1 /* 2 * Cell Broadband Engine Performance Monitor 3 * 4 * (C) Copyright IBM Corporation 2001,2006 5 * 6 * Author: 7 * David Erb (djerb@us.ibm.com) 8 * Kevin Corry (kevcorry@us.ibm.com) 9 * 10 * This program is free software; you can redistribute it and/or modify 11 * it under the terms of the GNU General Public License as published by 12 * the Free Software Foundation; either version 2, or (at your option) 13 * any later version. 14 * 15 * This program is distributed in the hope that it will be useful, 16 * but WITHOUT ANY WARRANTY; without even the implied warranty of 17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 18 * GNU General Public License for more details. 19 * 20 * You should have received a copy of the GNU General Public License 21 * along with this program; if not, write to the Free Software 22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. 23 */ 24 25 #include <linux/interrupt.h> 26 #include <linux/types.h> 27 #include <linux/export.h> 28 #include <asm/io.h> 29 #include <asm/irq_regs.h> 30 #include <asm/machdep.h> 31 #include <asm/pmc.h> 32 #include <asm/reg.h> 33 #include <asm/spu.h> 34 #include <asm/cell-regs.h> 35 36 #include "interrupt.h" 37 38 /* 39 * When writing to write-only mmio addresses, save a shadow copy. All of the 40 * registers are 32-bit, but stored in the upper-half of a 64-bit field in 41 * pmd_regs. 42 */ 43 44 #define WRITE_WO_MMIO(reg, x) \ 45 do { \ 46 u32 _x = (x); \ 47 struct cbe_pmd_regs __iomem *pmd_regs; \ 48 struct cbe_pmd_shadow_regs *shadow_regs; \ 49 pmd_regs = cbe_get_cpu_pmd_regs(cpu); \ 50 shadow_regs = cbe_get_cpu_pmd_shadow_regs(cpu); \ 51 out_be64(&(pmd_regs->reg), (((u64)_x) << 32)); \ 52 shadow_regs->reg = _x; \ 53 } while (0) 54 55 #define READ_SHADOW_REG(val, reg) \ 56 do { \ 57 struct cbe_pmd_shadow_regs *shadow_regs; \ 58 shadow_regs = cbe_get_cpu_pmd_shadow_regs(cpu); \ 59 (val) = shadow_regs->reg; \ 60 } while (0) 61 62 #define READ_MMIO_UPPER32(val, reg) \ 63 do { \ 64 struct cbe_pmd_regs __iomem *pmd_regs; \ 65 pmd_regs = cbe_get_cpu_pmd_regs(cpu); \ 66 (val) = (u32)(in_be64(&pmd_regs->reg) >> 32); \ 67 } while (0) 68 69 /* 70 * Physical counter registers. 71 * Each physical counter can act as one 32-bit counter or two 16-bit counters. 72 */ 73 74 u32 cbe_read_phys_ctr(u32 cpu, u32 phys_ctr) 75 { 76 u32 val_in_latch, val = 0; 77 78 if (phys_ctr < NR_PHYS_CTRS) { 79 READ_SHADOW_REG(val_in_latch, counter_value_in_latch); 80 81 /* Read the latch or the actual counter, whichever is newer. */ 82 if (val_in_latch & (1 << phys_ctr)) { 83 READ_SHADOW_REG(val, pm_ctr[phys_ctr]); 84 } else { 85 READ_MMIO_UPPER32(val, pm_ctr[phys_ctr]); 86 } 87 } 88 89 return val; 90 } 91 EXPORT_SYMBOL_GPL(cbe_read_phys_ctr); 92 93 void cbe_write_phys_ctr(u32 cpu, u32 phys_ctr, u32 val) 94 { 95 struct cbe_pmd_shadow_regs *shadow_regs; 96 u32 pm_ctrl; 97 98 if (phys_ctr < NR_PHYS_CTRS) { 99 /* Writing to a counter only writes to a hardware latch. 100 * The new value is not propagated to the actual counter 101 * until the performance monitor is enabled. 102 */ 103 WRITE_WO_MMIO(pm_ctr[phys_ctr], val); 104 105 pm_ctrl = cbe_read_pm(cpu, pm_control); 106 if (pm_ctrl & CBE_PM_ENABLE_PERF_MON) { 107 /* The counters are already active, so we need to 108 * rewrite the pm_control register to "re-enable" 109 * the PMU. 110 */ 111 cbe_write_pm(cpu, pm_control, pm_ctrl); 112 } else { 113 shadow_regs = cbe_get_cpu_pmd_shadow_regs(cpu); 114 shadow_regs->counter_value_in_latch |= (1 << phys_ctr); 115 } 116 } 117 } 118 EXPORT_SYMBOL_GPL(cbe_write_phys_ctr); 119 120 /* 121 * "Logical" counter registers. 