| 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
| 2 | /* |
| 3 | * PowerPC64 SLB support. |
| 4 | * |
| 5 | * Copyright (C) 2004 David Gibson <dwg@au.ibm.com>, IBM |
| 6 | * Based on earlier code written by: |
| 7 | * Dave Engebretsen and Mike Corrigan {engebret|mikejc}@us.ibm.com |
| 8 | * Copyright (c) 2001 Dave Engebretsen |
| 9 | * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM |
| 10 | */ |
| 11 | |
| 12 | #include <asm/interrupt.h> |
| 13 | #include <asm/mmu.h> |
| 14 | #include <asm/mmu_context.h> |
| 15 | #include <asm/paca.h> |
| 16 | #include <asm/lppaca.h> |
| 17 | #include <asm/ppc-opcode.h> |
| 18 | #include <asm/cputable.h> |
| 19 | #include <asm/cacheflush.h> |
| 20 | #include <asm/smp.h> |
| 21 | #include <linux/compiler.h> |
| 22 | #include <linux/context_tracking.h> |
| 23 | #include <linux/mm_types.h> |
| 24 | #include <linux/pgtable.h> |
| 25 | |
| 26 | #include <asm/udbg.h> |
| 27 | #include <asm/text-patching.h> |
| 28 | |
| 29 | #include "internal.h" |
| 30 | |
| 31 | |
| 32 | static long slb_allocate_user(struct mm_struct *mm, unsigned long ea); |
| 33 | |
| 34 | bool stress_slb_enabled __initdata; |
| 35 | |
| 36 | static int __init parse_stress_slb(char *p) |
| 37 | { |
| 38 | stress_slb_enabled = true; |
| 39 | return 0; |
| 40 | } |
| 41 | early_param("stress_slb" , parse_stress_slb); |
| 42 | |
| 43 | __ro_after_init DEFINE_STATIC_KEY_FALSE(stress_slb_key); |
| 44 | |
| 45 | bool no_slb_preload __initdata; |
| 46 | static int __init parse_no_slb_preload(char *p) |
| 47 | { |
| 48 | no_slb_preload = true; |
| 49 | return 0; |
| 50 | } |
| 51 | early_param("no_slb_preload" , parse_no_slb_preload); |
| 52 | __ro_after_init DEFINE_STATIC_KEY_FALSE(no_slb_preload_key); |
| 53 | |
| 54 | static void assert_slb_presence(bool present, unsigned long ea) |
| 55 | { |
| 56 | #ifdef CONFIG_DEBUG_VM |
| 57 | unsigned long tmp; |
| 58 | |
| 59 | WARN_ON_ONCE(mfmsr() & MSR_EE); |
| 60 | |
| 61 | if (!cpu_has_feature(CPU_FTR_ARCH_206)) |
| 62 | return; |
| 63 | |
| 64 | /* |
| 65 | * slbfee. requires bit 24 (PPC bit 39) be clear in RB. Hardware |
| 66 | * ignores all other bits from 0-27, so just clear them all. |
| 67 | */ |
| 68 | ea &= ~((1UL << SID_SHIFT) - 1); |
| 69 | asm volatile(__PPC_SLBFEE_DOT(%0, %1) : "=r" (tmp) : "r" (ea) : "cr0" ); |
| 70 | |
| 71 | WARN_ON(present == (tmp == 0)); |
| 72 | #endif |
| 73 | } |
| 74 | |
| 75 | static inline void slb_shadow_update(unsigned long ea, int ssize, |
| 76 | unsigned long flags, |
| 77 | enum slb_index index) |
| 78 | { |
| 79 | struct slb_shadow *p = get_slb_shadow(); |
| 80 | |
| 81 | /* |
| 82 | * Clear the ESID first so the entry is not valid while we are |
| 83 | * updating it. No write barriers are needed here, provided |
| 84 | * we only update the current CPU's SLB shadow buffer. |
| 85 | */ |
| 86 | WRITE_ONCE(p->save_area[index].esid, 0); |
| 87 | WRITE_ONCE(p->save_area[index].vsid, cpu_to_be64(mk_vsid_data(ea, ssize, flags))); |
| 88 | WRITE_ONCE(p->save_area[index].esid, cpu_to_be64(mk_esid_data(ea, ssize, index))); |
| 89 | } |
| 90 | |
| 91 | static inline void slb_shadow_clear(enum slb_index index) |
| 92 | { |
| 93 | WRITE_ONCE(get_slb_shadow()->save_area[index].esid, cpu_to_be64(index)); |
| 94 | } |
| 95 | |
| 96 | static inline void create_shadowed_slbe(unsigned long ea, int ssize, |
| 97 | unsigned long flags, |
| 98 | enum slb_index index) |
| 99 | { |
| 100 | /* |
| 101 | * Updating the shadow buffer before writing the SLB ensures |
