*
* ldxa [%g0] ASI_{D,I}MMU_TSB_8KB_PTR, %g1
* ldxa [%g0] ASI_{D,I}MMU, %g6
+ * sllx %g6, 22, %g6
+ * srlx %g6, 22, %g6
* ldda [%g1] ASI_NUCLEUS_QUAD_LDD, %g4
* cmp %g4, %g6
* bne,pn %xcc, tsb_miss_{d,i}tlb
* -------------------------------------------------
* 63 61 60 48 47 42 41 0
*
+ * But actually, since we use per-mm TSB's, we zero out the CONTEXT
+ * field.
+ *
* Like the powerpc hashtables we need to use locking in order to
* synchronize while we update the entries. PTE updates need locking
* as well.
* choose to use bit 47 in the tag. Also, since we never map anything
* at page zero in context zero, we use zero as an invalid tag entry.
* When the lock bit is set, this forces a tag comparison failure.
- *
- * Currently, we allocate an 8K TSB per-process and we use it for both
- * I-TLB and D-TLB misses. Perhaps at some point we'll add code that
- * monitors the number of active pages in the process as we get
- * major/minor faults, and grow the TSB in response. The only trick
- * in implementing that is synchronizing the freeing of the old TSB
- * wrt. parallel TSB updates occuring on other processors. On
- * possible solution is to use RCU for the freeing of the TSB.
*/
#define TSB_TAG_LOCK_BIT 47
#define TSB_TAG_LOCK_HIGH (1 << (TSB_TAG_LOCK_BIT - 32))
+#define TSB_TAG_INVALID_BIT 46
+#define TSB_TAG_INVALID_HIGH (1 << (TSB_TAG_INVALID_BIT - 32))
+
#define TSB_MEMBAR membar #StoreStore
+/* Some cpus support physical address quad loads. We want to use
+ * those if possible so we don't need to hard-lock the TSB mapping
+ * into the TLB. We encode some instruction patching in order to
+ * support this.
+ *
+ * The kernel TSB is locked into the TLB by virtue of being in the
+ * kernel image, so we don't play these games for swapper_tsb access.
+ */
+#ifndef __ASSEMBLY__
+struct tsb_ldquad_phys_patch_entry {
+ unsigned int addr;
+ unsigned int sun4u_insn;
+ unsigned int sun4v_insn;
+};
+extern struct tsb_ldquad_phys_patch_entry __tsb_ldquad_phys_patch,
+ __tsb_ldquad_phys_patch_end;
+
+struct tsb_phys_patch_entry {
+ unsigned int addr;
+ unsigned int insn;
+};
+extern struct tsb_phys_patch_entry __tsb_phys_patch, __tsb_phys_patch_end;
+#endif
+#define TSB_LOAD_QUAD(TSB, REG) \
+661: ldda [TSB] ASI_NUCLEUS_QUAD_LDD, REG; \
+ .section .tsb_ldquad_phys_patch, "ax"; \
+ .word 661b; \
+ ldda [TSB] ASI_QUAD_LDD_PHYS, REG; \
+ ldda [TSB] ASI_QUAD_LDD_PHYS_4V, REG; \
+ .previous
+
+#define TSB_LOAD_TAG_HIGH(TSB, REG) \
+661: lduwa [TSB] ASI_N, REG; \
+ .section .tsb_phys_patch, "ax"; \
+ .word 661b; \
+ lduwa [TSB] ASI_PHYS_USE_EC, REG; \
+ .previous
+
+#define TSB_LOAD_TAG(TSB, REG) \
+661: ldxa [TSB] ASI_N, REG; \
