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/cpuidle.h> 22 #include <asm/kvm_ppc.h> 23 #include <asm/machdep.h> 24 #include <asm/opal.h> 25 #include <asm/smp.h> 26 27 #include "subcore.h" 28 #include "powernv.h" 29 30 31 /* 32 * Split/unsplit procedure: 33 * 34 * A core can be in one of three states, unsplit, 2-way split, and 4-way split. 35 * 36 * The mapping to subcores_per_core is simple: 37 * 38 * State | subcores_per_core 39 * ------------|------------------ 40 * Unsplit | 1 41 * 2-way split | 2 42 * 4-way split | 4 43 * 44 * The core is split along thread boundaries, the mapping between subcores and 45 * threads is as follows: 46 * 47 * Unsplit: 48 * ---------------------------- 49 * Subcore | 0 | 50 * ---------------------------- 51 * Thread | 0 1 2 3 4 5 6 7 | 52 * ---------------------------- 53 * 54 * 2-way split: 55 * ------------------------------------- 56 * Subcore | 0 | 1 | 57 * ------------------------------------- 58 * Thread | 0 1 2 3 | 4 5 6 7 | 59 * ------------------------------------- 60 * 61 * 4-way split: 62 * ----------------------------------------- 63 * Subcore | 0 | 1 | 2 | 3 | 64 * ----------------------------------------- 65 * Thread | 0 1 | 2 3 | 4 5 | 6 7 | 66 * ----------------------------------------- 67 * 68 * 69 * Transitions 70 * ----------- 71 * 72 * It is not possible to transition between either of the split states, the 73 * core must first be unsplit. The legal transitions are: 74 * 75 * ----------- --------------- 76 * | | <----> | 2-way split | 77 * | | --------------- 78 * | Unsplit | 79 * | | --------------- 80 * | | <----> | 4-way split | 81 * ----------- --------------- 82 * 83 * Unsplitting 84 * ----------- 85 * 86 * Unsplitting is the simpler procedure. It requires thread 0 to request the 87 * unsplit while all other threads NAP. 88 * 89 * Thread 0 clears HID0_POWER8_DYNLPARDIS (Dynamic LPAR Disable). This tells 90 * the hardware that if all threads except 0 are napping, the hardware should 91 * unsplit the core. 92 * 93 * Non-zero threads are sent to a NAP loop, they don't exit the loop until they 94 * see the core unsplit. 95 * 96 * Core 0 spins waiting for the hardware to see all the other threads napping 97 * and perform the unsplit. 98 * 99 * Once thread 0 sees the unsplit, it IPIs the secondary threads to wake them 100 * out of NAP. They will then see the core unsplit and exit the NAP loop. 101 * 102 * Splitting 103 * --------- 104 * 105 * The basic splitting procedure is fairly straight forward. However it is 106 * complicated by the fact that after the split occurs, the newly created 107 * subcores are not in a fully initialised state. 108 * 109 * Most notably the subcores do not have the correct value for SDR1, which 110 * means they must not be running in virtual mode when the split occurs. The 111 * subcores have separate timebases SPRs but these are pre-synchronised by 112 * opal. 113 * 114 * To begin with secondary threads are sent to an assembly routine. There they 115 * switch to real mode, so they are immune to the uninitialised SDR1 value. 116 * Once in real mode they indicate that they are in real mode, and spin waiting 117 * to see the core split. 118 * 119 * Thread 0 waits to see that all secondaries are in real mode, and then begins 120 * the splitting procedure. It firstly sets HID0_POWER8_DYNLPARDIS, which 121 * prevents the hardware from unsplitting. Then it sets the appropriate HID bit 122 * to request the split, and spins waiting to see that the split has happened. 123 * 124 * Concurrently the secondaries will notice the split. When they do they set up 125 * their SPRs, notably SDR1, and then they can return to virtual mode and exit 126 * the procedure. 