1/* 2 * Copyright 2013, Michael (Ellerman|Neuling), IBM Corporation. 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public License 6 * as published by the Free Software Foundation; either version 7 * 2 of the License, or (at your option) any later version. 8 */ 9 10#define pr_fmt(fmt) "powernv: " fmt 11 12#include <linux/kernel.h> 13#include <linux/cpu.h> 14#include <linux/cpumask.h> 15#include <linux/device.h> 16#include <linux/gfp.h> 17#include <linux/smp.h> 18#include <linux/stop_machine.h> 19 20#include <asm/cputhreads.h> 21#include <asm/kvm_ppc.h> 22#include <asm/machdep.h> 23#include <asm/opal.h> 24#include <asm/smp.h> 25 26#include "subcore.h" 27#include "powernv.h" 28 29 30/* 31 * Split/unsplit procedure: 32 * 33 * A core can be in one of three states, unsplit, 2-way split, and 4-way split. 34 * 35 * The mapping to subcores_per_core is simple: 36 * 37 * State | subcores_per_core 38 * ------------|------------------ 39 * Unsplit | 1 40 * 2-way split | 2 41 * 4-way split | 4 42 * 43 * The core is split along thread boundaries, the mapping between subcores and 44 * threads is as follows: 45 * 46 * Unsplit: 47 * ---------------------------- 48 * Subcore | 0 | 49 * ---------------------------- 50 * Thread | 0 1 2 3 4 5 6 7 | 51 * ---------------------------- 52 * 53 * 2-way split: 54 * ------------------------------------- 55 * Subcore | 0 | 1 | 56 * ------------------------------------- 57 * Thread | 0 1 2 3 | 4 5 6 7 | 58 * ------------------------------------- 59 * 60 * 4-way split: 61 * ----------------------------------------- 62 * Subcore | 0 | 1 | 2 | 3 | 63 * ----------------------------------------- 64 * Thread | 0 1 | 2 3 | 4 5 | 6 7 | 65 * ----------------------------------------- 66 * 67 * 68 * Transitions 69 * ----------- 70 * 71 * It is not possible to transition between either of the split states, the 72 * core must first be unsplit. The legal transitions are: 73 * 74 * ----------- --------------- 75 * | | <----> | 2-way split | 76 * | | --------------- 77 * | Unsplit | 78 * | | --------------- 79 * | | <----> | 4-way split | 80 * ----------- --------------- 81 * 82 * Unsplitting 83 * ----------- 84 * 85 * Unsplitting is the simpler procedure. It requires thread 0 to request the 86 * unsplit while all other threads NAP. 87 * 88 * Thread 0 clears HID0_POWER8_DYNLPARDIS (Dynamic LPAR Disable). This tells 89 * the hardware that if all threads except 0 are napping, the hardware should 90 * unsplit the core. 91 * 92 * Non-zero threads are sent to a NAP loop, they don't exit the loop until they 93 * see the core unsplit. 94 * 95 * Core 0 spins waiting for the hardware to see all the other threads napping 96 * and perform the unsplit. 97 * 98 * Once thread 0 sees the unsplit, it IPIs the secondary threads to wake them 99 * out of NAP. They will then see the core unsplit and exit the NAP loop. 100 * 101 * Splitting 102 * --------- 103 * 104 * The basic splitting procedure is fairly straight forward. However it is 105 * complicated by the fact that after the split occurs, the newly created 106 * subcores are not in a fully initialised state. 107 * 108 * Most notably the subcores do not have the correct value for SDR1, which 109 * means they must not be running in virtual mode when the split occurs. The 110 * subcores have separate timebases SPRs but these are pre-synchronised by 111 * opal. 112 * 113 * To begin with secondary threads are sent to an assembly routine. There they 114 * switch to real mode, so they are immune to the uninitialised SDR1 value. 115 * Once in real mode they indicate that they are in real mode, and spin waiting 116 * to see the core split. 117 * 118 * Thread 0 waits to see that all secondaries are in real mode, and then begins 119 * the splitting procedure. It firstly sets HID0_POWER8_DYNLPARDIS, which 120 * prevents the hardware from unsplitting. Then it sets the appropriate HID bit 121 * to request the split, and spins waiting to see that the split has happened. 