122 * These will read/write 16-bits or 32-bits depending on the 123 * current size of the counter. Counters 4 - 7 are always 16-bit. 124 */ 125 126 u32 cbe_read_ctr(u32 cpu, u32 ctr) 127 { 128 u32 val; 129 u32 phys_ctr = ctr & (NR_PHYS_CTRS - 1); 130 131 val = cbe_read_phys_ctr(cpu, phys_ctr); 132 133 if (cbe_get_ctr_size(cpu, phys_ctr) == 16) 134 val = (ctr < NR_PHYS_CTRS) ? (val >> 16) : (val & 0xffff); 135 136 return val; 137 } 138 EXPORT_SYMBOL_GPL(cbe_read_ctr); 139 140 void cbe_write_ctr(u32 cpu, u32 ctr, u32 val) 141 { 142 u32 phys_ctr; 143 u32 phys_val; 144 145 phys_ctr = ctr & (NR_PHYS_CTRS - 1); 146 147 if (cbe_get_ctr_size(cpu, phys_ctr) == 16) { 148 phys_val = cbe_read_phys_ctr(cpu, phys_ctr); 149 150 if (ctr < NR_PHYS_CTRS) 151 val = (val << 16) | (phys_val & 0xffff); 152 else 153 val = (val & 0xffff) | (phys_val & 0xffff0000); 154 } 155 156 cbe_write_phys_ctr(cpu, phys_ctr, val); 157 } 158 EXPORT_SYMBOL_GPL(cbe_write_ctr); 159 160 /* 161 * Counter-control registers. 162 * Each "logical" counter has a corresponding control register. 163 */ 164 165 u32 cbe_read_pm07_control(u32 cpu, u32 ctr) 166 { 167 u32 pm07_control = 0; 168 169 if (ctr < NR_CTRS) 170 READ_SHADOW_REG(pm07_control, pm07_control[ctr]); 171 172 return pm07_control; 173 } 174 EXPORT_SYMBOL_GPL(cbe_read_pm07_control); 175 176 void cbe_write_pm07_control(u32 cpu, u32 ctr, u32 val) 177 { 178 if (ctr < NR_CTRS) 179 WRITE_WO_MMIO(pm07_control[ctr], val); 180 } 181 EXPORT_SYMBOL_GPL(cbe_write_pm07_control); 182 183 /* 184 * Other PMU control registers. Most of these are write-only. 185 */ 186 187 u32 cbe_read_pm(u32 cpu, enum pm_reg_name reg) 188 { 189 u32 val = 0; 190 191 switch (reg) { 192 case group_control: 193 READ_SHADOW_REG(val, group_control); 194 break; 195 196 case debug_bus_control: 197 READ_SHADOW_REG(val, debug_bus_control); 198 break; 199 200 case trace_address: 201 READ_MMIO_UPPER32(val, trace_address); 202 break; 203 204 case ext_tr_timer: 205 READ_SHADOW_REG(val, ext_tr_timer); 206 break; 207 208 case pm_status: 209 READ_MMIO_UPPER32(val, pm_status); 210 break; 211 212 case pm_control: 213 READ_SHADOW_REG(val, pm_control); 214 break; 215 216 case pm_interval: 217 READ_MMIO_UPPER32(val, pm_interval); 218 break; 219 220 case pm_start_stop: 221 READ_SHADOW_REG(val, pm_start_stop); 222 break; 223 } 224 225 return val; 226 } 227 EXPORT_SYMBOL_GPL(cbe_read_pm); 228 229 void cbe_write_pm(u32 cpu, enum pm_reg_name reg, u32 val) 230 { 231 switch (reg) { 232 case group_control: 233 WRITE_WO_MMIO(group_control, val); 234 break; 235 236 case debug_bus_control: 237 WRITE_WO_MMIO(debug_bus_control, val); 238 break; 239 240 case trace_address: 241 WRITE_WO_MMIO(trace_address, val); 242 break; 243 244 case ext_tr_timer: 245 WRITE_WO_MMIO(ext_tr_timer, val); 246 break; 247 248 case pm_status: 249 WRITE_WO_MMIO(pm_status, val); 250 break; 251 252 case pm_control: 253 WRITE_WO_MMIO(pm_control, val); 254 break; 255 256 case pm_interval: 257 WRITE_WO_MMIO(pm_interval, val); 258 break; 259 260 case pm_start_stop: 261 WRITE_WO_MMIO(pm_start_stop, val); 262 break; 263 } 264 } 265 EXPORT_SYMBOL_GPL(cbe_write_pm); 266 267 /* 268 * Get/set the size of a physical counter to either 16 or 32 bits. 269 */ 270 271 u32 cbe_get_ctr_size(u32 cpu, u32 phys_ctr) 272 { 273 u32 pm_ctrl, size = 0; 274 275 if (phys_ctr < NR_PHYS_CTRS) { 276 pm_ctrl = cbe_read_pm(cpu, pm_control); 277 size = (pm_ctrl & CBE_PM_16BIT_CTR(phys_ctr)) ? 