127 */ 128 129 /* Initialised at boot by subcore_init() */ 130 static int subcores_per_core; 131 132 /* 133 * Used to communicate to offline cpus that we want them to pop out of the 134 * offline loop and do a split or unsplit. 135 * 136 * 0 - no split happening 137 * 1 - unsplit in progress 138 * 2 - split to 2 in progress 139 * 4 - split to 4 in progress 140 */ 141 static int new_split_mode; 142 143 static cpumask_var_t cpu_offline_mask; 144 145 struct split_state { 146 u8 step; 147 u8 master; 148 }; 149 150 static DEFINE_PER_CPU(struct split_state, split_state); 151 152 static void wait_for_sync_step(int step) 153 { 154 int i, cpu = smp_processor_id(); 155 156 for (i = cpu + 1; i < cpu + threads_per_core; i++) 157 while(per_cpu(split_state, i).step < step) 158 barrier(); 159 160 /* Order the wait loop vs any subsequent loads/stores. */ 161 mb(); 162 } 163 164 static void update_hid_in_slw(u64 hid0) 165 { 166 u64 idle_states = pnv_get_supported_cpuidle_states(); 167 168 if (idle_states & OPAL_PM_WINKLE_ENABLED) { 169 /* OPAL call to patch slw with the new HID0 value */ 170 u64 cpu_pir = hard_smp_processor_id(); 171 172 opal_slw_set_reg(cpu_pir, SPRN_HID0, hid0); 173 } 174 } 175 176 static void unsplit_core(void) 177 { 178 u64 hid0, mask; 179 int i, cpu; 180 181 mask = HID0_POWER8_2LPARMODE | HID0_POWER8_4LPARMODE; 182 183 cpu = smp_processor_id(); 184 if (cpu_thread_in_core(cpu) != 0) { 185 while (mfspr(SPRN_HID0) & mask) 186 power7_idle_insn(PNV_THREAD_NAP); 187 188 per_cpu(split_state, cpu).step = SYNC_STEP_UNSPLIT; 189 return; 190 } 191 192 hid0 = mfspr(SPRN_HID0); 193 hid0 &= ~HID0_POWER8_DYNLPARDIS; 194 update_power8_hid0(hid0); 195 update_hid_in_slw(hid0); 196 197 while (mfspr(SPRN_HID0) & mask) 198 cpu_relax(); 199 200 /* Wake secondaries out of NAP */ 201 for (i = cpu + 1; i < cpu + threads_per_core; i++) 202 smp_send_reschedule(i); 203 204 wait_for_sync_step(SYNC_STEP_UNSPLIT); 205 } 206 207 static void split_core(int new_mode) 208 { 209 struct { u64 value; u64 mask; } split_parms[2] = { 210 { HID0_POWER8_1TO2LPAR, HID0_POWER8_2LPARMODE }, 211 { HID0_POWER8_1TO4LPAR, HID0_POWER8_4LPARMODE } 212 }; 213 int i, cpu; 214 u64 hid0; 215 216 /* Convert new_mode (2 or 4) into an index into our parms array */ 217 i = (new_mode >> 1) - 1; 218 BUG_ON(i < 0 || i > 1); 219 220 cpu = smp_processor_id(); 221 if (cpu_thread_in_core(cpu) != 0) { 222 split_core_secondary_loop(&per_cpu(split_state, cpu).step); 223 return; 224 } 225 226 wait_for_sync_step(SYNC_STEP_REAL_MODE); 227 228 /* Write new mode */ 229 hid0 = mfspr(SPRN_HID0); 230 hid0 |= HID0_POWER8_DYNLPARDIS | split_parms[i].value; 231 update_power8_hid0(hid0); 232 update_hid_in_slw(hid0); 233 234 /* Wait for it to happen */ 235 while (!(mfspr(SPRN_HID0) & split_parms[i].mask)) 236 cpu_relax(); 237 } 238 239 static void cpu_do_split(int new_mode) 240 { 241 /* 242 * At boot subcores_per_core will be 0, so we will always unsplit at 243 * boot. In the usual case where the core is already unsplit it's a 244 * nop, and this just ensures the kernel's notion of the mode is 245 * consistent with the hardware. 246 */ 247 if (subcores_per_core != 1) 248 unsplit_core(); 249 250 if (new_mode != 1) 251 split_core(new_mode); 252 253 mb(); 254 per_cpu(split_state, smp_processor_id()).step = SYNC_STEP_FINISHED; 255 } 256 257 bool cpu_core_split_required(void) 258 { 259 smp_rmb(); 260 261 if (!new_split_mode) 262 return false; 263 264 cpu_do_split(new_split_mode); 265 266 return true; 267 } 268 269 void update_subcore_sibling_mask(void) 270 { 271 int cpu; 272 /* 273 * sibling mask for the first cpu. Left shift this by required bits 274 * to get sibling mask for the rest of the cpus. 275 */ 276 int sibling_mask_first_cpu = (1 << threads_per_subcore) - 1; 277 278 for_each_possible_cpu(cpu) { 