122 * 123 * Concurrently the secondaries will notice the split. When they do they set up 124 * their SPRs, notably SDR1, and then they can return to virtual mode and exit 125 * the procedure. 126 */ 127 128/* Initialised at boot by subcore_init() */ 129static int subcores_per_core; 130 131/* 132 * Used to communicate to offline cpus that we want them to pop out of the 133 * offline loop and do a split or unsplit. 134 * 135 * 0 - no split happening 136 * 1 - unsplit in progress 137 * 2 - split to 2 in progress 138 * 4 - split to 4 in progress 139 */ 140static int new_split_mode; 141 142static cpumask_var_t cpu_offline_mask; 143 144struct split_state { 145 u8 step; 146 u8 master; 147}; 148 149static DEFINE_PER_CPU(struct split_state, split_state); 150 151static void wait_for_sync_step(int step) 152{ 153 int i, cpu = smp_processor_id(); 154 155 for (i = cpu + 1; i < cpu + threads_per_core; i++) 156 while(per_cpu(split_state, i).step < step) 157 barrier(); 158 159 /* Order the wait loop vs any subsequent loads/stores. */ 160 mb(); 161} 162 163static void update_hid_in_slw(u64 hid0) 164{ 165 u64 idle_states = pnv_get_supported_cpuidle_states(); 166 167 if (idle_states & OPAL_PM_WINKLE_ENABLED) { 168 /* OPAL call to patch slw with the new HID0 value */ 169 u64 cpu_pir = hard_smp_processor_id(); 170 171 opal_slw_set_reg(cpu_pir, SPRN_HID0, hid0); 172 } 173} 174 175static void unsplit_core(void) 176{ 177 u64 hid0, mask; 178 int i, cpu; 179 180 mask = HID0_POWER8_2LPARMODE | HID0_POWER8_4LPARMODE; 181 182 cpu = smp_processor_id(); 183 if (cpu_thread_in_core(cpu) != 0) { 184 while (mfspr(SPRN_HID0) & mask) 185 power7_nap(0); 186 187 per_cpu(split_state, cpu).step = SYNC_STEP_UNSPLIT; 188 return; 189 } 190 191 hid0 = mfspr(SPRN_HID0); 192 hid0 &= ~HID0_POWER8_DYNLPARDIS; 193 update_power8_hid0(hid0); 194 update_hid_in_slw(hid0); 195 196 while (mfspr(SPRN_HID0) & mask) 197 cpu_relax(); 198 199 /* Wake secondaries out of NAP */ 200 for (i = cpu + 1; i < cpu + threads_per_core; i++) 201 smp_send_reschedule(i); 202 203 wait_for_sync_step(SYNC_STEP_UNSPLIT); 204} 205 206static void split_core(int new_mode) 207{ 208 struct { u64 value; u64 mask; } split_parms[2] = { 209 { HID0_POWER8_1TO2LPAR, HID0_POWER8_2LPARMODE }, 210 { HID0_POWER8_1TO4LPAR, HID0_POWER8_4LPARMODE } 211 }; 212 int i, cpu; 213 u64 hid0; 214 215 /* Convert new_mode (2 or 4) into an index into our parms array */ 216 i = (new_mode >> 1) - 1; 217 BUG_ON(i < 0 || i > 1); 218 219 cpu = smp_processor_id(); 220 if (cpu_thread_in_core(cpu) != 0) { 221 split_core_secondary_loop(&per_cpu(split_state, cpu).step); 222 return; 223 } 224 225 wait_for_sync_step(SYNC_STEP_REAL_MODE); 226 227 /* Write new mode */ 228 hid0 = mfspr(SPRN_HID0); 229 hid0 |= HID0_POWER8_DYNLPARDIS | split_parms[i].value; 230 update_power8_hid0(hid0); 231 update_hid_in_slw(hid0); 232 233 /* Wait for it to happen */ 234 while (!(mfspr(SPRN_HID0) & split_parms[i].mask)) 235 cpu_relax(); 236} 237 238static void cpu_do_split(int new_mode) 239{ 240 /* 241 * At boot subcores_per_core will be 0, so we will always unsplit at 242 * boot. In the usual case where the core is already unsplit it's a 243 * nop, and this just ensures the kernel's notion of the mode is 244 * consistent with the hardware. 