1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * native hashtable management. 4 * 5 * SMP scalability work: 6 * Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM 7 */ 8 9 #undef DEBUG_LOW 10 11 #include <linux/spinlock.h> 12 #include <linux/bitops.h> 13 #include <linux/of.h> 14 #include <linux/processor.h> 15 #include <linux/threads.h> 16 #include <linux/smp.h> 17 #include <linux/pgtable.h> 18 19 #include <asm/machdep.h> 20 #include <asm/mmu.h> 21 #include <asm/mmu_context.h> 22 #include <asm/trace.h> 23 #include <asm/tlb.h> 24 #include <asm/cputable.h> 25 #include <asm/udbg.h> 26 #include <asm/kexec.h> 27 #include <asm/ppc-opcode.h> 28 #include <asm/feature-fixups.h> 29 30 #include <misc/cxl-base.h> 31 32 #ifdef DEBUG_LOW 33 #define DBG_LOW(fmt...) udbg_printf(fmt) 34 #else 35 #define DBG_LOW(fmt...) 36 #endif 37 38 #ifdef __BIG_ENDIAN__ 39 #define HPTE_LOCK_BIT 3 40 #else 41 #define HPTE_LOCK_BIT (56+3) 42 #endif 43 44 static DEFINE_RAW_SPINLOCK(native_tlbie_lock); 45 46 static inline void tlbiel_hash_set_isa206(unsigned int set, unsigned int is) 47 { 48 unsigned long rb; 49 50 rb = (set << PPC_BITLSHIFT(51)) | (is << PPC_BITLSHIFT(53)); 51 52 asm volatile("tlbiel %0" : : "r" (rb)); 53 } 54 55 /* 56 * tlbiel instruction for hash, set invalidation 57 * i.e., r=1 and is=01 or is=10 or is=11 58 */ 59 static __always_inline void tlbiel_hash_set_isa300(unsigned int set, unsigned int is, 60 unsigned int pid, 61 unsigned int ric, unsigned int prs) 62 { 63 unsigned long rb; 64 unsigned long rs; 65 unsigned int r = 0; /* hash format */ 66 67 rb = (set << PPC_BITLSHIFT(51)) | (is << PPC_BITLSHIFT(53)); 68 rs = ((unsigned long)pid << PPC_BITLSHIFT(31)); 69 70 asm volatile(PPC_TLBIEL(%0, %1, %2, %3, %4) 71 : : "r"(rb), "r"(rs), "i"(ric), "i"(prs), "r"(r) 72 : "memory"); 73 } 74 75 76 static void tlbiel_all_isa206(unsigned int num_sets, unsigned int is) 77 { 78 unsigned int set; 79 80 asm volatile("ptesync": : :"memory"); 81 82 for (set = 0; set < num_sets; set++) 83 tlbiel_hash_set_isa206(set, is); 84 85 ppc_after_tlbiel_barrier(); 86 } 87 88 static void tlbiel_all_isa300(unsigned int num_sets, unsigned int is) 89 { 90 unsigned int set; 91 92 asm volatile("ptesync": : :"memory"); 93 94 /* 95 * Flush the first set of the TLB, and any caching of partition table 96 * entries. Then flush the remaining sets of the TLB. Hash mode uses 97 * partition scoped TLB translations. 98 */ 99 tlbiel_hash_set_isa300(0, is, 0, 2, 0); 100 for (set = 1; set < num_sets; set++) 101 tlbiel_hash_set_isa300(set, is, 0, 0, 0); 102 103 /* 104 * Now invalidate the process table cache. 105 * 106 * From ISA v3.0B p. 1078: 107 * The following forms are invalid. 108 * * PRS=1, R=0, and RIC!=2 (The only process-scoped 109 * HPT caching is of the Process Table.) 110 */ 111 tlbiel_hash_set_isa300(0, is, 0, 2, 1); 112 113 ppc_after_tlbiel_barrier(); 114 115 asm volatile(PPC_ISA_3_0_INVALIDATE_ERAT "; isync" : : :"memory"); 116 } 117 118 void hash__tlbiel_all(unsigned int action) 119 { 120 unsigned int is; 121 122 switch (action) { 123 case TLB_INVAL_SCOPE_GLOBAL: 124 is = 3; 125 break; 126 case TLB_INVAL_SCOPE_LPID: 127 is = 2; 128 break; 129 default: 130 BUG(); 131 } 132 133 if (early_cpu_has_feature(CPU_FTR_ARCH_300)) 134 tlbiel_all_isa300(POWER9_TLB_SETS_HASH, is); 135 else if (early_cpu_has_feature(CPU_FTR_ARCH_207S)) 136 tlbiel_all_isa206(POWER8_TLB_SETS, is); 137 else if (early_cpu_has_feature(CPU_FTR_ARCH_206)) 138 tlbiel_all_isa206(POWER7_TLB_SETS, is); 139 else 140 WARN(1, "%s called on pre-POWER7 CPU\n", __func__); 141 } 142 143 static inline unsigned long ___tlbie(unsigned long vpn, int psize, 144 int apsize, int ssize) 145 { 146 unsigned long va; 147 unsigned int penc; 148 unsigned long sllp; 149 150 /* 151 * We need 14 to 65 bits of va for a tlibe of 4K page 152 * With vpn we ignore the lower VPN_SHIFT bits already. 153 * And top two bits are already ignored because we can 154 * only accomodate 76 bits in a 64 bit vpn with a VPN_SHIFT 155 * of 12. 156 */ 157 va = vpn << VPN_SHIFT; 158 /* 159 * clear top 16 bits of 64bit va, non SLS segment 160 * Older versions of the architecture (2.02 and earler) require the 161 * masking of the top 16 bits. 162 */ 163 if (mmu_has_feature(MMU_FTR_TLBIE_CROP_VA)) 164 va &= ~(0xffffULL << 48); 165 166 switch (psize) { 167 case MMU_PAGE_4K: 168 /* clear out bits after (52) [0....52.....63] */ 169 va &= ~((1ul << (64 - 52)) - 1); 170 va |= ssize << 8; 171 sllp = get_sllp_encoding(apsize); 172 va |= sllp << 5; 173 asm volatile(ASM_FTR_IFCLR("tlbie %0,0", PPC_TLBIE(%1,%0), %2) 174 : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206) 175 : "memory"); 176 break; 177 default: 178 /* We need 14 to 14 + i bits of va */ 179 penc = mmu_psize_defs[psize].penc[apsize]; 180 va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1); 181 va |= penc << 12; 182 va |= ssize << 8; 183 /* 184 * AVAL bits: 185 * We don't need all the bits, but rest of the bits 186 * must be ignored by the processor. 187 * vpn cover upto 65 bits of va. (0...65) and we need 188 * 58..64 bits of va. 189 */ 190 va |= (vpn & 0xfe); /* AVAL */ 191 va |= 1; /* L */ 192 asm volatile(ASM_FTR_IFCLR("tlbie %0,1", PPC_TLBIE(%1,%0), %2) 193 : : "r" (va), "r"(0), "i" (CPU_FTR_ARCH_206) 194 : "memory"); 195 break; 196 } 197 return va; 198 } 199 200 static inline void fixup_tlbie_vpn(unsigned long vpn, int psize, 201 int apsize, int ssize) 202 { 203 if (cpu_has_feature(CPU_FTR_P9_TLBIE_ERAT_BUG)) { 204 /* Radix flush for a hash guest */ 205 206 unsigned long rb,rs,prs,r,ric; 207 208 rb = PPC_BIT(52); /* IS = 2 */ 209 rs = 0; /* lpid = 0 */ 210 prs = 0; /* partition scoped */ 211 r = 1; /* radix format */ 212 ric = 0; /* RIC_FLSUH_TLB */ 213 214 /* 215 * Need the extra ptesync to make sure we don't 216 * re-order the tlbie 217 */ 218 asm volatile("ptesync": : :"memory"); 219 asm volatile(PPC_TLBIE_5(%0, %4, %3, %2, %1) 220 : : "r"(rb), "i"(r), "i"(prs), 221 "i"(ric), "r"(rs) : "memory"); 222 } 223 224 225 if (cpu_has_feature(CPU_FTR_P9_TLBIE_STQ_BUG)) { 226 /* Need the extra ptesync to ensure we don't reorder tlbie*/ 227 asm volatile("ptesync": : :"memory"); 228 ___tlbie(vpn, psize, apsize, ssize); 229 } 230 } 231 232 static inline void __tlbie(unsigned long vpn, int psize, int apsize, int ssize) 233 { 234 unsigned long rb; 235 236 rb = ___tlbie(vpn, psize, apsize, ssize); 237 trace_tlbie(0, 0, rb, 0, 0, 0, 0); 238 } 239 240 static inline void __tlbiel(unsigned long vpn, int psize, int apsize, int ssize) 241 { 242 unsigned long va; 243 unsigned int penc; 244 unsigned long sllp; 245 246 /* VPN_SHIFT can be atmost 12 */ 247 va = vpn << VPN_SHIFT; 248 /* 249 * clear top 16 bits of 64 bit va, non SLS segment 250 * Older versions of the architecture (2.02 and earler) require the 251 * masking of the top 16 bits. 252 */ 253 if (mmu_has_feature(MMU_FTR_TLBIE_CROP_VA)) 254 va &= ~(0xffffULL << 48); 255 256 switch (psize) { 257 case MMU_PAGE_4K: 258 /* clear out bits after(52) [0....52.....63] */ 259 va &= ~((1ul << (64 - 52)) - 1); 260 va |= ssize << 8; 261 sllp = get_sllp_encoding(apsize); 262 va |= sllp << 5; 263 asm volatile(ASM_FTR_IFSET("tlbiel %0", "tlbiel %0,0", %1) 264 : : "r" (va), "i" (CPU_FTR_ARCH_206) 265 : "memory"); 266 break; 267 default: 268 /* We need 14 to 14 + i bits of va */ 269 penc = mmu_psize_defs[psize].penc[apsize]; 270 va &= ~((1ul << mmu_psize_defs[apsize].shift) - 1); 271 va |= penc << 12; 272 va |= ssize << 8; 273 /* 274 * AVAL bits: 275 * We don't need all the bits, but rest of the bits 276 * must be ignored by the processor. 277 * vpn cover upto 65 bits of va. (0...65) and we need 278 * 58..64 bits of va. 279 */ 280 va |= (vpn & 0xfe); 281 va |= 1; /* L */ 282 asm volatile(ASM_FTR_IFSET("tlbiel %0", "tlbiel %0,1", %1) 283 : : "r" (va), "i" (CPU_FTR_ARCH_206) 284 : "memory"); 285 break; 286 } 287 trace_tlbie(0, 1, va, 0, 0, 0, 0); 288 289 } 290 291 static inline void tlbie(unsigned long vpn, int psize, int apsize, 292 int ssize, int local) 293 { 294 unsigned int use_local; 295 int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE); 296 297 use_local = local && mmu_has_feature(MMU_FTR_TLBIEL) && !cxl_ctx_in_use(); 298 299 if (use_local) 300 use_local = mmu_psize_defs[psize].tlbiel; 301 if (lock_tlbie && !use_local) 302 raw_spin_lock(&native_tlbie_lock); 303 asm volatile("ptesync": : :"memory"); 304 if (use_local) { 305 __tlbiel(vpn, psize, apsize, ssize); 306 ppc_after_tlbiel_barrier(); 307 } else { 308 __tlbie(vpn, psize, apsize, ssize); 309 fixup_tlbie_vpn(vpn, psize, apsize, ssize); 310 asm volatile("eieio; tlbsync; ptesync": : :"memory"); 311 } 312 if (lock_tlbie && !use_local) 313 raw_spin_unlock(&native_tlbie_lock); 314 } 315 316 static inline void native_lock_hpte(struct hash_pte *hptep) 317 { 318 unsigned long *word = (unsigned long *)&hptep->v; 319 320 while (1) { 321 if (!test_and_set_bit_lock(HPTE_LOCK_BIT, word)) 322 break; 323 spin_begin(); 324 while(test_bit(HPTE_LOCK_BIT, word)) 325 spin_cpu_relax(); 326 spin_end(); 327 } 328 } 329 330 static inline void native_unlock_hpte(struct hash_pte *hptep) 331 { 332 unsigned long *word = (unsigned long *)&hptep->v; 333 334 clear_bit_unlock(HPTE_LOCK_BIT, word); 335 } 336 337 static long native_hpte_insert(unsigned long hpte_group, unsigned long vpn, 338 unsigned long pa, unsigned long rflags, 339 unsigned long vflags, int psize, int apsize, int ssize) 340 { 341 struct hash_pte *hptep = htab_address + hpte_group; 342 unsigned long hpte_v, hpte_r; 343 int i; 344 345 if (!(vflags & HPTE_V_BOLTED)) { 346 DBG_LOW(" insert(group=%lx, vpn=%016lx, pa=%016lx," 347 " rflags=%lx, vflags=%lx, psize=%d)\n", 348 hpte_group, vpn, pa, rflags, vflags, psize); 349 } 350 351 for (i = 0; i < HPTES_PER_GROUP; i++) { 352 if (! (be64_to_cpu(hptep->v) & HPTE_V_VALID)) { 353 /* retry with lock held */ 354 native_lock_hpte(hptep); 355 if (! (be64_to_cpu(hptep->v) & HPTE_V_VALID)) 356 break; 357 native_unlock_hpte(hptep); 358 } 359 360 hptep++; 361 } 362 363 if (i == HPTES_PER_GROUP) 364 return -1; 365 366 hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID; 367 hpte_r = hpte_encode_r(pa, psize, apsize) | rflags; 368 369 if (!