| 102 | * we don't get a stale entry here if we get preempted by PHYP |
| 103 | * between these two statements. |
| 104 | */ |
| 105 | slb_shadow_update(ea, ssize, flags, index: index); |
| 106 | |
| 107 | assert_slb_presence(present: false, ea); |
| 108 | asm volatile("slbmte %0,%1" : |
| 109 | : "r" (mk_vsid_data(ea, ssize, flags)), |
| 110 | "r" (mk_esid_data(ea, ssize, index)) |
| 111 | : "memory" ); |
| 112 | } |
| 113 | |
| 114 | /* |
| 115 | * Insert bolted entries into SLB (which may not be empty, so don't clear |
| 116 | * slb_cache_ptr). |
| 117 | */ |
| 118 | void __slb_restore_bolted_realmode(void) |
| 119 | { |
| 120 | struct slb_shadow *p = get_slb_shadow(); |
| 121 | enum slb_index index; |
| 122 | |
| 123 | /* No isync needed because realmode. */ |
| 124 | for (index = 0; index < SLB_NUM_BOLTED; index++) { |
| 125 | asm volatile("slbmte %0,%1" : |
| 126 | : "r" (be64_to_cpu(p->save_area[index].vsid)), |
| 127 | "r" (be64_to_cpu(p->save_area[index].esid))); |
| 128 | } |
| 129 | |
| 130 | assert_slb_presence(present: true, ea: local_paca->kstack); |
| 131 | } |
| 132 | |
| 133 | /* |
| 134 | * Insert the bolted entries into an empty SLB. |
| 135 | */ |
| 136 | void slb_restore_bolted_realmode(void) |
| 137 | { |
| 138 | __slb_restore_bolted_realmode(); |
| 139 | get_paca()->slb_cache_ptr = 0; |
| 140 | |
| 141 | get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1; |
| 142 | get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap; |
| 143 | } |
| 144 | |
| 145 | /* |
| 146 | * This flushes all SLB entries including 0, so it must be realmode. |
| 147 | */ |
| 148 | void slb_flush_all_realmode(void) |
| 149 | { |
| 150 | asm volatile("slbmte %0,%0; slbia" : : "r" (0)); |
| 151 | } |
| 152 | |
| 153 | static __always_inline void __slb_flush_and_restore_bolted(bool preserve_kernel_lookaside) |
| 154 | { |
| 155 | struct slb_shadow *p = get_slb_shadow(); |
| 156 | unsigned long ksp_esid_data, ksp_vsid_data; |
| 157 | u32 ih; |
| 158 | |
| 159 | /* |
| 160 | * SLBIA IH=1 on ISA v2.05 and newer processors may preserve lookaside |
| 161 | * information created with Class=0 entries, which we use for kernel |
| 162 | * SLB entries (the SLB entries themselves are still invalidated). |
| 163 | * |
| 164 | * Older processors will ignore this optimisation. Over-invalidation |
| 165 | * is fine because we never rely on lookaside information existing. |
| 166 | */ |
| 167 | if (preserve_kernel_lookaside) |
| 168 | ih = 1; |
| 169 | else |
| 170 | ih = 0; |
| 171 | |
| 172 | ksp_esid_data = be64_to_cpu(p->save_area[KSTACK_INDEX].esid); |
| 173 | ksp_vsid_data = be64_to_cpu(p->save_area[KSTACK_INDEX].vsid); |
| 174 | |
| 175 | asm volatile(PPC_SLBIA(%0)" \n" |
| 176 | "slbmte %1, %2 \n" |
| 177 | :: "i" (ih), |
| 178 | "r" (ksp_vsid_data), |
| 179 | "r" (ksp_esid_data) |
| 180 | : "memory" ); |
| 181 | } |
| 182 | |
| 183 | /* |
| 184 | * This flushes non-bolted entries, it can be run in virtual mode. Must |
| 185 | * be called with interrupts disabled. |
| 186 | */ |
| 187 | void slb_flush_and_restore_bolted(void) |
| 188 | { |
| 189 | BUILD_BUG_ON(SLB_NUM_BOLTED != 2); |
| 190 | |
| 191 | WARN_ON(!irqs_disabled()); |
| 192 | |
| 193 | /* |
| 194 | * We can't take a PMU exception in the following code, so hard |
| 195 | * disable interrupts. |
| 196 | */ |
| 197 | hard_irq_disable(); |
| 198 | |
| 199 | isync(); |
| 200 | __slb_flush_and_restore_bolted(preserve_kernel_lookaside: false); |
| 201 | isync(); |
| 202 | |