245 */ 246 if (subcores_per_core != 1) 247 unsplit_core(); 248 249 if (new_mode != 1) 250 split_core(new_mode); 251 252 mb(); 253 per_cpu(split_state, smp_processor_id()).step = SYNC_STEP_FINISHED; 254} 255 256bool cpu_core_split_required(void) 257{ 258 smp_rmb(); 259 260 if (!new_split_mode) 261 return false; 262 263 cpu_do_split(new_split_mode); 264 265 return true; 266} 267 268void update_subcore_sibling_mask(void) 269{ 270 int cpu; 271 /* 272 * sibling mask for the first cpu. Left shift this by required bits 273 * to get sibling mask for the rest of the cpus. 274 */ 275 int sibling_mask_first_cpu = (1 << threads_per_subcore) - 1; 276 277 for_each_possible_cpu(cpu) { 278 int tid = cpu_thread_in_core(cpu); 279 int offset = (tid / threads_per_subcore) * threads_per_subcore; 280 int mask = sibling_mask_first_cpu << offset; 281 282 paca[cpu].subcore_sibling_mask = mask; 283 284 } 285} 286 287static int cpu_update_split_mode(void *data) 288{ 289 int cpu, new_mode = *(int *)data; 290 291 if (this_cpu_ptr(&split_state)->master) { 292 new_split_mode = new_mode; 293 smp_wmb(); 294 295 cpumask_andnot(cpu_offline_mask, cpu_present_mask, 296 cpu_online_mask); 297 298 /* This should work even though the cpu is offline */ 299 for_each_cpu(cpu, cpu_offline_mask) 300 smp_send_reschedule(cpu); 301 } 302 303 cpu_do_split(new_mode); 304 305 if (this_cpu_ptr(&split_state)->master) { 306 /* Wait for all cpus to finish before we touch subcores_per_core */ 307 for_each_present_cpu(cpu) { 308 if (cpu >= setup_max_cpus) 309 break; 310 311 while(per_cpu(split_state, cpu).step < SYNC_STEP_FINISHED) 312 barrier(); 313 } 314 315 new_split_mode = 0; 316 317 /* Make the new mode public */ 318 subcores_per_core = new_mode; 319 threads_per_subcore = threads_per_core / subcores_per_core; 320 update_subcore_sibling_mask(); 321 322 /* Make sure the new mode is written before we exit */ 323 mb(); 324 } 325 326 return 0; 327} 328 329static int set_subcores_per_core(int new_mode) 330{ 331 struct split_state *state; 332 int cpu; 333 334 if (kvm_hv_mode_active()) { 335 pr_err("Unable to change split core mode while KVM active.\n"); 336 return -EBUSY; 337 } 338 339 /* 340 * We are only called at boot, or from the sysfs write. If that ever 341 * changes we'll need a lock here. 342 */ 343 BUG_ON(new_mode < 1 || new_mode > 4 || new_mode == 3); 344 345 for_each_present_cpu(cpu) { 346 state = &per_cpu(split_state, cpu); 347 state->step = SYNC_STEP_INITIAL; 348 state->master = 0; 349 } 350 351 get_online_cpus(); 352 353 /* This cpu will update the globals before exiting stop machine */ 354 this_cpu_ptr(&split_state)->master = 1; 355 356 /* Ensure state is consistent before we call the other cpus */ 357 mb(); 358 359 stop_machine(cpu_update_split_mode, &new_mode, cpu_online_mask); 360 361 put_online_cpus(); 362 363 return 0; 364} 365 366static ssize_t __used store_subcores_per_core(struct device *dev, 367 struct device_attribute *attr, const char *buf, 368 size_t count) 369{ 370 unsigned long val; 371 int rc; 372 373 /* We are serialised by the attribute lock */ 374 375 rc = sscanf(buf, "%lx", &val); 376 if (rc != 1) 377 return -EINVAL; 378 379 switch (val) { 380 case 1: 381 case 2: 382 case 4: 383 if (subcores_per_core == val) 384 /* Nothing to do */ 385 goto out; 386 break; 387 default: 388 return -EINVAL; 389 } 390 391 rc = set_subcores_per_core(val); 392 if (rc) 393 return rc; 394 395out: 396 return count; 397} 398 399static ssize_t show_subcores_per_core(struct device *dev, 400 struct device_attribute *attr, char *buf) 401{ 402 return sprintf(buf, "%x\n", subcores_per_core); 403} 404 405static DEVICE_ATTR(subcores_per_core, 0644, 406 show_subcores_per_core, store_subcores_per_core); 407 408static int subcore_init(void) 409{ 410 if (!cpu_has_feature(CPU_FTR_ARCH_207S)) 411 return 0; 412 413 /* 414 * We need all threads in a core to be present to split/unsplit so 415 * continue only if max_cpus are aligned to threads_per_core. 416 */ 417 if (setup_max_cpus % threads_per_core) 418 return 0; 419 420 BUG_ON(!alloc_cpumask_var(&cpu_offline_mask, GFP_KERNEL)); 421 422 set_subcores_per_core(1); 423 424 return device_create_file(cpu_subsys.dev_root, 425 &dev_attr_subcores_per_core); 426} 427machine_device_initcall(powernv, subcore_init); 428