(vflags & HPTE_V_BOLTED)) { 370 DBG_LOW(" i=%x hpte_v=%016lx, hpte_r=%016lx\n", 371 i, hpte_v, hpte_r); 372 } 373 374 if (cpu_has_feature(CPU_FTR_ARCH_300)) { 375 hpte_r = hpte_old_to_new_r(hpte_v, hpte_r); 376 hpte_v = hpte_old_to_new_v(hpte_v); 377 } 378 379 hptep->r = cpu_to_be64(hpte_r); 380 /* Guarantee the second dword is visible before the valid bit */ 381 eieio(); 382 /* 383 * Now set the first dword including the valid bit 384 * NOTE: this also unlocks the hpte 385 */ 386 hptep->v = cpu_to_be64(hpte_v); 387 388 __asm__ __volatile__ ("ptesync" : : : "memory"); 389 390 return i | (!!(vflags & HPTE_V_SECONDARY) << 3); 391 } 392 393 static long native_hpte_remove(unsigned long hpte_group) 394 { 395 struct hash_pte *hptep; 396 int i; 397 int slot_offset; 398 unsigned long hpte_v; 399 400 DBG_LOW(" remove(group=%lx)\n", hpte_group); 401 402 /* pick a random entry to start at */ 403 slot_offset = mftb() & 0x7; 404 405 for (i = 0; i < HPTES_PER_GROUP; i++) { 406 hptep = htab_address + hpte_group + slot_offset; 407 hpte_v = be64_to_cpu(hptep->v); 408 409 if ((hpte_v & HPTE_V_VALID) && !(hpte_v & HPTE_V_BOLTED)) { 410 /* retry with lock held */ 411 native_lock_hpte(hptep); 412 hpte_v = be64_to_cpu(hptep->v); 413 if ((hpte_v & HPTE_V_VALID) 414 && !(hpte_v & HPTE_V_BOLTED)) 415 break; 416 native_unlock_hpte(hptep); 417 } 418 419 slot_offset++; 420 slot_offset &= 0x7; 421 } 422 423 if (i == HPTES_PER_GROUP) 424 return -1; 425 426 /* Invalidate the hpte. NOTE: this also unlocks it */ 427 hptep->v = 0; 428 429 return i; 430 } 431 432 static long native_hpte_updatepp(unsigned long slot, unsigned long newpp, 433 unsigned long vpn, int bpsize, 434 int apsize, int ssize, unsigned long flags) 435 { 436 struct hash_pte *hptep = htab_address + slot; 437 unsigned long hpte_v, want_v; 438 int ret = 0, local = 0; 439 440 want_v = hpte_encode_avpn(vpn, bpsize, ssize); 441 442 DBG_LOW(" update(vpn=%016lx, avpnv=%016lx, group=%lx, newpp=%lx)", 443 vpn, want_v & HPTE_V_AVPN, slot, newpp); 444 445 hpte_v = hpte_get_old_v(hptep); 446 /* 447 * We need to invalidate the TLB always because hpte_remove doesn't do 448 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less 449 * random entry from it. When we do that we don't invalidate the TLB 450 * (hpte_remove) because we assume the old translation is still 451 * technically "valid". 452 */ 453 if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID)) { 454 DBG_LOW(" -> miss\n"); 455 ret = -1; 456 } else { 457 native_lock_hpte(hptep); 458 /* recheck with locks held */ 459 hpte_v = hpte_get_old_v(hptep); 460 if (unlikely(!HPTE_V_COMPARE(hpte_v, want_v) || 461 !(hpte_v & HPTE_V_VALID))) { 462 ret = -1; 463 } else { 464 DBG_LOW(" -> hit\n"); 465 /* Update the HPTE */ 466 hptep->r = cpu_to_be64((be64_to_cpu(hptep->r) & 467 ~(HPTE_R_PPP | HPTE_R_N)) | 468 (newpp & (HPTE_R_PPP | HPTE_R_N | 469 HPTE_R_C))); 470 } 471 native_unlock_hpte(hptep); 472 } 473 474 if (flags & HPTE_LOCAL_UPDATE) 475 local = 1; 476 /* 477 * Ensure it is out of the tlb too if it is not a nohpte fault 478 */ 479 if (!