16 : 32; 278 } 279 280 return size; 281 } 282 EXPORT_SYMBOL_GPL(cbe_get_ctr_size); 283 284 void cbe_set_ctr_size(u32 cpu, u32 phys_ctr, u32 ctr_size) 285 { 286 u32 pm_ctrl; 287 288 if (phys_ctr < NR_PHYS_CTRS) { 289 pm_ctrl = cbe_read_pm(cpu, pm_control); 290 switch (ctr_size) { 291 case 16: 292 pm_ctrl |= CBE_PM_16BIT_CTR(phys_ctr); 293 break; 294 295 case 32: 296 pm_ctrl &= ~CBE_PM_16BIT_CTR(phys_ctr); 297 break; 298 } 299 cbe_write_pm(cpu, pm_control, pm_ctrl); 300 } 301 } 302 EXPORT_SYMBOL_GPL(cbe_set_ctr_size); 303 304 /* 305 * Enable/disable the entire performance monitoring unit. 306 * When we enable the PMU, all pending writes to counters get committed. 307 */ 308 309 void cbe_enable_pm(u32 cpu) 310 { 311 struct cbe_pmd_shadow_regs *shadow_regs; 312 u32 pm_ctrl; 313 314 shadow_regs = cbe_get_cpu_pmd_shadow_regs(cpu); 315 shadow_regs->counter_value_in_latch = 0; 316 317 pm_ctrl = cbe_read_pm(cpu, pm_control) | CBE_PM_ENABLE_PERF_MON; 318 cbe_write_pm(cpu, pm_control, pm_ctrl); 319 } 320 EXPORT_SYMBOL_GPL(cbe_enable_pm); 321 322 void cbe_disable_pm(u32 cpu) 323 { 324 u32 pm_ctrl; 325 pm_ctrl = cbe_read_pm(cpu, pm_control) & ~CBE_PM_ENABLE_PERF_MON; 326 cbe_write_pm(cpu, pm_control, pm_ctrl); 327 } 328 EXPORT_SYMBOL_GPL(cbe_disable_pm); 329 330 /* 331 * Reading from the trace_buffer. 332 * The trace buffer is two 64-bit registers. Reading from 333 * the second half automatically increments the trace_address. 334 */ 335 336 void cbe_read_trace_buffer(u32 cpu, u64 *buf) 337 { 338 struct cbe_pmd_regs __iomem *pmd_regs = cbe_get_cpu_pmd_regs(cpu); 339 340 *buf++ = in_be64(&pmd_regs->trace_buffer_0_63); 341 *buf++ = in_be64(&pmd_regs->trace_buffer_64_127); 342 } 343 EXPORT_SYMBOL_GPL(cbe_read_trace_buffer); 344 345 /* 346 * Enabling/disabling interrupts for the entire performance monitoring unit. 347 */ 348 349 u32 cbe_get_and_clear_pm_interrupts(u32 cpu) 350 { 351 /* Reading pm_status clears the interrupt bits. */ 352 return cbe_read_pm(cpu, pm_status); 353 } 354 EXPORT_SYMBOL_GPL(cbe_get_and_clear_pm_interrupts); 355 356 void cbe_enable_pm_interrupts(u32 cpu, u32 thread, u32 mask) 357 { 358 /* Set which node and thread will handle the next interrupt. */ 359 iic_set_interrupt_routing(cpu, thread, 0); 360 361 /* Enable the interrupt bits in the pm_status register. */ 362 if (mask) 363 cbe_write_pm(cpu, pm_status, mask); 364 } 365 EXPORT_SYMBOL_GPL(cbe_enable_pm_interrupts); 366 367 void cbe_disable_pm_interrupts(u32 cpu) 368 { 369 cbe_get_and_clear_pm_interrupts(cpu); 370 cbe_write_pm(cpu, pm_status, 0); 371 } 372 EXPORT_SYMBOL_GPL(cbe_disable_pm_interrupts); 373 374 static irqreturn_t cbe_pm_irq(int irq, void *dev_id) 375 { 376 perf_irq(get_irq_regs()); 377 return IRQ_HANDLED; 378 } 379 380 static int __init cbe_init_pm_irq(void) 381 { 382 unsigned int irq; 383 int rc, node; 384 385 for_each_online_node(node) { 386 irq = irq_create_mapping(NULL, IIC_IRQ_IOEX_PMI | 387 (node << IIC_IRQ_NODE_SHIFT)); 388 if (irq == NO_IRQ) { 389 printk("ERROR: Unable to allocate irq for node %d\n", 390 node); 391 return -EINVAL; 392 } 393 394 rc = request_irq(irq, cbe_pm_irq, 395 0, "cbe-pmu-0", NULL); 396 if (rc) { 397 printk("ERROR: Request for irq on node %d failed\n", 398 node); 399 return rc; 400 } 401 } 402 403 return 0; 404 } 405 machine_arch_initcall(cell, cbe_init_pm_irq); 406 407 void cbe_sync_irq(int node) 408 { 409 unsigned int irq; 410 411 irq = irq_find_mapping(NULL, 412 IIC_IRQ_IOEX_PMI 413 | (node << IIC_IRQ_NODE_SHIFT)); 414 415 if (irq == NO_IRQ) { 416 printk(KERN_WARNING "ERROR, unable to get existing irq %d " \ 417 "for node %d\n", irq, node); 418 return; 419 } 420 421 synchronize_irq(irq); 422 } 423 EXPORT_SYMBOL_GPL(cbe_sync_irq); 424 425