arch/sparc/include/asm/tsb.h

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