| 203 | assert_slb_presence(present: true, ea: get_paca()->kstack); |
| 204 | |
| 205 | get_paca()->slb_cache_ptr = 0; |
| 206 | |
| 207 | get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1; |
| 208 | get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap; |
| 209 | } |
| 210 | |
| 211 | void slb_save_contents(struct slb_entry *slb_ptr) |
| 212 | { |
| 213 | int i; |
| 214 | unsigned long e, v; |
| 215 | |
| 216 | /* Save slb_cache_ptr value. */ |
| 217 | get_paca()->slb_save_cache_ptr = get_paca()->slb_cache_ptr; |
| 218 | |
| 219 | if (!slb_ptr) |
| 220 | return; |
| 221 | |
| 222 | for (i = 0; i < mmu_slb_size; i++) { |
| 223 | asm volatile("slbmfee %0,%1" : "=r" (e) : "r" (i)); |
| 224 | asm volatile("slbmfev %0,%1" : "=r" (v) : "r" (i)); |
| 225 | slb_ptr->esid = e; |
| 226 | slb_ptr->vsid = v; |
| 227 | slb_ptr++; |
| 228 | } |
| 229 | } |
| 230 | |
| 231 | void slb_dump_contents(struct slb_entry *slb_ptr) |
| 232 | { |
| 233 | int i, n; |
| 234 | unsigned long e, v; |
| 235 | unsigned long llp; |
| 236 | |
| 237 | if (!slb_ptr) |
| 238 | return; |
| 239 | |
| 240 | pr_err("SLB contents of cpu 0x%x\n" , smp_processor_id()); |
| 241 | |
| 242 | for (i = 0; i < mmu_slb_size; i++) { |
| 243 | e = slb_ptr->esid; |
| 244 | v = slb_ptr->vsid; |
| 245 | slb_ptr++; |
| 246 | |
| 247 | if (!e && !v) |
| 248 | continue; |
| 249 | |
| 250 | pr_err("%02d %016lx %016lx %s\n" , i, e, v, |
| 251 | (e & SLB_ESID_V) ? "VALID" : "NOT VALID" ); |
| 252 | |
| 253 | if (!(e & SLB_ESID_V)) |
| 254 | continue; |
| 255 | |
| 256 | llp = v & SLB_VSID_LLP; |
| 257 | if (v & SLB_VSID_B_1T) { |
| 258 | pr_err(" 1T ESID=%9lx VSID=%13lx LLP:%3lx\n" , |
| 259 | GET_ESID_1T(e), |
| 260 | (v & ~SLB_VSID_B) >> SLB_VSID_SHIFT_1T, llp); |
| 261 | } else { |
| 262 | pr_err(" 256M ESID=%9lx VSID=%13lx LLP:%3lx\n" , |
| 263 | GET_ESID(e), |
| 264 | (v & ~SLB_VSID_B) >> SLB_VSID_SHIFT, llp); |
| 265 | } |
| 266 | } |
| 267 | |
| 268 | if (!early_cpu_has_feature(CPU_FTR_ARCH_300)) { |
| 269 | /* RR is not so useful as it's often not used for allocation */ |
| 270 | pr_err("SLB RR allocator index %d\n" , get_paca()->stab_rr); |
| 271 | |
| 272 | /* Dump slb cache entires as well. */ |
| 273 | pr_err("SLB cache ptr value = %d\n" , get_paca()->slb_save_cache_ptr); |
| 274 | pr_err("Valid SLB cache entries:\n" ); |
| 275 | n = min_t(int, get_paca()->slb_save_cache_ptr, SLB_CACHE_ENTRIES); |
| 276 | for (i = 0; i < n; i++) |
| 277 | pr_err("%02d EA[0-35]=%9x\n" , i, get_paca()->slb_cache[i]); |
| 278 | pr_err("Rest of SLB cache entries:\n" ); |
| 279 | for (i = n; i < SLB_CACHE_ENTRIES; i++) |
| 280 | pr_err("%02d EA[0-35]=%9x\n" , i, get_paca()->slb_cache[i]); |
| 281 | } |
| 282 | } |
| 283 | |
| 284 | void slb_vmalloc_update(void) |
| 285 | { |
| 286 | /* |
| 287 | * vmalloc is not bolted, so just have to flush non-bolted. |
| 288 | */ |
| 289 | slb_flush_and_restore_bolted(); |
| 290 | } |
| 291 | |
| 292 | static bool preload_hit(struct thread_info *ti, unsigned long esid) |
| 293 | { |
| 294 | unsigned char i; |
| 295 | |
| 296 | for (i = 0; i < ti->slb_preload_nr; i++) { |
| 297 | unsigned char idx; |
| 298 | |
| 299 | idx = (ti->slb_preload_tail + i) % SLB_PRELOAD_NR; |
| 300 | if (esid == ti->slb_preload_esid[idx]) |
| 301 | return true; |
| 302 | } |
| 303 | return false; |
| 304 | } |
| 305 | |
| 306 | static void preload_add(struct thread_info *ti, unsigned long ea) |
| 307 | { |
| 308 | unsigned char idx; |
| 309 | unsigned long esid; |
| 310 | |
| 311 | if (slb_preload_disabled()) |