(flags & HPTE_NOHPTE_UPDATE)) 480 tlbie(vpn, bpsize, apsize, ssize, local); 481 482 return ret; 483 } 484 485 static long __native_hpte_find(unsigned long want_v, unsigned long slot) 486 { 487 struct hash_pte *hptep; 488 unsigned long hpte_v; 489 unsigned long i; 490 491 for (i = 0; i < HPTES_PER_GROUP; i++) { 492 493 hptep = htab_address + slot; 494 hpte_v = hpte_get_old_v(hptep); 495 if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) 496 /* HPTE matches */ 497 return slot; 498 ++slot; 499 } 500 501 return -1; 502 } 503 504 static long native_hpte_find(unsigned long vpn, int psize, int ssize) 505 { 506 unsigned long hpte_group; 507 unsigned long want_v; 508 unsigned long hash; 509 long slot; 510 511 hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize); 512 want_v = hpte_encode_avpn(vpn, psize, ssize); 513 514 /* 515 * We try to keep bolted entries always in primary hash 516 * But in some case we can find them in secondary too. 517 */ 518 hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP; 519 slot = __native_hpte_find(want_v, hpte_group); 520 if (slot < 0) { 521 /* Try in secondary */ 522 hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP; 523 slot = __native_hpte_find(want_v, hpte_group); 524 if (slot < 0) 525 return -1; 526 } 527 528 return slot; 529 } 530 531 /* 532 * Update the page protection bits. Intended to be used to create 533 * guard pages for kernel data structures on pages which are bolted 534 * in the HPT. Assumes pages being operated on will not be stolen. 535 * 536 * No need to lock here because we should be the only user. 537 */ 538 static void native_hpte_updateboltedpp(unsigned long newpp, unsigned long ea, 539 int psize, int ssize) 540 { 541 unsigned long vpn; 542 unsigned long vsid; 543 long slot; 544 struct hash_pte *hptep; 545 546 vsid = get_kernel_vsid(ea, ssize); 547 vpn = hpt_vpn(ea, vsid, ssize); 548 549 slot = native_hpte_find(vpn, psize, ssize); 550 if (slot == -1) 551 panic("could not find page to bolt\n"); 552 hptep = htab_address + slot; 553 554 /* Update the HPTE */ 555 hptep->r = cpu_to_be64((be64_to_cpu(hptep->r) & 556 ~(HPTE_R_PPP | HPTE_R_N)) | 557 (newpp & (HPTE_R_PPP | HPTE_R_N))); 558 /* 559 * Ensure it is out of the tlb too. Bolted entries base and 560 * actual page size will be same. 561 */ 562 tlbie(vpn, psize, psize, ssize, 0); 563 } 564 565 /* 566 * Remove a bolted kernel entry. Memory hotplug uses this. 567 * 568 * No need to lock here because we should be the only user. 569 */ 570 static int native_hpte_removebolted(unsigned long ea, int psize, int ssize) 571 { 572 unsigned long vpn; 573 unsigned long vsid; 574 long slot; 575 struct hash_pte *hptep; 576 577 vsid = get_kernel_vsid(ea, ssize); 578 vpn = hpt_vpn(ea, vsid, ssize); 579 580 slot = native_hpte_find(vpn, psize, ssize); 581 if (slot == -1) 582 return -ENOENT; 583 584 hptep = htab_address + slot; 585 586 VM_WARN_ON(!