMIO_WR 70 #define FAB_PORT_ID GENMASK_ULL(21, 20) 71 #define FAB_PORT_FILTER BIT_ULL(23) 72 #define FAB_PORT_FILTER_DISABLE 0 73 #define FAB_PORT_FILTER_ENABLE 1 74 #define FAB_CNTR 0x28 75 #define FAB_CNTR_EVNT_CNTR GENMASK_ULL(59, 0) 76 #define FAB_CNTR_EVNT GENMASK_ULL(63, 60) 77 78 /* 79 * Performance Counter Registers for Clock. 80 * 81 * Clock Counter can't be reset or frozen by SW. 82 */ 83 #define CLK_CNTR 0x30 84 #define BASIC_EVNT_CLK 0x0 85 #define BASIC_EVNT_MAX BASIC_EVNT_CLK 86 87 /* 88 * Performance Counter Registers for IOMMU / VT-D. 89 * 90 * VT-D Events are listed below as VTD_EVNT_* and VTD_SIP_EVNT_* 91 */ 92 #define VTD_CTRL 0x38 93 #define VTD_RESET_CNTR BIT_ULL(0) 94 #define VTD_FREEZE_CNTR BIT_ULL(8) 95 #define VTD_CTRL_EVNT GENMASK_ULL(19, 16) 96 #define VTD_EVNT_AFU_MEM_RD_TRANS 0x0 97 #define VTD_EVNT_AFU_MEM_WR_TRANS 0x1 98 #define VTD_EVNT_AFU_DEVTLB_RD_HIT 0x2 99 #define VTD_EVNT_AFU_DEVTLB_WR_HIT 0x3 100 #define VTD_EVNT_DEVTLB_4K_FILL 0x4 101 #define VTD_EVNT_DEVTLB_2M_FILL 0x5 102 #define VTD_EVNT_DEVTLB_1G_FILL 0x6 103 #define VTD_EVNT_MAX VTD_EVNT_DEVTLB_1G_FILL 104 #define VTD_CNTR 0x40 105 #define VTD_CNTR_EVNT_CNTR GENMASK_ULL(47, 0) 106 #define VTD_CNTR_EVNT GENMASK_ULL(63, 60) 107 108 #define VTD_SIP_CTRL 0x48 109 #define VTD_SIP_RESET_CNTR BIT_ULL(0) 110 #define VTD_SIP_FREEZE_CNTR BIT_ULL(8) 111 #define VTD_SIP_CTRL_EVNT GENMASK_ULL(19, 16) 112 #define VTD_SIP_EVNT_IOTLB_4K_HIT 0x0 113 #define VTD_SIP_EVNT_IOTLB_2M_HIT 0x1 114 #define VTD_SIP_EVNT_IOTLB_1G_HIT 0x2 115 #define VTD_SIP_EVNT_SLPWC_L3_HIT 0x3 116 #define VTD_SIP_EVNT_SLPWC_L4_HIT 0x4 117 #define VTD_SIP_EVNT_RCC_HIT 0x5 118 #define VTD_SIP_EVNT_IOTLB_4K_MISS 0x6 119 #define VTD_SIP_EVNT_IOTLB_2M_MISS 0x7 120 #define VTD_SIP_EVNT_IOTLB_1G_MISS 0x8 121 #define VTD_SIP_EVNT_SLPWC_L3_MISS 0x9 122 #define VTD_SIP_EVNT_SLPWC_L4_MISS 0xa 123 #define VTD_SIP_EVNT_RCC_MISS 0xb 124 #define VTD_SIP_EVNT_MAX VTD_SIP_EVNT_SLPWC_L4_MISS 125 #define VTD_SIP_CNTR 0X50 126 #define VTD_SIP_CNTR_EVNT_CNTR GENMASK_ULL(47, 0) 127 #define VTD_SIP_CNTR_EVNT GENMASK_ULL(63, 60) 128 129 #define PERF_TIMEOUT 30 130 131 #define PERF_MAX_PORT_NUM 1U 132 133 /** 134 * struct fme_perf_priv - priv data structure for fme perf driver 135 * 136 * @dev: parent device. 137 * @ioaddr: mapped base address of mmio region. 138 * @pmu: pmu data structure for fme perf counters. 139 * @id: id of this fme performance report private feature. 140 * @fab_users: current user number on fabric counters. 141 * @fab_port_id: used to indicate current working mode of fabric counters. 142 * @fab_lock: lock to protect fabric counters working mode. 143 * @cpu: active CPU to which the PMU is bound for accesses. 144 * @cpuhp_node: node for CPU hotplug notifier link. 145 * @cpuhp_state: state for CPU hotplug notification; 146 */ 147 struct fme_perf_priv { 148 struct device *dev; 149 void __iomem *ioaddr; 150 struct pmu pmu; 151 u16 id; 152 153 u32 fab_users; 154 u32 fab_port_id; 155 spinlock_t fab_lock; 156 157 unsigned int cpu; 158 struct hlist_node node; 159 enum cpuhp_state cpuhp_state; 160 }; 161 162 /** 163 * struct fme_perf_event_ops - callbacks for fme perf events 164 * 165 * @event_init: callback invoked during event init. 166 * @event_destroy: callback invoked during event destroy. 167 * @read_counter: callback to read hardware counters. 