| 312 | return; |
| 313 | |
| 314 | if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) { |
| 315 | /* EAs are stored >> 28 so 256MB segments don't need clearing */ |
| 316 | if (ea & ESID_MASK_1T) |
| 317 | ea &= ESID_MASK_1T; |
| 318 | } |
| 319 | |
| 320 | esid = ea >> SID_SHIFT; |
| 321 | |
| 322 | if (preload_hit(ti, esid)) |
| 323 | return; |
| 324 | |
| 325 | idx = (ti->slb_preload_tail + ti->slb_preload_nr) % SLB_PRELOAD_NR; |
| 326 | ti->slb_preload_esid[idx] = esid; |
| 327 | if (ti->slb_preload_nr == SLB_PRELOAD_NR) |
| 328 | ti->slb_preload_tail = (ti->slb_preload_tail + 1) % SLB_PRELOAD_NR; |
| 329 | else |
| 330 | ti->slb_preload_nr++; |
| 331 | } |
| 332 | |
| 333 | static void preload_age(struct thread_info *ti) |
| 334 | { |
| 335 | if (!ti->slb_preload_nr) |
| 336 | return; |
| 337 | ti->slb_preload_nr--; |
| 338 | ti->slb_preload_tail = (ti->slb_preload_tail + 1) % SLB_PRELOAD_NR; |
| 339 | } |
| 340 | |
| 341 | static void slb_cache_slbie_kernel(unsigned int index) |
| 342 | { |
| 343 | unsigned long slbie_data = get_paca()->slb_cache[index]; |
| 344 | unsigned long ksp = get_paca()->kstack; |
| 345 | |
| 346 | slbie_data <<= SID_SHIFT; |
| 347 | slbie_data |= 0xc000000000000000ULL; |
| 348 | if ((ksp & slb_esid_mask(mmu_kernel_ssize)) == slbie_data) |
| 349 | return; |
| 350 | slbie_data |= mmu_kernel_ssize << SLBIE_SSIZE_SHIFT; |
| 351 | |
| 352 | asm volatile("slbie %0" : : "r" (slbie_data)); |
| 353 | } |
| 354 | |
| 355 | static void slb_cache_slbie_user(unsigned int index) |
| 356 | { |
| 357 | unsigned long slbie_data = get_paca()->slb_cache[index]; |
| 358 | |
| 359 | slbie_data <<= SID_SHIFT; |
| 360 | slbie_data |= user_segment_size(slbie_data) << SLBIE_SSIZE_SHIFT; |
| 361 | slbie_data |= SLBIE_C; /* user slbs have C=1 */ |
| 362 | |
| 363 | asm volatile("slbie %0" : : "r" (slbie_data)); |
| 364 | } |
| 365 | |
| 366 | /* Flush all user entries from the segment table of the current processor. */ |
| 367 | void switch_slb(struct task_struct *tsk, struct mm_struct *mm) |
| 368 | { |
| 369 | struct thread_info *ti = task_thread_info(tsk); |
| 370 | unsigned char i; |
| 371 | |
| 372 | /* |
| 373 | * We need interrupts hard-disabled here, not just soft-disabled, |
| 374 | * so that a PMU interrupt can't occur, which might try to access |
| 375 | * user memory (to get a stack trace) and possible cause an SLB miss |
| 376 | * which would update the slb_cache/slb_cache_ptr fields in the PACA. |
| 377 | */ |
| 378 | hard_irq_disable(); |
| 379 | isync(); |
| 380 | if (stress_slb()) { |
| 381 | __slb_flush_and_restore_bolted(preserve_kernel_lookaside: false); |
| 382 | isync(); |
| 383 | get_paca()->slb_cache_ptr = 0; |
| 384 | get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1; |
| 385 | |
| 386 | } else if (cpu_has_feature(CPU_FTR_ARCH_300)) { |
| 387 | /* |
| 388 | * SLBIA IH=3 invalidates all Class=1 SLBEs and their |
| 389 | * associated lookaside structures, which matches what |
| 390 | * switch_slb wants. So ARCH_300 does not use the slb |
| 391 | * cache. |
| 392 | */ |
| 393 | asm volatile(PPC_SLBIA(3)); |
| 394 | |
| 395 | } else { |
| 396 | unsigned long offset = get_paca()->slb_cache_ptr; |
| 397 | |
| 398 | if (!mmu_has_feature(MMU_FTR_NO_SLBIE_B) && |
| 399 | offset <= SLB_CACHE_ENTRIES) { |
| 400 | /* |
| 401 | * Could assert_slb_presence(true) here, but |
| 402 | * hypervisor or machine check could have come |
| 403 | * in and removed the entry at this point. |
| 404 | */ |
