1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef _SPARC64_TSB_H 3 #define _SPARC64_TSB_H 4 5 /* The sparc64 TSB is similar to the powerpc hashtables. It's a 6 * power-of-2 sized table of TAG/PTE pairs. The cpu precomputes 7 * pointers into this table for 8K and 64K page sizes, and also a 8 * comparison TAG based upon the virtual address and context which 9 * faults. 10 * 11 * TLB miss trap handler software does the actual lookup via something 12 * of the form: 13 * 14 * ldxa [%g0] ASI_{D,I}MMU_TSB_8KB_PTR, %g1 15 * ldxa [%g0] ASI_{D,I}MMU, %g6 16 * sllx %g6, 22, %g6 17 * srlx %g6, 22, %g6 18 * ldda [%g1] ASI_NUCLEUS_QUAD_LDD, %g4 19 * cmp %g4, %g6 20 * bne,pn %xcc, tsb_miss_{d,i}tlb 21 * mov FAULT_CODE_{D,I}TLB, %g3 22 * stxa %g5, [%g0] ASI_{D,I}TLB_DATA_IN 23 * retry 24 * 25 * 26 * Each 16-byte slot of the TSB is the 8-byte tag and then the 8-byte 27 * PTE. The TAG is of the same layout as the TLB TAG TARGET mmu 28 * register which is: 29 * 30 * ------------------------------------------------- 31 * | - | CONTEXT | - | VADDR bits 63:22 | 32 * ------------------------------------------------- 33 * 63 61 60 48 47 42 41 0 34 * 35 * But actually, since we use per-mm TSB's, we zero out the CONTEXT 36 * field. 37 * 38 * Like the powerpc hashtables we need to use locking in order to 39 * synchronize while we update the entries. PTE updates need locking 40 * as well. 41 * 42 * We need to carefully choose a lock bits for the TSB entry. We 43 * choose to use bit 47 in the tag. Also, since we never map anything 44 * at page zero in context zero, we use zero as an invalid tag entry. 45 * When the lock bit is set, this forces a tag comparison failure. 46 */ 47 48 #define TSB_TAG_LOCK_BIT 47 49 #define TSB_TAG_LOCK_HIGH (1 << (TSB_TAG_LOCK_BIT - 32)) 50 51 #define TSB_TAG_INVALID_BIT 46 52 #define TSB_TAG_INVALID_HIGH (1 << (TSB_TAG_INVALID_BIT - 32)) 53 54 /* Some cpus support physical address quad loads. We want to use 55 * those if possible so we don't need to hard-lock the TSB mapping 56 * into the TLB. We encode some instruction patching in order to 57 * support this. 58 * 59 * The kernel TSB is locked into the TLB by virtue of being in the 60 * kernel image, so we don't play these games for swapper_tsb access. 