168 */ 169 struct fme_perf_event_ops { 170 int (*event_init)(struct fme_perf_priv *priv, u32 event, u32 portid); 171 void (*event_destroy)(struct fme_perf_priv *priv, u32 event, 172 u32 portid); 173 u64 (*read_counter)(struct fme_perf_priv *priv, u32 event, u32 portid); 174 }; 175 176 #define to_fme_perf_priv(_pmu) container_of(_pmu, struct fme_perf_priv, pmu) 177 178 static ssize_t cpumask_show(struct device *dev, 179 struct device_attribute *attr, char *buf) 180 { 181 struct pmu *pmu = dev_get_drvdata(dev); 182 struct fme_perf_priv *priv; 183 184 priv = to_fme_perf_priv(pmu); 185 186 return cpumap_print_to_pagebuf(true, buf, cpumask_of(priv->cpu)); 187 } 188 static DEVICE_ATTR_RO(cpumask); 189 190 static struct attribute *fme_perf_cpumask_attrs[] = { 191 &dev_attr_cpumask.attr, 192 NULL, 193 }; 194 195 static const struct attribute_group fme_perf_cpumask_group = { 196 .attrs = fme_perf_cpumask_attrs, 197 }; 198 199 #define FME_EVENT_MASK GENMASK_ULL(11, 0) 200 #define FME_EVENT_SHIFT 0 201 #define FME_EVTYPE_MASK GENMASK_ULL(15, 12) 202 #define FME_EVTYPE_SHIFT 12 203 #define FME_EVTYPE_BASIC 0 204 #define FME_EVTYPE_CACHE 1 205 #define FME_EVTYPE_FABRIC 2 206 #define FME_EVTYPE_VTD 3 207 #define FME_EVTYPE_VTD_SIP 4 208 #define FME_EVTYPE_MAX FME_EVTYPE_VTD_SIP 209 #define FME_PORTID_MASK GENMASK_ULL(23, 16) 210 #define FME_PORTID_SHIFT 16 211 #define FME_PORTID_ROOT (0xffU) 212 213 #define get_event(_config) FIELD_GET(FME_EVENT_MASK, _config) 214 #define get_evtype(_config) FIELD_GET(FME_EVTYPE_MASK, _config) 215 #define get_portid(_config) FIELD_GET(FME_PORTID_MASK, _config) 216 217 PMU_FORMAT_ATTR(event, "config:0-11"); 218 PMU_FORMAT_ATTR(evtype, "config:12-15"); 219 PMU_FORMAT_ATTR(portid, "config:16-23"); 220 221 static struct attribute *fme_perf_format_attrs[] = { 222 &format_attr_event.attr, 223 &format_attr_evtype.attr, 224 &format_attr_portid.attr, 225 NULL, 226 }; 227 228 static const struct attribute_group fme_perf_format_group = { 229 .name = "format", 230 .attrs = fme_perf_format_attrs, 231 }; 232 233 /* 234 * There are no default events, but we need to create 235 * "events" group (with empty attrs) before updating 236 * it with detected events (using pmu->attr_update). 237 */ 238 static struct attribute *fme_perf_events_attrs_empty[] = { 239 NULL, 240 }; 241 242 static const struct attribute_group fme_perf_events_group = { 243 .name = "events", 244 .attrs = fme_perf_events_attrs_empty, 245 }; 246 247 static const struct attribute_group *fme_perf_groups[] = { 248 &fme_perf_format_group, 249 &fme_perf_cpumask_group, 250 &fme_perf_events_group, 251 NULL, 252 }; 253 254 static bool is_portid_root(u32 portid) 255 { 256 return portid == FME_PORTID_ROOT; 257 } 258 259 static bool is_portid_port(u32 portid) 260 { 261 return portid < PERF_MAX_PORT_NUM; 262 } 263 264 static bool is_portid_root_or_port(u32 portid) 265 { 266 return is_portid_root(portid) || is_portid_port(portid); 267 } 268 269 static u64 fme_read_perf_cntr_reg(void __iomem *addr) 270 { 271 u32 low; 272 u64 v; 273 274 /* 275 * For 64bit counter registers, the counter may increases and carries 276 * out of bit [31] between 2 32bit reads. So add extra reads to help 277 * to prevent this issue. This only happens in platforms which don't 278 * support 64bit read - readq is split into 2 readl. 279 */ 280 do { 281 v = readq(addr); 282 low = readl(addr); 283 } while (((u32)v) > low); 284 285 return v; 286 } 287 288 static int basic_event_init(struct fme_perf_priv *priv, u32 event, u32 portid) 289 { 290 if (event <= BASIC_EVNT_MAX && is_portid_root(portid)) 291 return 0; 292 293 return -EINVAL; 294 } 295 296 static u64 basic_read_event_counter(struct fme_perf_priv *priv, 297 u32 event, u32 portid) 298 { 299 void __iomem *base = priv->ioaddr; 300 301 return fme_read_perf_cntr_reg(base + CLK_CNTR); 302 } 303 304 static int cache_event_init(struct fme_perf_priv *priv, u32 event, u32 portid) 305 { 306 if (priv->id == FME_FEATURE_ID_GLOBAL_IPERF && 307 event <= CACHE_EVNT_MAX && is_portid_root(portid)) 308 return 0; 309 310 return -EINVAL; 311 } 312 313 static u64 cache_read_event_counter(struct fme_perf_priv *priv, 314 u32 event, u32 portid) 315 { 316 void __iomem *base = priv->ioaddr; 317 u64 v, count; 318 u8 channel; 319 320 if (event == CACHE_EVNT_WR_HIT || event == CACHE_EVNT_WR_MISS || 321 event == CACHE_EVNT_DATA_WR_PORT_CONTEN || 322 event == CACHE_EVNT_TAG_WR_PORT_CONTEN) 323 channel = CACHE_CHANNEL_WR; 324 else 325 channel = CACHE_CHANNEL_RD; 326 327 /* set channel access type and cache event code. */ 328 v = readq(base + CACHE_CTRL); 329 v &= ~(CACHE_CHANNEL_SEL | CACHE_CTRL_EVNT); 330 v |= FIELD_PREP(CACHE_CHANNEL_SEL, channel); 331 v |= FIELD_PREP(CACHE_CTRL_EVNT, event); 332 writeq(v, base + CACHE_CTRL); 333 334 if (readq_poll_timeout_atomic(base + CACHE_CNTR0, v, 335 FIELD_GET(CACHE_CNTR_EVNT, v) == event, 336 1, PERF_TIMEOUT)) { 337 dev_err(priv->dev, "timeout, unmatched cache event code in counter register.\n"); 338 return 0; 339 } 340 341 v = fme_read_perf_cntr_reg(base + CACHE_CNTR0); 342 count = FIELD_GET(CACHE_CNTR_EVNT_CNTR, v); 343 v = fme_read_perf_cntr_reg(base + CACHE_CNTR1); 344 count += FIELD_GET(CACHE_CNTR_EVNT_CNTR, v); 345 346 return count; 347 } 348 349 static bool is_fabric_event_supported(struct fme_perf_priv *priv, u32 event, 350 u32 portid) 351 { 352 if (event > FAB_EVNT_MAX || !is_portid_root_or_port(portid)) 353 return false; 354 355 if (priv->id == FME_FEATURE_ID_GLOBAL_DPERF && 356 (event == FAB_EVNT_PCIE1_RD || event == FAB_EVNT_UPI_RD || 357 event == FAB_EVNT_PCIE1_WR || event == FAB_EVNT_UPI_WR)) 358 return false; 359 360 return true; 361 } 362 363 static int fabric_event_init(struct fme_perf_priv *priv, u32 event, u32 portid) 364 { 365 void __iomem *base = priv->ioaddr; 366 int ret = 0; 367 u64 v; 368 369 if (!is_fabric_event_supported(priv, event, portid)) 370 return -EINVAL; 371 372 /* 373 * as fabric counter set only can be in either overall or port mode. 374 * In overall mode, it counts overall data for FPGA, and in port mode, 375 * it is configured to monitor on one individual port. 376 * 377 * so every time, a new event is initialized, driver checks 378 * current working mode and if someone is using this counter set. 379 */ 380 spin_lock(&priv->fab_lock); 381 if (priv->fab_users && priv->fab_port_id != portid) { 382 dev_dbg(priv->dev, "conflict fabric event monitoring mode.\n"); 383 ret = -EOPNOTSUPP; 384 goto exit; 385 } 386 387 priv->fab_users++; 388 389 /* 390 * skip if current working mode matches, otherwise change the working 391 * mode per input port_id, to monitor overall data or another port. 392 */ 393 if (priv->fab_port_id == portid) 394 goto exit; 395 396 priv->fab_port_id = portid; 397 398 v = readq(base + FAB_CTRL); 399 v &= ~(FAB_PORT_FILTER | FAB_PORT_ID); 400 401 if (is_portid_root(portid)) { 402 v |= FIELD_PREP(FAB_PORT_FILTER, FAB_PORT_FILTER_DISABLE); 403 } else { 404 v |= FIELD_PREP(FAB_PORT_FILTER, FAB_PORT_FILTER_ENABLE); 405 v |= FIELD_PREP(FAB_PORT_ID, portid); 406 } 407 writeq(v, base + FAB_CTRL); 408 409 exit: 410 spin_unlock(&priv->fab_lock); 411 return ret; 412 } 413 414 static void fabric_event_destroy(struct fme_perf_priv *priv, u32 event, 415 u32 portid) 416 { 417 spin_lock(&priv->fab_lock); 418 priv->fab_users--; 419 spin_unlock(&priv->fab_lock); 420 } 421 422 static u64 fabric_read_event_counter(struct fme_perf_priv *priv, u32 event, 423 u32 portid) 424 { 425 void __iomem *base = priv->ioaddr; 426 u64 v; 427 428 v = readq(base + FAB_CTRL); 429 v &= ~FAB_CTRL_EVNT; 430 v |= FIELD_PREP(FAB_CTRL_EVNT, event); 431 writeq(v, base + FAB_CTRL); 432 433 if (readq_poll_timeout_atomic(base + FAB_CNTR, v, 434 FIELD_GET(FAB_CNTR_EVNT, v) == event, 435 1, PERF_TIMEOUT)) { 436 dev_err(priv->dev, "timeout, unmatched fab event code in counter register.