| 405 | |
| 406 | for (i = 0; i < offset; i++) |
| 407 | slb_cache_slbie_user(index: i); |
| 408 | |
| 409 | /* Workaround POWER5 < DD2.1 issue */ |
| 410 | if (!cpu_has_feature(CPU_FTR_ARCH_207S) && offset == 1) |
| 411 | slb_cache_slbie_user(index: 0); |
| 412 | |
| 413 | } else { |
| 414 | /* Flush but retain kernel lookaside information */ |
| 415 | __slb_flush_and_restore_bolted(preserve_kernel_lookaside: true); |
| 416 | isync(); |
| 417 | |
| 418 | get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1; |
| 419 | } |
| 420 | |
| 421 | get_paca()->slb_cache_ptr = 0; |
| 422 | } |
| 423 | get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap; |
| 424 | |
| 425 | copy_mm_to_paca(mm); |
| 426 | |
| 427 | if (slb_preload_disabled()) |
| 428 | return; |
| 429 | |
| 430 | /* |
| 431 | * We gradually age out SLBs after a number of context switches to |
| 432 | * reduce reload overhead of unused entries (like we do with FP/VEC |
| 433 | * reload). Each time we wrap 256 switches, take an entry out of the |
| 434 | * SLB preload cache. |
| 435 | */ |
| 436 | tsk->thread.load_slb++; |
| 437 | if (!tsk->thread.load_slb) { |
| 438 | unsigned long pc = KSTK_EIP(tsk); |
| 439 | |
| 440 | preload_age(ti); |
| 441 | preload_add(ti, ea: pc); |
| 442 | } |
| 443 | |
| 444 | for (i = 0; i < ti->slb_preload_nr; i++) { |
| 445 | unsigned char idx; |
| 446 | unsigned long ea; |
| 447 | |
| 448 | idx = (ti->slb_preload_tail + i) % SLB_PRELOAD_NR; |
| 449 | ea = (unsigned long)ti->slb_preload_esid[idx] << SID_SHIFT; |
| 450 | |
| 451 | slb_allocate_user(mm, ea); |
| 452 | } |
| 453 | |
| 454 | /* |
| 455 | * Synchronize slbmte preloads with possible subsequent user memory |
| 456 | * address accesses by the kernel (user mode won't happen until |
| 457 | * rfid, which is safe). |
| 458 | */ |
| 459 | isync(); |
| 460 | } |
| 461 | |
| 462 | void slb_set_size(u16 size) |
| 463 | { |
| 464 | mmu_slb_size = size; |
| 465 | } |
| 466 | |
| 467 | void slb_initialize(void) |
| 468 | { |
| 469 | unsigned long linear_llp, vmalloc_llp, io_llp; |
| 470 | unsigned long lflags; |
| 471 | static int slb_encoding_inited; |
| 472 | #ifdef CONFIG_SPARSEMEM_VMEMMAP |
| 473 | unsigned long vmemmap_llp; |
| 474 | #endif |
| 475 | |
| 476 | /* Prepare our SLB miss handler based on our page size */ |
| 477 | linear_llp = mmu_psize_defs[mmu_linear_psize].sllp; |
| 478 | io_llp = mmu_psize_defs[mmu_io_psize].sllp; |
| 479 | vmalloc_llp = mmu_psize_defs[mmu_vmalloc_psize].sllp; |
| 480 | get_paca()->vmalloc_sllp = SLB_VSID_KERNEL | vmalloc_llp; |
| 481 | #ifdef CONFIG_SPARSEMEM_VMEMMAP |
| 482 | vmemmap_llp = mmu_psize_defs[mmu_vmemmap_psize].sllp; |
| 483 | #endif |
| 484 | if (!slb_encoding_inited) { |
| 485 | slb_encoding_inited = 1; |
| 486 | pr_devel("SLB: linear LLP = %04lx\n" , linear_llp); |
| 487 | pr_devel("SLB: io LLP = %04lx\n" , io_llp); |
| 488 | #ifdef CONFIG_SPARSEMEM_VMEMMAP |
| 489 | pr_devel("SLB: vmemmap LLP = %04lx\n" , vmemmap_llp); |
| 490 | #endif |
| 491 | } |
| 492 | |
| 493 | get_paca()->stab_rr = SLB_NUM_BOLTED - 1; |
| 494 | get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1; |
| 495 | get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap; |
| 496 | |
| 497 | lflags = SLB_VSID_KERNEL | linear_llp; |
| 498 | |
| 499 | /* Invalidate the entire SLB (even entry 0) & all the ERATS */ |
| 500 | asm volatile("isync" :::"memory" ); |
| 501 | asm volatile("slbmte %0,%0" ::"r" (0) : "memory" ); |
| 502 | asm volatile("isync; slbia; isync" :::"memory" ); |