279 int tid = cpu_thread_in_core(cpu); 280 int offset = (tid / threads_per_subcore) * threads_per_subcore; 281 int mask = sibling_mask_first_cpu << offset; 282 283 paca[cpu].subcore_sibling_mask = mask; 284 285 } 286 } 287 288 static int cpu_update_split_mode(void *data) 289 { 290 int cpu, new_mode = *(int *)data; 291 292 if (this_cpu_ptr(&split_state)->master) { 293 new_split_mode = new_mode; 294 smp_wmb(); 295 296 cpumask_andnot(cpu_offline_mask, cpu_present_mask, 297 cpu_online_mask); 298 299 /* This should work even though the cpu is offline */ 300 for_each_cpu(cpu, cpu_offline_mask) 301 smp_send_reschedule(cpu); 302 } 303 304 cpu_do_split(new_mode); 305 306 if (this_cpu_ptr(&split_state)->master) { 307 /* Wait for all cpus to finish before we touch subcores_per_core */ 308 for_each_present_cpu(cpu) { 309 if (cpu >= setup_max_cpus) 310 break; 311 312 while(per_cpu(split_state, cpu).step < SYNC_STEP_FINISHED) 313 barrier(); 314 } 315 316 new_split_mode = 0; 317 318 /* Make the new mode public */ 319 subcores_per_core = new_mode; 320 threads_per_subcore = threads_per_core / subcores_per_core; 321 update_subcore_sibling_mask(); 322 323 /* Make sure the new mode is written before we exit */ 324 mb(); 325 } 326 327 return 0; 328 } 329 330 static int set_subcores_per_core(int new_mode) 331 { 332 struct split_state *state; 333 int cpu; 334 335 if (kvm_hv_mode_active()) { 336 pr_err("Unable to change split core mode while KVM active.\n"); 337 return -EBUSY; 338 } 339 340 /* 341 * We are only called at boot, or from the sysfs write. If that ever 342 * changes we'll need a lock here. 343 */ 344 BUG_ON(new_mode < 1 || new_mode > 4 || new_mode == 3); 345 346 for_each_present_cpu(cpu) { 347 state = &per_cpu(split_state, cpu); 348 state->step = SYNC_STEP_INITIAL; 349 state->master = 0; 350 } 351 352 cpus_read_lock(); 353 354 /* This cpu will update the globals before exiting stop machine */ 355 this_cpu_ptr(&split_state)->master = 1; 356 357 /* Ensure state is consistent before we call the other cpus */ 358 mb(); 359 360 stop_machine_cpuslocked(cpu_update_split_mode, &new_mode, 361 cpu_online_mask); 362 363 cpus_read_unlock(); 364 365 return 0; 366 } 367 368 static ssize_t __used store_subcores_per_core(struct device *dev, 369 struct device_attribute *attr, const char *buf, 370 size_t count) 371 { 372 unsigned long val; 373 int rc; 374 375 /* We are serialised by the attribute lock */ 376 377 rc = sscanf(buf, "%lx", &val); 378 if (rc != 1) 379 return -EINVAL; 380 381 switch (val) { 382 case 1: 383 case 2: 384 case 4: 385 if (subcores_per_core == val) 386 /* Nothing to do */ 387 goto out; 388 break; 389 default: 390 return -EINVAL; 391 } 392 393 rc = set_subcores_per_core(val); 394 if (rc) 395 return rc; 396 397 out: 398 return count; 399 } 400 401 static ssize_t show_subcores_per_core(struct device *dev, 402 struct device_attribute *attr, char *buf) 403 { 404 return sprintf(buf, "%x\n", subcores_per_core); 405 } 406 407 static DEVICE_ATTR(subcores_per_core, 0644, 408 show_subcores_per_core, store_subcores_per_core); 409 410 static int subcore_init(void) 411 { 412 unsigned pvr_ver; 413 414 pvr_ver = PVR_VER(mfspr(SPRN_PVR)); 415 416 if (pvr_ver != PVR_POWER8 && 417 pvr_ver != PVR_POWER8E && 418 pvr_ver != PVR_POWER8NVL) 419 return 0; 420 421 /* 422 * We need all threads in a core to be present to split/unsplit so 423 * continue only if max_cpus are aligned to threads_per_core. 424 */ 425 if (setup_max_cpus % threads_per_core) 426 return 0; 427 428 BUG_ON(!alloc_cpumask_var(&cpu_offline_mask, GFP_KERNEL)); 429 430 set_subcores_per_core(1); 431 432 return device_create_file(cpu_subsys.dev_root, 433 &dev_attr_subcores_per_core); 434 } 435 machine_device_initcall(powernv, subcore_init); 436