\n"); 437 return 0; 438 } 439 440 v = fme_read_perf_cntr_reg(base + FAB_CNTR); 441 return FIELD_GET(FAB_CNTR_EVNT_CNTR, v); 442 } 443 444 static int vtd_event_init(struct fme_perf_priv *priv, u32 event, u32 portid) 445 { 446 if (priv->id == FME_FEATURE_ID_GLOBAL_IPERF && 447 event <= VTD_EVNT_MAX && is_portid_port(portid)) 448 return 0; 449 450 return -EINVAL; 451 } 452 453 static u64 vtd_read_event_counter(struct fme_perf_priv *priv, u32 event, 454 u32 portid) 455 { 456 void __iomem *base = priv->ioaddr; 457 u64 v; 458 459 event += (portid * (VTD_EVNT_MAX + 1)); 460 461 v = readq(base + VTD_CTRL); 462 v &= ~VTD_CTRL_EVNT; 463 v |= FIELD_PREP(VTD_CTRL_EVNT, event); 464 writeq(v, base + VTD_CTRL); 465 466 if (readq_poll_timeout_atomic(base + VTD_CNTR, v, 467 FIELD_GET(VTD_CNTR_EVNT, v) == event, 468 1, PERF_TIMEOUT)) { 469 dev_err(priv->dev, "timeout, unmatched vtd event code in counter register.\n"); 470 return 0; 471 } 472 473 v = fme_read_perf_cntr_reg(base + VTD_CNTR); 474 return FIELD_GET(VTD_CNTR_EVNT_CNTR, v); 475 } 476 477 static int vtd_sip_event_init(struct fme_perf_priv *priv, u32 event, u32 portid) 478 { 479 if (priv->id == FME_FEATURE_ID_GLOBAL_IPERF && 480 event <= VTD_SIP_EVNT_MAX && is_portid_root(portid)) 481 return 0; 482 483 return -EINVAL; 484 } 485 486 static u64 vtd_sip_read_event_counter(struct fme_perf_priv *priv, u32 event, 487 u32 portid) 488 { 489 void __iomem *base = priv->ioaddr; 490 u64 v; 491 492 v = readq(base + VTD_SIP_CTRL); 493 v &= ~VTD_SIP_CTRL_EVNT; 494 v |= FIELD_PREP(VTD_SIP_CTRL_EVNT, event); 495 writeq(v, base + VTD_SIP_CTRL); 496 497 if (readq_poll_timeout_atomic(base + VTD_SIP_CNTR, v, 498 FIELD_GET(VTD_SIP_CNTR_EVNT, v) == event, 499 1, PERF_TIMEOUT)) { 500 dev_err(priv->dev, "timeout, unmatched vtd sip event code in counter register\n"); 501 return 0; 502 } 503 504 v = fme_read_perf_cntr_reg(base + VTD_SIP_CNTR); 505 return FIELD_GET(VTD_SIP_CNTR_EVNT_CNTR, v); 506 } 507 508 static struct fme_perf_event_ops fme_perf_event_ops[] = { 509 [FME_EVTYPE_BASIC] = {.event_init = basic_event_init, 510 .read_counter = basic_read_event_counter,}, 511 [FME_EVTYPE_CACHE] = {.event_init = cache_event_init, 512 .read_counter = cache_read_event_counter,}, 513 [FME_EVTYPE_FABRIC] = {.event_init = fabric_event_init, 514 .event_destroy = fabric_event_destroy, 515 .read_counter = fabric_read_event_counter,}, 516 [FME_EVTYPE_VTD] = {.event_init = vtd_event_init, 517 .read_counter = vtd_read_event_counter,}, 518 [FME_EVTYPE_VTD_SIP] = {.event_init = vtd_sip_event_init, 519 .read_counter = vtd_sip_read_event_counter,}, 520 }; 521 522 static ssize_t fme_perf_event_show(struct device *dev, 523 struct device_attribute *attr, char *buf) 524 { 525 struct dev_ext_attribute *eattr; 526 unsigned long config; 527 char *ptr = buf; 528 529 eattr = container_of(attr, struct dev_ext_attribute, attr); 530 config = (unsigned long)eattr->var; 531 532 ptr += sprintf(ptr, "event=0x%02x", (unsigned int)get_event(config)); 533 ptr += sprintf(ptr, ",evtype=0x%02x", (unsigned int)get_evtype(config)); 534 535 if (is_portid_root(get_portid(config))) 536 ptr += sprintf(ptr, ",portid=0x%02x\n", FME_PORTID_ROOT); 537 else 538 ptr += sprintf(ptr, ",portid=?\n"); 539 540 return (ssize_t)(ptr - buf); 541 } 542 543 #define FME_EVENT_ATTR(_name) \ 544 __ATTR(_name, 0444, fme_perf_event_show, NULL) 545 546 #define FME_PORT_EVENT_CONFIG(_event, _type) \ 547 (void *)((((_event) << FME_EVENT_SHIFT) & FME_EVENT_MASK) | \ 548 (((_type) << FME_EVTYPE_SHIFT) & FME_EVTYPE_MASK)) 549 550 #define FME_EVENT_CONFIG(_event, _type) \ 551 (void *)((((_event) << FME_EVENT_SHIFT) & FME_EVENT_MASK) | \ 552 (((_type) << FME_EVTYPE_SHIFT) & FME_EVTYPE_MASK) | \ 553 (FME_PORTID_ROOT << FME_PORTID_SHIFT)) 554 555 /* FME Perf Basic Events */ 556 #define FME_EVENT_BASIC(_name, _event) \ 557 static struct dev_ext_attribute fme_perf_event_##_name = { \ 558 .attr = FME_EVENT_ATTR(_name), \ 559 .var = FME_EVENT_CONFIG(_event, FME_EVTYPE_BASIC), \ 560 } 561 562 FME_EVENT_BASIC(clock, BASIC_EVNT_CLK); 563 564 static struct attribute *fme_perf_basic_events_attrs[] = { 565 &fme_perf_event_clock.attr.attr, 566 NULL, 567 }; 568 569 static const struct attribute_group fme_perf_basic_events_group = { 570 .name = "events", 571 .attrs = fme_perf_basic_events_attrs, 572 }; 573 574 /* FME Perf Cache Events */ 575 #define FME_EVENT_CACHE(_name, _event) \ 576 static struct dev_ext_attribute fme_perf_event_cache_##_name = { \ 577 .attr = FME_EVENT_ATTR(cache_##_name), \ 578 .var = FME_EVENT_CONFIG(_event, FME_EVTYPE_CACHE), \ 579 } 580 581 FME_EVENT_CACHE(read_hit, CACHE_EVNT_RD_HIT); 582 FME_EVENT_CACHE(read_miss, CACHE_EVNT_RD_MISS); 583 FME_EVENT_CACHE(write_hit, CACHE_EVNT_WR_HIT); 584 FME_EVENT_CACHE(write_miss, CACHE_EVNT_WR_MISS); 585 FME_EVENT_CACHE(hold_request, CACHE_EVNT_HOLD_REQ); 586 FME_EVENT_CACHE(tx_req_stall, CACHE_EVNT_TX_REQ_STALL); 587 FME_EVENT_CACHE(rx_req_stall, CACHE_EVNT_RX_REQ_STALL); 588 FME_EVENT_CACHE(eviction, CACHE_EVNT_EVICTIONS); 589 FME_EVENT_CACHE(data_write_port_contention, CACHE_EVNT_DATA_WR_PORT_CONTEN); 590 FME_EVENT_CACHE(tag_write_port_contention, CACHE_EVNT_TAG_WR_PORT_CONTEN); 591 592 static struct attribute *fme_perf_cache_events_attrs[] = { 593 &fme_perf_event_cache_read_hit.attr.attr, 594 &fme_perf_event_cache_read_miss.attr.attr, 595 &fme_perf_event_cache_write_hit.attr.attr, 596 &fme_perf_event_cache_write_miss.attr.attr, 597 &fme_perf_event_cache_hold_request.attr.attr, 598 &fme_perf_event_cache_tx_req_stall.attr.attr, 599 &fme_perf_event_cache_rx_req_stall.attr.attr, 600 &fme_perf_event_cache_eviction.attr.attr, 601 &fme_perf_event_cache_data_write_port_contention.attr.attr, 602 &fme_perf_event_cache_tag_write_port_contention.attr.attr, 603 NULL, 604 }; 605 606 static umode_t fme_perf_events_visible(struct kobject *kobj, 607 struct attribute *attr, int n) 608 { 609 struct pmu *pmu = dev_get_drvdata(kobj_to_dev(kobj)); 610 struct fme_perf_priv *priv = to_fme_perf_priv(pmu); 611 612 return (priv->id == FME_FEATURE_ID_GLOBAL_IPERF) ? attr->mode : 0; 613 } 614 615 static const struct attribute_group fme_perf_cache_events_group = { 616 .name = "events", 617 .attrs = fme_perf_cache_events_attrs, 618 .is_visible = fme_perf_events_visible, 619 }; 620 621 /* FME Perf Fabric Events */ 622 #define FME_EVENT_FABRIC(_name, _event) \ 623 static struct dev_ext_attribute fme_perf_event_fab_##_name = { \ 624 .attr = FME_EVENT_ATTR(fab_##_name), \ 625 .var = FME_EVENT_CONFIG(_event, FME_EVTYPE_FABRIC), \ 626 } 627 628 #define FME_EVENT_FABRIC_PORT(_name, _event) \ 629 static struct dev_ext_attribute fme_perf_event_fab_port_##_name = { \ 630 .attr = FME_EVENT_ATTR(fab_port_##_name), \ 631 .var = FME_PORT_EVENT_CONFIG(_event, FME_EVTYPE_FABRIC), \ 632 } 633 634 FME_EVENT_FABRIC(pcie0_read, FAB_EVNT_PCIE0_RD); 635 FME_EVENT_FABRIC(pcie0_write, FAB_EVNT_PCIE0_WR); 636 FME_EVENT_FABRIC(pcie1_read, FAB_EVNT_PCIE1_RD); 637 FME_EVENT_FABRIC(pcie1_write, FAB_EVNT_PCIE1_WR); 638 FME_EVENT_FABRIC(upi_read, FAB_EVNT_UPI_RD); 639 FME_EVENT_FABRIC(upi_write, FAB_EVNT_UPI_WR); 640 FME_EVENT_FABRIC(mmio_read, FAB_EVNT_MMIO_RD); 641 FME_EVENT_FABRIC(mmio_write, FAB_EVNT_MMIO_WR); 642 643 FME_EVENT_FABRIC_PORT(pcie0_read, FAB_EVNT_PCIE0_RD); 644 FME_EVENT_FABRIC_PORT(pcie0_write, FAB_EVNT_PCIE0_WR); 645 FME_EVENT_FABRIC_PORT(pcie1_read, FAB_EVNT_PCIE1_RD); 646 FME_EVENT_FABRIC_PORT(pcie1_write, FAB_EVNT_PCIE1_WR); 647 FME_EVENT_FABRIC_PORT(upi_read, FAB_EVNT_UPI_RD); 648 FME_EVENT_FABRIC_PORT(upi_write, FAB_EVNT_UPI_WR); 649 FME_EVENT_FABRIC_PORT(mmio_read, FAB_EVNT_MMIO_RD); 650 FME_EVENT_FABRIC_PORT(mmio_write, FAB_EVNT_MMIO_WR); 651 652 static struct attribute *fme_perf_fabric_events_attrs[] = { 653 &fme_perf_event_fab_pcie0_read.attr.attr, 654 &fme_perf_event_fab_pcie0_write.attr.attr, 655 &fme_perf_event_fab_pcie1_read.attr.attr, 656 &fme_perf_event_fab_pcie1_write.attr.attr, 657 &fme_perf_event_fab_upi_read.attr.attr, 658 &fme_perf_event_fab_upi_write.attr.attr, 659 &fme_perf_event_fab_mmio_read.attr.attr, 660 &fme_perf_event_fab_mmio_write.attr.attr, 