| 503 | create_shadowed_slbe(PAGE_OFFSET, mmu_kernel_ssize, lflags, LINEAR_INDEX); |
| 504 | |
| 505 | /* |
| 506 | * For the boot cpu, we're running on the stack in init_thread_union, |
| 507 | * which is in the first segment of the linear mapping, and also |
| 508 | * get_paca()->kstack hasn't been initialized yet. |
| 509 | * For secondary cpus, we need to bolt the kernel stack entry now. |
| 510 | */ |
| 511 | slb_shadow_clear(KSTACK_INDEX); |
| 512 | if (raw_smp_processor_id() != boot_cpuid && |
| 513 | (get_paca()->kstack & slb_esid_mask(mmu_kernel_ssize)) > PAGE_OFFSET) |
| 514 | create_shadowed_slbe(get_paca()->kstack, |
| 515 | mmu_kernel_ssize, lflags, KSTACK_INDEX); |
| 516 | |
| 517 | asm volatile("isync" :::"memory" ); |
| 518 | } |
| 519 | |
| 520 | static void slb_cache_update(unsigned long esid_data) |
| 521 | { |
| 522 | int slb_cache_index; |
| 523 | |
| 524 | if (cpu_has_feature(CPU_FTR_ARCH_300)) |
| 525 | return; /* ISAv3.0B and later does not use slb_cache */ |
| 526 | |
| 527 | if (stress_slb()) |
| 528 | return; |
| 529 | |
| 530 | /* |
| 531 | * Now update slb cache entries |
| 532 | */ |
| 533 | slb_cache_index = local_paca->slb_cache_ptr; |
| 534 | if (slb_cache_index < SLB_CACHE_ENTRIES) { |
| 535 | /* |
| 536 | * We have space in slb cache for optimized switch_slb(). |
| 537 | * Top 36 bits from esid_data as per ISA |
| 538 | */ |
| 539 | local_paca->slb_cache[slb_cache_index++] = esid_data >> SID_SHIFT; |
| 540 | local_paca->slb_cache_ptr++; |
| 541 | } else { |
| 542 | /* |
| 543 | * Our cache is full and the current cache content strictly |
| 544 | * doesn't indicate the active SLB contents. Bump the ptr |
| 545 | * so that switch_slb() will ignore the cache. |
| 546 | */ |
| 547 | local_paca->slb_cache_ptr = SLB_CACHE_ENTRIES + 1; |
| 548 | } |
| 549 | } |
| 550 | |
| 551 | static enum slb_index alloc_slb_index(bool kernel) |
| 552 | { |
| 553 | enum slb_index index; |
| 554 | |
| 555 | /* |
| 556 | * The allocation bitmaps can become out of synch with the SLB |
| 557 | * when the _switch code does slbie when bolting a new stack |
| 558 | * segment and it must not be anywhere else in the SLB. This leaves |
| 559 | * a kernel allocated entry that is unused in the SLB. With very |
| 560 | * large systems or small segment sizes, the bitmaps could slowly |
| 561 | * fill with these entries. They will eventually be cleared out |
| 562 | * by the round robin allocator in that case, so it's probably not |
| 563 | * worth accounting for. |
| 564 | */ |
| 565 | |
| 566 | /* |
| 567 | * SLBs beyond 32 entries are allocated with stab_rr only |
| 568 | * POWER7/8/9 have 32 SLB entries, this could be expanded if a |
| 569 | * future CPU has more. |
| 570 | */ |
| 571 | if (local_paca->slb_used_bitmap != U32_MAX) { |
| 572 | index = ffz(local_paca->slb_used_bitmap); |
| 573 | local_paca->slb_used_bitmap |= 1U << index; |
| 574 | if (kernel) |
| 575 | local_paca->slb_kern_bitmap |= 1U << index; |
| 576 | } else { |
| 577 | /* round-robin replacement of slb starting at SLB_NUM_BOLTED. */ |
| 578 | index = local_paca->stab_rr; |
| 579 | if (index < (mmu_slb_size - 1)) |
| 580 | index++; |
| 581 | else |
| 582 | index = SLB_NUM_BOLTED; |
| 583 | local_paca->stab_rr = index; |
| 584 | if (index < 32) { |
| 585 | if (kernel) |
| 586 | local_paca->slb_kern_bitmap |= 1U << index; |
| 587 | else |
| 588 | local_paca->slb_kern_bitmap &= ~(1U << index); |
| 589 | } |
| 590 | } |
| 591 | BUG_ON(index < SLB_NUM_BOLTED); |