61 */ 62 #ifndef __ASSEMBLY__ 63 struct tsb_ldquad_phys_patch_entry { 64 unsigned int addr; 65 unsigned int sun4u_insn; 66 unsigned int sun4v_insn; 67 }; 68 extern struct tsb_ldquad_phys_patch_entry __tsb_ldquad_phys_patch, 69 __tsb_ldquad_phys_patch_end; 70 71 struct tsb_phys_patch_entry { 72 unsigned int addr; 73 unsigned int insn; 74 }; 75 extern struct tsb_phys_patch_entry __tsb_phys_patch, __tsb_phys_patch_end; 76 #endif 77 #define TSB_LOAD_QUAD(TSB, REG) \ 78 661: ldda [TSB] ASI_NUCLEUS_QUAD_LDD, REG; \ 79 .section .tsb_ldquad_phys_patch, "ax"; \ 80 .word 661b; \ 81 ldda [TSB] ASI_QUAD_LDD_PHYS, REG; \ 82 ldda [TSB] ASI_QUAD_LDD_PHYS_4V, REG; \ 83 .previous 84 85 #define TSB_LOAD_TAG_HIGH(TSB, REG) \ 86 661: lduwa [TSB] ASI_N, REG; \ 87 .section .tsb_phys_patch, "ax"; \ 88 .word 661b; \ 89 lduwa [TSB] ASI_PHYS_USE_EC, REG; \ 90 .previous 91 92 #define TSB_LOAD_TAG(TSB, REG) \ 93 661: ldxa [TSB] ASI_N, REG; \ 94 .section .tsb_phys_patch, "ax"; \ 95 .word 661b; \ 96 ldxa [TSB] ASI_PHYS_USE_EC, REG; \ 97 .previous 98 99 #define TSB_CAS_TAG_HIGH(TSB, REG1, REG2) \ 100 661: casa [TSB] ASI_N, REG1, REG2; \ 101 .section .tsb_phys_patch, "ax"; \ 102 .word 661b; \ 103 casa [TSB] ASI_PHYS_USE_EC, REG1, REG2; \ 104 .previous 105 106 #define TSB_CAS_TAG(TSB, REG1, REG2) \ 107 661: casxa [TSB] ASI_N, REG1, REG2; \ 108 .section .tsb_phys_patch, "ax"; \ 109 .word 661b; \ 110 casxa [TSB] ASI_PHYS_USE_EC, REG1, REG2; \ 111 .previous 112 113 #define TSB_STORE(ADDR, VAL) \ 114 661: stxa VAL, [ADDR] ASI_N; \ 115 .section .tsb_phys_patch, "ax"; \ 116 .word 661b; \ 117 stxa VAL, [ADDR] ASI_PHYS_USE_EC; \ 118 .previous 119 120 #define TSB_LOCK_TAG(TSB, REG1, REG2) \ 121 99: TSB_LOAD_TAG_HIGH(TSB, REG1); \ 122 sethi %hi(TSB_TAG_LOCK_HIGH), REG2;\ 123 andcc REG1, REG2, %g0; \ 124 bne,pn %icc, 99b; \ 125 nop; \ 126 TSB_CAS_TAG_HIGH(TSB, REG1, REG2); \ 127 cmp REG1, REG2; \ 128 bne,pn %icc, 99b; \ 129 nop; \ 130 131 #define TSB_WRITE(TSB, TTE, TAG) \ 132 add TSB, 0x8, TSB; \ 133 TSB_STORE(TSB, TTE); \ 134 sub TSB, 0x8, TSB; \ 135 TSB_STORE(TSB, TAG); 136 137 /* Do a kernel page table walk. Leaves valid PTE value in 138 * REG1. Jumps to FAIL_LABEL on early page table walk 139 * termination. VADDR will not be clobbered, but REG2 will. 140 * 141 * There are two masks we must apply to propagate bits from 142 * the virtual address into the PTE physical address field 143 * when dealing with huge pages. This is because the page 144 * table boundaries do not match the huge page size(s) the 145 * hardware supports. 146 * 147 * In these cases we propagate the bits that are below the 148 * page table level where we saw the huge page mapping, but 149 * are still within the relevant physical bits for the huge 150 * page size in question. So for PMD mappings (which fall on 151 * bit 23, for 8MB per PMD) we must propagate bit 22 for a 152 * 4MB huge page. For huge PUDs (which fall on bit 33, for 153 * 8GB per PUD), we have to accommodate 256MB and 2GB huge 154 * pages. So for those we propagate bits 32 to 28. 