661 &fme_perf_event_fab_port_pcie0_read.attr.attr, 662 &fme_perf_event_fab_port_pcie0_write.attr.attr, 663 &fme_perf_event_fab_port_pcie1_read.attr.attr, 664 &fme_perf_event_fab_port_pcie1_write.attr.attr, 665 &fme_perf_event_fab_port_upi_read.attr.attr, 666 &fme_perf_event_fab_port_upi_write.attr.attr, 667 &fme_perf_event_fab_port_mmio_read.attr.attr, 668 &fme_perf_event_fab_port_mmio_write.attr.attr, 669 NULL, 670 }; 671 672 static umode_t fme_perf_fabric_events_visible(struct kobject *kobj, 673 struct attribute *attr, int n) 674 { 675 struct pmu *pmu = dev_get_drvdata(kobj_to_dev(kobj)); 676 struct fme_perf_priv *priv = to_fme_perf_priv(pmu); 677 struct dev_ext_attribute *eattr; 678 unsigned long var; 679 680 eattr = container_of(attr, struct dev_ext_attribute, attr.attr); 681 var = (unsigned long)eattr->var; 682 683 if (is_fabric_event_supported(priv, get_event(var), get_portid(var))) 684 return attr->mode; 685 686 return 0; 687 } 688 689 static const struct attribute_group fme_perf_fabric_events_group = { 690 .name = "events", 691 .attrs = fme_perf_fabric_events_attrs, 692 .is_visible = fme_perf_fabric_events_visible, 693 }; 694 695 /* FME Perf VTD Events */ 696 #define FME_EVENT_VTD_PORT(_name, _event) \ 697 static struct dev_ext_attribute fme_perf_event_vtd_port_##_name = { \ 698 .attr = FME_EVENT_ATTR(vtd_port_##_name), \ 699 .var = FME_PORT_EVENT_CONFIG(_event, FME_EVTYPE_VTD), \ 700 } 701 702 FME_EVENT_VTD_PORT(read_transaction, VTD_EVNT_AFU_MEM_RD_TRANS); 703 FME_EVENT_VTD_PORT(write_transaction, VTD_EVNT_AFU_MEM_WR_TRANS); 704 FME_EVENT_VTD_PORT(devtlb_read_hit, VTD_EVNT_AFU_DEVTLB_RD_HIT); 705 FME_EVENT_VTD_PORT(devtlb_write_hit, VTD_EVNT_AFU_DEVTLB_WR_HIT); 706 FME_EVENT_VTD_PORT(devtlb_4k_fill, VTD_EVNT_DEVTLB_4K_FILL); 707 FME_EVENT_VTD_PORT(devtlb_2m_fill, VTD_EVNT_DEVTLB_2M_FILL); 708 FME_EVENT_VTD_PORT(devtlb_1g_fill, VTD_EVNT_DEVTLB_1G_FILL); 709 710 static struct attribute *fme_perf_vtd_events_attrs[] = { 711 &fme_perf_event_vtd_port_read_transaction.attr.attr, 712 &fme_perf_event_vtd_port_write_transaction.attr.attr, 713 &fme_perf_event_vtd_port_devtlb_read_hit.attr.attr, 714 &fme_perf_event_vtd_port_devtlb_write_hit.attr.attr, 715 &fme_perf_event_vtd_port_devtlb_4k_fill.attr.attr, 716 &fme_perf_event_vtd_port_devtlb_2m_fill.attr.attr, 717 &fme_perf_event_vtd_port_devtlb_1g_fill.attr.attr, 718 NULL, 719 }; 720 721 static const struct attribute_group fme_perf_vtd_events_group = { 722 .name = "events", 723 .attrs = fme_perf_vtd_events_attrs, 724 .is_visible = fme_perf_events_visible, 725 }; 726 727 /* FME Perf VTD SIP Events */ 728 #define FME_EVENT_VTD_SIP(_name, _event) \ 729 static struct dev_ext_attribute fme_perf_event_vtd_sip_##_name = { \ 730 .attr = FME_EVENT_ATTR(vtd_sip_##_name), \ 731 .var = FME_EVENT_CONFIG(_event, FME_EVTYPE_VTD_SIP), \ 732 } 733 734 FME_EVENT_VTD_SIP(iotlb_4k_hit, VTD_SIP_EVNT_IOTLB_4K_HIT); 735 FME_EVENT_VTD_SIP(iotlb_2m_hit, VTD_SIP_EVNT_IOTLB_2M_HIT); 736 FME_EVENT_VTD_SIP(iotlb_1g_hit, VTD_SIP_EVNT_IOTLB_1G_HIT); 737 FME_EVENT_VTD_SIP(slpwc_l3_hit, VTD_SIP_EVNT_SLPWC_L3_HIT); 738 FME_EVENT_VTD_SIP(slpwc_l4_hit, VTD_SIP_EVNT_SLPWC_L4_HIT); 739 FME_EVENT_VTD_SIP(rcc_hit, VTD_SIP_EVNT_RCC_HIT); 740 FME_EVENT_VTD_SIP(iotlb_4k_miss, VTD_SIP_EVNT_IOTLB_4K_MISS); 741 FME_EVENT_VTD_SIP(iotlb_2m_miss, VTD_SIP_EVNT_IOTLB_2M_MISS); 742 FME_EVENT_VTD_SIP(iotlb_1g_miss, VTD_SIP_EVNT_IOTLB_1G_MISS); 743 FME_EVENT_VTD_SIP(slpwc_l3_miss, VTD_SIP_EVNT_SLPWC_L3_MISS); 744 FME_EVENT_VTD_SIP(slpwc_l4_miss, VTD_SIP_EVNT_SLPWC_L4_MISS); 745 FME_EVENT_VTD_SIP(rcc_miss, VTD_SIP_EVNT_RCC_MISS); 746 747 static struct attribute *fme_perf_vtd_sip_events_attrs[] = { 748 &fme_perf_event_vtd_sip_iotlb_4k_hit.attr.attr, 749 &fme_perf_event_vtd_sip_iotlb_2m_hit.attr.attr, 750 &fme_perf_event_vtd_sip_iotlb_1g_hit.attr.attr, 751 &fme_perf_event_vtd_sip_slpwc_l3_hit.attr.attr, 