| 592 | |
| 593 | return index; |
| 594 | } |
| 595 | |
| 596 | static long slb_insert_entry(unsigned long ea, unsigned long context, |
| 597 | unsigned long flags, int ssize, bool kernel) |
| 598 | { |
| 599 | unsigned long vsid; |
| 600 | unsigned long vsid_data, esid_data; |
| 601 | enum slb_index index; |
| 602 | |
| 603 | vsid = get_vsid(context, ea, ssize); |
| 604 | if (!vsid) |
| 605 | return -EFAULT; |
| 606 | |
| 607 | /* |
| 608 | * There must not be a kernel SLB fault in alloc_slb_index or before |
| 609 | * slbmte here or the allocation bitmaps could get out of whack with |
| 610 | * the SLB. |
| 611 | * |
| 612 | * User SLB faults or preloads take this path which might get inlined |
| 613 | * into the caller, so add compiler barriers here to ensure unsafe |
| 614 | * memory accesses do not come between. |
| 615 | */ |
| 616 | barrier(); |
| 617 | |
| 618 | index = alloc_slb_index(kernel); |
| 619 | |
| 620 | vsid_data = __mk_vsid_data(vsid, ssize, flags); |
| 621 | esid_data = mk_esid_data(ea, ssize, index); |
| 622 | |
| 623 | /* |
| 624 | * No need for an isync before or after this slbmte. The exception |
| 625 | * we enter with and the rfid we exit with are context synchronizing. |
| 626 | * User preloads should add isync afterwards in case the kernel |
| 627 | * accesses user memory before it returns to userspace with rfid. |
| 628 | */ |
| 629 | assert_slb_presence(present: false, ea); |
| 630 | if (stress_slb()) { |
| 631 | int slb_cache_index = local_paca->slb_cache_ptr; |
| 632 | |
| 633 | /* |
| 634 | * stress_slb() does not use slb cache, repurpose as a |
| 635 | * cache of inserted (non-bolted) kernel SLB entries. All |
| 636 | * non-bolted kernel entries are flushed on any user fault, |
| 637 | * or if there are already 3 non-boled kernel entries. |
| 638 | */ |
| 639 | BUILD_BUG_ON(SLB_CACHE_ENTRIES < 3); |
| 640 | if (!kernel || slb_cache_index == 3) { |
| 641 | int i; |
| 642 | |
| 643 | for (i = 0; i < slb_cache_index; i++) |
| 644 | slb_cache_slbie_kernel(index: i); |
| 645 | slb_cache_index = 0; |
| 646 | } |
| 647 | |
| 648 | if (kernel) |
| 649 | local_paca->slb_cache[slb_cache_index++] = esid_data >> SID_SHIFT; |
| 650 | local_paca->slb_cache_ptr = slb_cache_index; |
| 651 | } |
| 652 | asm volatile("slbmte %0, %1" : : "r" (vsid_data), "r" (esid_data)); |
| 653 | |
| 654 | barrier(); |
| 655 | |
| 656 | if (!kernel) |
| 657 | slb_cache_update(esid_data); |
| 658 | |
| 659 | return 0; |
| 660 | } |
| 661 | |
| 662 | static long slb_allocate_kernel(unsigned long ea, unsigned long id) |
| 663 | { |
| 664 | unsigned long context; |
| 665 | unsigned long flags; |
| 666 | int ssize; |
| 667 | |
| 668 | if (id == LINEAR_MAP_REGION_ID) { |
| 669 | |
| 670 | /* We only support upto H_MAX_PHYSMEM_BITS */ |
| 671 | if ((ea & EA_MASK) > (1UL << H_MAX_PHYSMEM_BITS)) |
| 672 | return -EFAULT; |
| 673 | |
| 674 | flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_linear_psize].sllp; |
| 675 | |
| 676 | #ifdef CONFIG_SPARSEMEM_VMEMMAP |
| 677 | } else if (id == VMEMMAP_REGION_ID) { |
| 678 | |
| 679 | if (ea >= H_VMEMMAP_END) |
| 680 | return -EFAULT; |
| 681 | |
| 682 | flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_vmemmap_psize].sllp; |
| 683 | #endif |
| 684 | } else if (id == VMALLOC_REGION_ID) { |
| 685 | |
| 686 | if (ea >= H_VMALLOC_END) |
| 687 | return -EFAULT; |
| 688 | |
| 689 | flags = local_paca->vmalloc_sllp; |
| 690 | |
| 691 | } else if (id == IO_REGION_ID) { |
| 692 | |
| 693 | if (ea >= H_KERN_IO_END) |