155 */ 156 #define KERN_PGTABLE_WALK(VADDR, REG1, REG2, FAIL_LABEL) \ 157 sethi %hi(swapper_pg_dir), REG1; \ 158 or REG1, %lo(swapper_pg_dir), REG1; \ 159 sllx VADDR, 64 - (PGDIR_SHIFT + PGDIR_BITS), REG2; \ 160 srlx REG2, 64 - PAGE_SHIFT, REG2; \ 161 andn REG2, 0x7, REG2; \ 162 ldx [REG1 + REG2], REG1; \ 163 brz,pn REG1, FAIL_LABEL; \ 164 sllx VADDR, 64 - (PUD_SHIFT + PUD_BITS), REG2; \ 165 srlx REG2, 64 - PAGE_SHIFT, REG2; \ 166 andn REG2, 0x7, REG2; \ 167 ldxa [REG1 + REG2] ASI_PHYS_USE_EC, REG1; \ 168 brz,pn REG1, FAIL_LABEL; \ 169 sethi %uhi(_PAGE_PUD_HUGE), REG2; \ 170 brz,pn REG1, FAIL_LABEL; \ 171 sllx REG2, 32, REG2; \ 172 andcc REG1, REG2, %g0; \ 173 sethi %hi(0xf8000000), REG2; \ 174 bne,pt %xcc, 697f; \ 175 sllx REG2, 1, REG2; \ 176 sllx VADDR, 64 - (PMD_SHIFT + PMD_BITS), REG2; \ 177 srlx REG2, 64 - PAGE_SHIFT, REG2; \ 178 andn REG2, 0x7, REG2; \ 179 ldxa [REG1 + REG2] ASI_PHYS_USE_EC, REG1; \ 180 sethi %uhi(_PAGE_PMD_HUGE), REG2; \ 181 brz,pn REG1, FAIL_LABEL; \ 182 sllx REG2, 32, REG2; \ 183 andcc REG1, REG2, %g0; \ 184 be,pn %xcc, 698f; \ 185 sethi %hi(0x400000), REG2; \ 186 697: brgez,pn REG1, FAIL_LABEL; \ 187 andn REG1, REG2, REG1; \ 188 and VADDR, REG2, REG2; \ 189 ba,pt %xcc, 699f; \ 190 or REG1, REG2, REG1; \ 191 698: sllx VADDR, 64 - PMD_SHIFT, REG2; \ 192 srlx REG2, 64 - PAGE_SHIFT, REG2; \ 193 andn REG2, 0x7, REG2; \ 194 ldxa [REG1 + REG2] ASI_PHYS_USE_EC, REG1; \ 195 brgez,pn REG1, FAIL_LABEL; \ 196 nop; \ 197 699: 198 199 /* PUD has been loaded into REG1, interpret the value, seeing 200 * if it is a HUGE PUD or a normal one. If it is not valid 201 * then jump to FAIL_LABEL. If it is a HUGE PUD, and it 202 * translates to a valid PTE, branch to PTE_LABEL. 203 * 204 * We have to propagate bits [32:22] from the virtual address 205 * to resolve at 4M granularity. 206 */ 207 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE) 208 #define USER_PGTABLE_CHECK_PUD_HUGE(VADDR, REG1, REG2, FAIL_LABEL, PTE_LABEL) \ 209 700: ba 700f; \ 210 nop; \ 211 .section .pud_huge_patch, "ax"; \ 212 .word 700b; \ 213 nop; \ 214 .previous; \ 215 brz,pn REG1, FAIL_LABEL; \ 216 sethi %uhi(_PAGE_PUD_HUGE), REG2; \ 217 sllx REG2, 32, REG2; \ 218 andcc REG1, REG2, %g0; \ 219 be,pt %xcc, 700f; \ 220 sethi %hi(0xffe00000), REG2; \ 221 sllx REG2, 1, REG2; \ 222 brgez,pn REG1, FAIL_LABEL; \ 223 andn REG1, REG2, REG1; \ 224 and VADDR, REG2, REG2; \ 225 brlz,pt REG1, PTE_LABEL; \ 226 or REG1, REG2, REG1; \ 227 700: 228 #else 229 #define USER_PGTABLE_CHECK_PUD_HUGE(VADDR, REG1, REG2, FAIL_LABEL, PTE_LABEL) \ 230 brz,pn REG1, FAIL_LABEL; \ 231 nop; 232 #endif 233 234 /* PMD has been loaded into REG1, interpret the value, seeing 235 * if it is a HUGE PMD or a normal one. If it is not valid 236 * then jump to FAIL_LABEL. If it is a HUGE PMD, and it 237 * translates to a valid PTE, branch to PTE_LABEL. 