752 &fme_perf_event_vtd_sip_slpwc_l4_hit.attr.attr, 753 &fme_perf_event_vtd_sip_rcc_hit.attr.attr, 754 &fme_perf_event_vtd_sip_iotlb_4k_miss.attr.attr, 755 &fme_perf_event_vtd_sip_iotlb_2m_miss.attr.attr, 756 &fme_perf_event_vtd_sip_iotlb_1g_miss.attr.attr, 757 &fme_perf_event_vtd_sip_slpwc_l3_miss.attr.attr, 758 &fme_perf_event_vtd_sip_slpwc_l4_miss.attr.attr, 759 &fme_perf_event_vtd_sip_rcc_miss.attr.attr, 760 NULL, 761 }; 762 763 static const struct attribute_group fme_perf_vtd_sip_events_group = { 764 .name = "events", 765 .attrs = fme_perf_vtd_sip_events_attrs, 766 .is_visible = fme_perf_events_visible, 767 }; 768 769 static const struct attribute_group *fme_perf_events_groups[] = { 770 &fme_perf_basic_events_group, 771 &fme_perf_cache_events_group, 772 &fme_perf_fabric_events_group, 773 &fme_perf_vtd_events_group, 774 &fme_perf_vtd_sip_events_group, 775 NULL, 776 }; 777 778 static struct fme_perf_event_ops *get_event_ops(u32 evtype) 779 { 780 if (evtype > FME_EVTYPE_MAX) 781 return NULL; 782 783 return &fme_perf_event_ops[evtype]; 784 } 785 786 static void fme_perf_event_destroy(struct perf_event *event) 787 { 788 struct fme_perf_event_ops *ops = get_event_ops(event->hw.event_base); 789 struct fme_perf_priv *priv = to_fme_perf_priv(event->pmu); 790 791 if (ops->event_destroy) 792 ops->event_destroy(priv, event->hw.idx, event->hw.config_base); 793 } 794 795 static int fme_perf_event_init(struct perf_event *event) 796 { 797 struct fme_perf_priv *priv = to_fme_perf_priv(event->pmu); 798 struct hw_perf_event *hwc = &event->hw; 799 struct fme_perf_event_ops *ops; 800 u32 eventid, evtype, portid; 801 802 /* test the event attr type check for PMU enumeration */ 803 if (event->attr.type != event->pmu->type) 804 return -ENOENT; 805 806 /* 807 * fme counters are shared across all cores. 808 * Therefore, it does not support per-process mode. 809 * Also, it does not support event sampling mode. 810 */ 811 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK) 812 return -EINVAL; 813 814 if (event->cpu < 0) 815 return -EINVAL; 816 817 if (event->cpu != priv->cpu) 818 return -EINVAL; 819 820 eventid = get_event(event->attr.config); 821 portid = get_portid(event->attr.config); 822 evtype = get_evtype(event->attr.config); 823 if (evtype > FME_EVTYPE_MAX) 824 return -EINVAL; 825 826 hwc->event_base = evtype; 827 hwc->idx = (int)eventid; 828 hwc->config_base = portid; 829 830 event->destroy = fme_perf_event_destroy; 831 832 dev_dbg(priv->dev, "%s event=0x%x, evtype=0x%x, portid=0x%x,\n", 833 __func__, eventid, evtype, portid); 834 835 ops = get_event_ops(evtype); 836 if (ops->event_init) 837 return ops->event_init(priv, eventid, portid); 838 839 return 0; 840 } 841 842 static void fme_perf_event_update(struct perf_event *event) 843 { 844 struct fme_perf_event_ops *ops = get_event_ops(event->hw.event_base); 845 struct fme_perf_priv *priv = to_fme_perf_priv(event->pmu); 846 struct hw_perf_event *hwc = &event->hw; 847 u64 now, prev, delta; 848 849 now = ops->read_counter(priv, (u32)hwc->idx, hwc->config_base); 850 prev = local64_read(&hwc->prev_count); 851 delta = now - prev; 852 853 local64_add(delta, &event->count); 854 } 855 856 static void fme_perf_event_start(struct perf_event *event, int flags) 857 { 858 struct fme_perf_event_ops *ops = get_event_ops(event->hw.event_base); 859 struct fme_perf_priv *priv = to_fme_perf_priv(event->pmu); 860 struct hw_perf_event *hwc = &event->hw; 861 u64 count; 862 863 count = ops->read_counter(priv, (u32)hwc->idx, hwc->config_base); 864 local64_set(&hwc->prev_count, count); 865 } 866 867 static void fme_perf_event_stop(struct perf_event *event, int flags) 868 { 869 fme_perf_event_update(event); 870 } 871 872 static int fme_perf_event_add(struct perf_event *event, int flags) 873 { 874 if (flags & PERF_EF_START) 875 