(be64_to_cpu(hptep->v) & HPTE_V_BOLTED)); 587 588 /* Invalidate the hpte */ 589 hptep->v = 0; 590 591 /* Invalidate the TLB */ 592 tlbie(vpn, psize, psize, ssize, 0); 593 return 0; 594 } 595 596 597 static void native_hpte_invalidate(unsigned long slot, unsigned long vpn, 598 int bpsize, int apsize, int ssize, int local) 599 { 600 struct hash_pte *hptep = htab_address + slot; 601 unsigned long hpte_v; 602 unsigned long want_v; 603 unsigned long flags; 604 605 local_irq_save(flags); 606 607 DBG_LOW(" invalidate(vpn=%016lx, hash: %lx)\n", vpn, slot); 608 609 want_v = hpte_encode_avpn(vpn, bpsize, ssize); 610 hpte_v = hpte_get_old_v(hptep); 611 612 if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) { 613 native_lock_hpte(hptep); 614 /* recheck with locks held */ 615 hpte_v = hpte_get_old_v(hptep); 616 617 if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) 618 /* Invalidate the hpte. NOTE: this also unlocks it */ 619 hptep->v = 0; 620 else 621 native_unlock_hpte(hptep); 622 } 623 /* 624 * We need to invalidate the TLB always because hpte_remove doesn't do 625 * a tlb invalidate. If a hash bucket gets full, we "evict" a more/less 626 * random entry from it. When we do that we don't invalidate the TLB 627 * (hpte_remove) because we assume the old translation is still 628 * technically "valid". 629 */ 630 tlbie(vpn, bpsize, apsize, ssize, local); 631 632 local_irq_restore(flags); 633 } 634 635 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 636 static void native_hugepage_invalidate(unsigned long vsid, 637 unsigned long addr, 638 unsigned char *hpte_slot_array, 639 int psize, int ssize, int local) 640 { 641 int i; 642 struct hash_pte *hptep; 643 int actual_psize = MMU_PAGE_16M; 644 unsigned int max_hpte_count, valid; 645 unsigned long flags, s_addr = addr; 646 unsigned long hpte_v, want_v, shift; 647 unsigned long hidx, vpn = 0, hash, slot; 648 649 shift = mmu_psize_defs[psize].shift; 650 max_hpte_count = 1U << (PMD_SHIFT - shift); 651 652 local_irq_save(flags); 653 for (i = 0; i < max_hpte_count; i++) { 654 valid = hpte_valid(hpte_slot_array, i); 655 if (!valid) 656 continue; 657 hidx = hpte_hash_index(hpte_slot_array, i); 658 659 /* get the vpn */ 660 addr = s_addr + (i * (1ul << shift)); 661 vpn = hpt_vpn(addr, vsid, ssize); 662 hash = hpt_hash(vpn, shift, ssize); 663 if (hidx & _PTEIDX_SECONDARY) 664 hash = ~hash; 665 666 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP; 667 slot += hidx & _PTEIDX_GROUP_IX; 668 669 hptep = htab_address + slot; 670 want_v = hpte_encode_avpn(vpn, psize, ssize); 671 hpte_v = hpte_get_old_v(hptep); 672 673 /* Even if we miss, we need to invalidate the TLB */ 674 if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) { 675 /* recheck with locks held */ 676 native_lock_hpte(hptep); 677 hpte_v = hpte_get_old_v(hptep); 678 679 if (HPTE_V_COMPARE(hpte_v, want_v) && (hpte_v & HPTE_V_VALID)) { 680 /* 681 * Invalidate the hpte. NOTE: this also unlocks it 682 */ 683 684 hptep->v = 0; 685 } else 686 native_unlock_hpte(hptep); 687 } 688 /* 689 * We need to do tlb invalidate for all the address, tlbie 690 * instruction compares entry_VA in tlb with the VA specified 691 * here 692 */ 693 tlbie(vpn, psize, actual_psize, ssize, local); 694 } 695 local_irq_restore(flags); 696 } 697 #else 698 static void native_hugepage_invalidate(unsigned long vsid, 699 unsigned long addr, 700 unsigned char *hpte_slot_array, 701 int psize, int ssize, int local) 702 { 703 WARN(1, "%s called without THP support\n", __func__); 704 } 705 #endif 706 707 static void hpte_decode(struct hash_pte *hpte, unsigned long slot, 708 int *psize, int *apsize, int *ssize, unsigned long *vpn) 709 { 710 unsigned long avpn, pteg, vpi; 711 unsigned long hpte_v = be64_to_cpu(hpte->v); 712 unsigned long hpte_r = be64_to_cpu(hpte->r); 713 unsigned long vsid, seg_off; 714 int size, a_size, shift; 715 /* Look at the 8 bit LP value */ 716 unsigned int lp = (hpte_r >> LP_SHIFT) & ((1 << LP_BITS) - 1); 717 718 if (cpu_has_feature(CPU_FTR_ARCH_300)) { 719 hpte_v = hpte_new_to_old_v(hpte_v, hpte_r); 720 hpte_r = hpte_new_to_old_r(hpte_r); 721 } 722 if (!