238 * 239 * We have to propagate the 4MB bit of the virtual address 240 * because we are fabricating 8MB pages using 4MB hw pages. 241 */ 242 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE) 243 #define USER_PGTABLE_CHECK_PMD_HUGE(VADDR, REG1, REG2, FAIL_LABEL, PTE_LABEL) \ 244 brz,pn REG1, FAIL_LABEL; \ 245 sethi %uhi(_PAGE_PMD_HUGE), REG2; \ 246 sllx REG2, 32, REG2; \ 247 andcc REG1, REG2, %g0; \ 248 be,pt %xcc, 700f; \ 249 sethi %hi(4 * 1024 * 1024), REG2; \ 250 brgez,pn REG1, FAIL_LABEL; \ 251 andn REG1, REG2, REG1; \ 252 and VADDR, REG2, REG2; \ 253 brlz,pt REG1, PTE_LABEL; \ 254 or REG1, REG2, REG1; \ 255 700: 256 #else 257 #define USER_PGTABLE_CHECK_PMD_HUGE(VADDR, REG1, REG2, FAIL_LABEL, PTE_LABEL) \ 258 brz,pn REG1, FAIL_LABEL; \ 259 nop; 260 #endif 261 262 /* Do a user page table walk in MMU globals. Leaves final, 263 * valid, PTE value in REG1. Jumps to FAIL_LABEL on early 264 * page table walk termination or if the PTE is not valid. 265 * 266 * Physical base of page tables is in PHYS_PGD which will not 267 * be modified. 268 * 269 * VADDR will not be clobbered, but REG1 and REG2 will. 270 */ 271 #define USER_PGTABLE_WALK_TL1(VADDR, PHYS_PGD, REG1, REG2, FAIL_LABEL) \ 272 sllx VADDR, 64 - (PGDIR_SHIFT + PGDIR_BITS), REG2; \ 273 srlx REG2, 64 - PAGE_SHIFT, REG2; \ 274 andn REG2, 0x7, REG2; \ 275 ldxa [PHYS_PGD + REG2] ASI_PHYS_USE_EC, REG1; \ 276 brz,pn REG1, FAIL_LABEL; \ 277 sllx VADDR, 64 - (PUD_SHIFT + PUD_BITS), REG2; \ 278 srlx REG2, 64 - PAGE_SHIFT, REG2; \ 279 andn REG2, 0x7, REG2; \ 280 ldxa [REG1 + REG2] ASI_PHYS_USE_EC, REG1; \ 281 USER_PGTABLE_CHECK_PUD_HUGE(VADDR, REG1, REG2, FAIL_LABEL, 800f) \ 282 brz,pn REG1, FAIL_LABEL; \ 283 sllx VADDR, 64 - (PMD_SHIFT + PMD_BITS), REG2; \ 284 srlx REG2, 64 - PAGE_SHIFT, REG2; \ 285 andn REG2, 0x7, REG2; \ 286 ldxa [REG1 + REG2] ASI_PHYS_USE_EC, REG1; \ 287 USER_PGTABLE_CHECK_PMD_HUGE(VADDR, REG1, REG2, FAIL_LABEL, 800f) \ 288 sllx VADDR, 64 - PMD_SHIFT, REG2; \ 289 srlx REG2, 64 - PAGE_SHIFT, REG2; \ 290 andn REG2, 0x7, REG2; \ 291 add REG1, REG2, REG1; \ 292 ldxa [REG1] ASI_PHYS_USE_EC, REG1; \ 293 brgez,pn REG1, FAIL_LABEL; \ 294 nop; \ 295 800: 296 297 /* Lookup a OBP mapping on VADDR in the prom_trans[] table at TL>0. 298 * If no entry is found, FAIL_LABEL will be branched to. On success 299 * the resulting PTE value will be left in REG1. VADDR is preserved 300 * by this routine. 