| 694 | return -EFAULT; |
| 695 | |
| 696 | flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_io_psize].sllp; |
| 697 | |
| 698 | } else { |
| 699 | return -EFAULT; |
| 700 | } |
| 701 | |
| 702 | ssize = MMU_SEGSIZE_1T; |
| 703 | if (!mmu_has_feature(MMU_FTR_1T_SEGMENT)) |
| 704 | ssize = MMU_SEGSIZE_256M; |
| 705 | |
| 706 | context = get_kernel_context(ea); |
| 707 | |
| 708 | return slb_insert_entry(ea, context, flags, ssize, kernel: true); |
| 709 | } |
| 710 | |
| 711 | static long slb_allocate_user(struct mm_struct *mm, unsigned long ea) |
| 712 | { |
| 713 | unsigned long context; |
| 714 | unsigned long flags; |
| 715 | int bpsize; |
| 716 | int ssize; |
| 717 | |
| 718 | /* |
| 719 | * consider this as bad access if we take a SLB miss |
| 720 | * on an address above addr limit. |
| 721 | */ |
| 722 | if (ea >= mm_ctx_slb_addr_limit(&mm->context)) |
| 723 | return -EFAULT; |
| 724 | |
| 725 | context = get_user_context(&mm->context, ea); |
| 726 | if (!context) |
| 727 | return -EFAULT; |
| 728 | |
| 729 | if (unlikely(ea >= H_PGTABLE_RANGE)) { |
| 730 | WARN_ON(1); |
| 731 | return -EFAULT; |
| 732 | } |
| 733 | |
| 734 | ssize = user_segment_size(ea); |
| 735 | |
| 736 | bpsize = get_slice_psize(mm, ea); |
| 737 | flags = SLB_VSID_USER | mmu_psize_defs[bpsize].sllp; |
| 738 | |
| 739 | return slb_insert_entry(ea, context, flags, ssize, kernel: false); |
| 740 | } |
| 741 | |
| 742 | DEFINE_INTERRUPT_HANDLER_RAW(do_slb_fault) |
| 743 | { |
| 744 | unsigned long ea = regs->dar; |
| 745 | unsigned long id = get_region_id(ea); |
| 746 | |
| 747 | /* IRQs are not reconciled here, so can't check irqs_disabled */ |
| 748 | VM_WARN_ON(mfmsr() & MSR_EE); |
| 749 | |
| 750 | if (regs_is_unrecoverable(regs)) |
| 751 | return -EINVAL; |
| 752 | |
| 753 | /* |
| 754 | * SLB kernel faults must be very careful not to touch anything that is |
| 755 | * not bolted. E.g., PACA and global variables are okay, mm->context |
| 756 | * stuff is not. SLB user faults may access all of memory (and induce |
| 757 | * one recursive SLB kernel fault), so the kernel fault must not |
| 758 | * trample on the user fault state at those points. |
| 759 | */ |
| 760 | |
| 761 | /* |
| 762 | * This is a raw interrupt handler, for performance, so that |
| 763 | * fast_interrupt_return can be used. The handler must not touch local |
| 764 | * irq state, or schedule. We could test for usermode and upgrade to a |
| 765 | * normal process context (synchronous) interrupt for those, which |
| 766 | * would make them first-class kernel code and able to be traced and |
| 767 | * instrumented, although performance would suffer a bit, it would |
| 768 | * probably be a good tradeoff. |
| 769 | */ |
| 770 | if (id >= LINEAR_MAP_REGION_ID) { |
| 771 | long err; |
| 772 | #ifdef CONFIG_DEBUG_VM |
| 773 | /* Catch recursive kernel SLB faults. */ |
| 774 | BUG_ON(local_paca->in_kernel_slb_handler); |
| 775 | local_paca->in_kernel_slb_handler = 1; |
| 776 | #endif |
| 777 | err = slb_allocate_kernel(ea, id); |
| 778 | #ifdef CONFIG_DEBUG_VM |
| 779 | local_paca->in_kernel_slb_handler = 0; |
| 780 | #endif |
| 781 | return err; |
| 782 | } else { |
| 783 | struct mm_struct *mm = current->mm; |
| 784 | long err; |
| 785 | |
| 786 | if (unlikely(!mm)) |
| 787 | return -EFAULT; |
| 788 | |
| 789 | err = slb_allocate_user(mm, ea); |
| 790 | if (!err) |
| 791 | preload_add(current_thread_info(), ea); |
| 792 | |
| 793 | return err; |
| 794 | } |
| 795 | } |
| 796 | |