(hpte_v & HPTE_V_LARGE)) { 723 size = MMU_PAGE_4K; 724 a_size = MMU_PAGE_4K; 725 } else { 726 size = hpte_page_sizes[lp] & 0xf; 727 a_size = hpte_page_sizes[lp] >> 4; 728 } 729 /* This works for all page sizes, and for 256M and 1T segments */ 730 *ssize = hpte_v >> HPTE_V_SSIZE_SHIFT; 731 shift = mmu_psize_defs[size].shift; 732 733 avpn = (HPTE_V_AVPN_VAL(hpte_v) & ~mmu_psize_defs[size].avpnm); 734 pteg = slot / HPTES_PER_GROUP; 735 if (hpte_v & HPTE_V_SECONDARY) 736 pteg = ~pteg; 737 738 switch (*ssize) { 739 case MMU_SEGSIZE_256M: 740 /* We only have 28 - 23 bits of seg_off in avpn */ 741 seg_off = (avpn & 0x1f) << 23; 742 vsid = avpn >> 5; 743 /* We can find more bits from the pteg value */ 744 if (shift < 23) { 745 vpi = (vsid ^ pteg) & htab_hash_mask; 746 seg_off |= vpi << shift; 747 } 748 *vpn = vsid << (SID_SHIFT - VPN_SHIFT) | seg_off >> VPN_SHIFT; 749 break; 750 case MMU_SEGSIZE_1T: 751 /* We only have 40 - 23 bits of seg_off in avpn */ 752 seg_off = (avpn & 0x1ffff) << 23; 753 vsid = avpn >> 17; 754 if (shift < 23) { 755 vpi = (vsid ^ (vsid << 25) ^ pteg) & htab_hash_mask; 756 seg_off |= vpi << shift; 757 } 758 *vpn = vsid << (SID_SHIFT_1T - VPN_SHIFT) | seg_off >> VPN_SHIFT; 759 break; 760 default: 761 *vpn = size = 0; 762 } 763 *psize = size; 764 *apsize = a_size; 765 } 766 767 /* 768 * clear all mappings on kexec. All cpus are in real mode (or they will 769 * be when they isi), and we are the only one left. We rely on our kernel 770 * mapping being 0xC0's and the hardware ignoring those two real bits. 771 * 772 * This must be called with interrupts disabled. 773 * 774 * Taking the native_tlbie_lock is unsafe here due to the possibility of 775 * lockdep being on. On pre POWER5 hardware, not taking the lock could 776 * cause deadlock. POWER5 and newer not taking the lock is fine. This only 777 * gets called during boot before secondary CPUs have come up and during 778 * crashdump and all bets are off anyway. 779 * 780 * TODO: add batching support when enabled. remember, no dynamic memory here, 781 * although there is the control page available... 782 */ 783 static void native_hpte_clear(void) 784 { 785 unsigned long vpn = 0; 786 unsigned long slot, slots; 787 struct hash_pte *hptep = htab_address; 788 unsigned long hpte_v; 789 unsigned long pteg_count; 790 int psize, apsize, ssize; 791 792 pteg_count = htab_hash_mask + 1; 793 794 slots = pteg_count * HPTES_PER_GROUP; 795 796 for (slot = 0; slot < slots; slot++, hptep++) { 797 /* 798 * we could lock the pte here, but we are the only cpu 799 * running, right? and for crash dump, we probably 800 * don't want to wait for a maybe bad cpu. 801 */ 802 hpte_v = be64_to_cpu(hptep->v); 803 804 /* 805 * Call __tlbie() here rather than tlbie() since we can't take the 806 * native_tlbie_lock. 807 */ 808 if (hpte_v & HPTE_V_VALID) { 809 hpte_decode(hptep, slot, &psize, &apsize, &ssize, &vpn); 810 hptep->v = 0; 811 ___tlbie(vpn, psize, apsize, ssize); 812 } 813 } 814 815 asm volatile("eieio; tlbsync; ptesync":::"memory"); 816 } 817 818 /* 819 * Batched hash table flush, we batch the tlbie's to avoid taking/releasing 820 * the lock all the time 821 */ 822 static void native_flush_hash_range(unsigned long number, int local) 823 { 824 unsigned long vpn = 0; 825 unsigned long hash, index, hidx, shift, slot; 826 struct hash_pte *hptep; 827 unsigned long hpte_v; 828 unsigned long want_v; 829 unsigned long flags; 830 real_pte_t pte; 831 struct ppc64_tlb_batch *batch = this_cpu_ptr(&ppc64_tlb_batch); 832 unsigned long psize = batch->psize; 833 int ssize = batch->ssize; 834 int i; 835 unsigned int use_local; 836 837 use_local = local && mmu_has_feature(MMU_FTR_TLBIEL) && 838 mmu_psize_defs[psize].tlbiel && !cxl_ctx_in_use(); 839 840 local_irq_save(flags); 841 842 for (i = 0; i < number; i++) { 843 vpn = batch->vpn[i]; 844 pte = batch->pte[i]; 845 846 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) { 847 hash = hpt_hash(vpn, shift, ssize); 848 hidx = __rpte_to_hidx(pte, index); 849 if (hidx & _PTEIDX_SECONDARY) 850 hash = ~hash; 851 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP; 852 slot += hidx & _PTEIDX_GROUP_IX; 853 hptep = htab_address + slot; 854 want_v = hpte_encode_avpn(vpn, psize, ssize); 855 hpte_v = hpte_get_old_v(hptep); 856 857 if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID)) 858 continue; 859 /* lock and try again */ 860 native_lock_hpte(hptep); 861 hpte_v = hpte_get_old_v(hptep); 862 863 if (!HPTE_V_COMPARE(hpte_v, want_v) || !(hpte_v & HPTE_V_VALID)) 864 native_unlock_hpte(hptep); 865 else 866 hptep->v = 0; 867 868 } pte_iterate_hashed_end(); 869 } 870 871 if (use_local) { 872 asm volatile("ptesync":::"memory"); 873 for (i = 0; i < number; i++) { 874 vpn = batch->vpn[i]; 875 pte = batch->pte[i]; 876 877 pte_iterate_hashed_subpages(pte, psize, 878 vpn, index, shift) { 879 __tlbiel(vpn, psize, psize, ssize); 880 } pte_iterate_hashed_end(); 881 } 882 ppc_after_tlbiel_barrier(); 883 } else { 884 int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE); 885 886 if (lock_tlbie) 887 raw_spin_lock(&native_tlbie_lock); 888 889 asm volatile("ptesync":::"memory"); 890 for (i = 0; i < number; i++) { 891 vpn = batch->vpn[i]; 892 pte = batch->pte[i]; 893 894 pte_iterate_hashed_subpages(pte, psize, 895 vpn, index, shift) { 896 __tlbie(vpn, psize, psize, ssize); 897 } pte_iterate_hashed_end(); 898 } 899 /* 900 * Just do one more with the last used values. 901 */ 902 fixup_tlbie_vpn(vpn, psize, psize, ssize); 903 asm volatile("eieio; tlbsync; ptesync":::"memory"); 904 905 if (lock_tlbie) 906 raw_spin_unlock(&native_tlbie_lock); 907 } 908 909 local_irq_restore(flags); 910 } 911 912 void __init hpte_init_native(void) 913 { 914 mmu_hash_ops.hpte_invalidate = native_hpte_invalidate; 915 mmu_hash_ops.hpte_updatepp = native_hpte_updatepp; 916 mmu_hash_ops.hpte_updateboltedpp = native_hpte_updateboltedpp; 917 mmu_hash_ops.hpte_removebolted = native_hpte_removebolted; 918 mmu_hash_ops.hpte_insert = native_hpte_insert; 919 mmu_hash_ops.hpte_remove = native_hpte_remove; 920 mmu_hash_ops.hpte_clear_all = native_hpte_clear; 921 mmu_hash_ops.flush_hash_range = native_flush_hash_range; 922 mmu_hash_ops.hugepage_invalidate = native_hugepage_invalidate; 923 } 924