fme_perf_event_start(event, flags); 876 877 return 0; 878 } 879 880 static void fme_perf_event_del(struct perf_event *event, int flags) 881 { 882 fme_perf_event_stop(event, PERF_EF_UPDATE); 883 } 884 885 static void fme_perf_event_read(struct perf_event *event) 886 { 887 fme_perf_event_update(event); 888 } 889 890 static void fme_perf_setup_hardware(struct fme_perf_priv *priv) 891 { 892 void __iomem *base = priv->ioaddr; 893 u64 v; 894 895 /* read and save current working mode for fabric counters */ 896 v = readq(base + FAB_CTRL); 897 898 if (FIELD_GET(FAB_PORT_FILTER, v) == FAB_PORT_FILTER_DISABLE) 899 priv->fab_port_id = FME_PORTID_ROOT; 900 else 901 priv->fab_port_id = FIELD_GET(FAB_PORT_ID, v); 902 } 903 904 static int fme_perf_pmu_register(struct platform_device *pdev, 905 struct fme_perf_priv *priv) 906 { 907 struct pmu *pmu = &priv->pmu; 908 char *name; 909 int ret; 910 911 spin_lock_init(&priv->fab_lock); 912 913 fme_perf_setup_hardware(priv); 914 915 pmu->task_ctx_nr = perf_invalid_context; 916 pmu->attr_groups = fme_perf_groups; 917 pmu->attr_update = fme_perf_events_groups; 918 pmu->event_init = fme_perf_event_init; 919 pmu->add = fme_perf_event_add; 920 pmu->del = fme_perf_event_del; 921 pmu->start = fme_perf_event_start; 922 pmu->stop = fme_perf_event_stop; 923 pmu->read = fme_perf_event_read; 924 pmu->capabilities = PERF_PMU_CAP_NO_INTERRUPT | 925 PERF_PMU_CAP_NO_EXCLUDE; 926 927 name = devm_kasprintf(priv->dev, GFP_KERNEL, "dfl_fme%d", pdev->id); 928 929 ret = perf_pmu_register(pmu, name, -1); 930 if (ret) 931 return ret; 932 933 return 0; 934 } 935 936 static void fme_perf_pmu_unregister(struct fme_perf_priv *priv) 937 { 938 perf_pmu_unregister(&priv->pmu); 939 } 940 941 static int fme_perf_offline_cpu(unsigned int cpu, struct hlist_node *node) 942 { 943 struct fme_perf_priv *priv; 944 int target; 945 946 priv = hlist_entry_safe(node, struct fme_perf_priv, node); 947 948 if (cpu != priv->cpu) 949 return 0; 950 951 target = cpumask_any_but(cpu_online_mask, cpu); 952 if (target >= nr_cpu_ids) 953 return 0; 954 955 priv->cpu = target; 956 perf_pmu_migrate_context(&priv->pmu, cpu, target); 957 958 return 0; 959 } 960 961 static int fme_perf_init(struct platform_device *pdev, 962 struct dfl_feature *feature) 963 { 964 struct fme_perf_priv *priv; 965 int ret; 966 967 priv = devm_kzalloc(&pdev->dev, sizeof(*priv), GFP_KERNEL); 968 if (!priv) 969 return -ENOMEM; 970 971 priv->dev = &pdev->dev; 972 priv->ioaddr = feature->ioaddr; 973 priv->id = feature->id; 974 priv->cpu = raw_smp_processor_id(); 975 976 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, 977 "perf/fpga/dfl_fme:online", 978 NULL, fme_perf_offline_cpu); 979 if (ret < 0) 980 return ret; 981 982 priv->cpuhp_state = ret; 983 984 /* Register the pmu instance for cpu hotplug */ 985 ret = cpuhp_state_add_instance_nocalls(priv->cpuhp_state, &priv->node); 986 if (ret) 987 goto cpuhp_instance_err; 988 989 ret = fme_perf_pmu_register(pdev, priv); 990 if (ret) 991 goto pmu_register_err; 992 993 feature->priv = priv; 994 return 0; 995 996 pmu_register_err: 997 cpuhp_state_remove_instance_nocalls(priv->cpuhp_state, &priv->node); 998 cpuhp_instance_err: 999 cpuhp_remove_multi_state(priv->cpuhp_state); 1000 return ret; 1001 } 1002 1003 static void fme_perf_uinit(struct platform_device *pdev, 1004 struct dfl_feature *feature) 1005 { 1006 struct fme_perf_priv *priv = feature->priv; 1007 1008 fme_perf_pmu_unregister(priv); 1009 cpuhp_state_remove_instance_nocalls(priv->cpuhp_state, &priv->node); 1010 cpuhp_remove_multi_state(priv->cpuhp_state); 1011 } 1012 1013 const struct dfl_feature_id fme_perf_id_table[] = { 1014 {.id = FME_FEATURE_ID_GLOBAL_IPERF,}, 1015 {.id = FME_FEATURE_ID_GLOBAL_DPERF,}, 1016 {0,} 1017 }; 1018 1019 const struct dfl_feature_ops fme_perf_ops = { 1020 .init = fme_perf_init, 1021 .uinit = fme_perf_uinit, 1022 }; 1023