301 */ 302 #define OBP_TRANS_LOOKUP(VADDR, REG1, REG2, REG3, FAIL_LABEL) \ 303 sethi %hi(prom_trans), REG1; \ 304 or REG1, %lo(prom_trans), REG1; \ 305 97: ldx [REG1 + 0x00], REG2; \ 306 brz,pn REG2, FAIL_LABEL; \ 307 nop; \ 308 ldx [REG1 + 0x08], REG3; \ 309 add REG2, REG3, REG3; \ 310 cmp REG2, VADDR; \ 311 bgu,pt %xcc, 98f; \ 312 cmp VADDR, REG3; \ 313 bgeu,pt %xcc, 98f; \ 314 ldx [REG1 + 0x10], REG3; \ 315 sub VADDR, REG2, REG2; \ 316 ba,pt %xcc, 99f; \ 317 add REG3, REG2, REG1; \ 318 98: ba,pt %xcc, 97b; \ 319 add REG1, (3 * 8), REG1; \ 320 99: 321 322 /* We use a 32K TSB for the whole kernel, this allows to 323 * handle about 16MB of modules and vmalloc mappings without 324 * incurring many hash conflicts. 325 */ 326 #define KERNEL_TSB_SIZE_BYTES (32 * 1024) 327 #define KERNEL_TSB_NENTRIES \ 328 (KERNEL_TSB_SIZE_BYTES / 16) 329 #define KERNEL_TSB4M_NENTRIES 4096 330 331 /* Do a kernel TSB lookup at tl>0 on VADDR+TAG, branch to OK_LABEL 332 * on TSB hit. REG1, REG2, REG3, and REG4 are used as temporaries 333 * and the found TTE will be left in REG1. REG3 and REG4 must 334 * be an even/odd pair of registers. 335 * 336 * VADDR and TAG will be preserved and not clobbered by this macro. 337 */ 338 #define KERN_TSB_LOOKUP_TL1(VADDR, TAG, REG1, REG2, REG3, REG4, OK_LABEL) \ 339 661: sethi %uhi(swapper_tsb), REG1; \ 340 sethi %hi(swapper_tsb), REG2; \ 341 or REG1, %ulo(swapper_tsb), REG1; \ 342 or REG2, %lo(swapper_tsb), REG2; \ 343 .section .swapper_tsb_phys_patch, "ax"; \ 344 .word 661b; \ 345 .previous; \ 346 sllx REG1, 32, REG1; \ 347 or REG1, REG2, REG1; \ 348 srlx VADDR, PAGE_SHIFT, REG2; \ 349 and REG2, (KERNEL_TSB_NENTRIES - 1), REG2; \ 350 sllx REG2, 4, REG2; \ 351 add REG1, REG2, REG2; \ 352 TSB_LOAD_QUAD(REG2, REG3); \ 353 cmp REG3, TAG; \ 354 be,a,pt %xcc, OK_LABEL; \ 355 mov REG4, REG1; 356 357 #ifndef CONFIG_DEBUG_PAGEALLOC 358 /* This version uses a trick, the TAG is already (VADDR >> 22) so 359 * we can make use of that for the index computation. 360 */ 361 #define KERN_TSB4M_LOOKUP_TL1(TAG, REG1, REG2, REG3, REG4, OK_LABEL) \ 362 661: sethi %uhi(swapper_4m_tsb), REG1; \ 363 sethi %hi(swapper_4m_tsb), REG2; \ 364 or REG1, %ulo(swapper_4m_tsb), REG1; \ 365 or REG2, %lo(swapper_4m_tsb), REG2; \ 366 .section .swapper_4m_tsb_phys_patch, "ax"; \ 367 .word 661b; \ 368 .previous; \ 369 sllx REG1, 32, REG1; \ 370 or REG1, REG2, REG1; \ 371 and TAG, (KERNEL_TSB4M_NENTRIES - 1), REG2; \ 372 sllx REG2, 4, REG2; \ 373 add REG1, REG2, REG2; \ 374 TSB_LOAD_QUAD(REG2, REG3); \ 375 cmp REG3, TAG; \ 376 be,a,pt %xcc, OK_LABEL; \ 377 mov REG4, REG1; 378 #endif 379 380 #endif /* !(_SPARC64_TSB_H) */ 381