+ .section .tsb_phys_patch, "ax"; \
+ .word 661b; \
+ ldxa [TSB] ASI_PHYS_USE_EC, REG; \
+ .previous
+
+#define TSB_CAS_TAG_HIGH(TSB, REG1, REG2) \
+661: casa [TSB] ASI_N, REG1, REG2; \
+ .section .tsb_phys_patch, "ax"; \
+ .word 661b; \
+ casa [TSB] ASI_PHYS_USE_EC, REG1, REG2; \
+ .previous
+
+#define TSB_CAS_TAG(TSB, REG1, REG2) \
+661: casxa [TSB] ASI_N, REG1, REG2; \
+ .section .tsb_phys_patch, "ax"; \
+ .word 661b; \
+ casxa [TSB] ASI_PHYS_USE_EC, REG1, REG2; \
+ .previous
+
+#define TSB_STORE(ADDR, VAL) \
+661: stxa VAL, [ADDR] ASI_N; \
+ .section .tsb_phys_patch, "ax"; \
+ .word 661b; \
+ stxa VAL, [ADDR] ASI_PHYS_USE_EC; \
+ .previous
+
#define TSB_LOCK_TAG(TSB, REG1, REG2) \
+99: TSB_LOAD_TAG_HIGH(TSB, REG1); \
+ sethi %hi(TSB_TAG_LOCK_HIGH), REG2;\
+ andcc REG1, REG2, %g0; \
+ bne,pn %icc, 99b; \
+ nop; \
+ TSB_CAS_TAG_HIGH(TSB, REG1, REG2); \
+ cmp REG1, REG2; \
+ bne,pn %icc, 99b; \
+ nop; \
+ TSB_MEMBAR
+
+#define TSB_WRITE(TSB, TTE, TAG) \
+ add TSB, 0x8, TSB; \
+ TSB_STORE(TSB, TTE); \
+ sub TSB, 0x8, TSB; \
+ TSB_MEMBAR; \
+ TSB_STORE(TSB, TAG);
+
+#define KTSB_LOAD_QUAD(TSB, REG) \
+ ldda [TSB] ASI_NUCLEUS_QUAD_LDD, REG;
+
+#define KTSB_STORE(ADDR, VAL) \
+ stxa VAL, [ADDR] ASI_N;
+
+#define KTSB_LOCK_TAG(TSB, REG1, REG2) \
99: lduwa [TSB] ASI_N, REG1; \
sethi %hi(TSB_TAG_LOCK_HIGH), REG2;\
andcc REG1, REG2, %g0; \
nop; \
TSB_MEMBAR
-#define TSB_WRITE(TSB, TTE, TAG) \
- stx TTE, [TSB + 0x08]; \
- TSB_MEMBAR; \
- stx TAG, [TSB + 0x00];
+#define KTSB_WRITE(TSB, TTE, TAG) \
+ add TSB, 0x8, TSB; \
+ stxa TTE, [TSB] ASI_N; \
+ sub TSB, 0x8, TSB; \
+ TSB_MEMBAR; \
+ stxa TAG, [TSB] ASI_N;
/* Do a kernel page table walk. Leaves physical PTE pointer in
* REG1. Jumps to FAIL_LABEL on early page table walk termination.
add REG1, (3 * 8), REG1; \
99:
+ /* We use a 32K TSB for the whole kernel, this allows to
+ * handle about 16MB of modules and vmalloc mappings without
+ * incurring many hash conflicts.
+ */
+#define KERNEL_TSB_SIZE_BYTES (32 * 1024)
+#define KERNEL_TSB_NENTRIES \
+ (KERNEL_TSB_SIZE_BYTES / 16)
+#define KERNEL_TSB4M_NENTRIES 4096
+
/* Do a kernel TSB lookup at tl>0 on VADDR+TAG, branch to OK_LABEL
* on TSB hit. REG1, REG2, REG3, and REG4 are used as temporaries
* and the found TTE will be left in REG1. REG3 and REG4 must
*
* VADDR and TAG will be preserved and not clobbered by this macro.
*/
- /* XXX non-8K base page size support... */
#define KERN_TSB_LOOKUP_TL1(VADDR, TAG, REG1, REG2, REG3, REG4, OK_LABEL) \
sethi %hi(swapper_tsb), REG1; \
or REG1, %lo(swapper_tsb), REG1; \
- srlx VADDR, 13, REG2; \
- and REG2, (512 - 1), REG2; \
+ srlx VADDR, PAGE_SHIFT, REG2; \
+ and REG2, (KERNEL_TSB_NENTRIES - 1), REG2; \
+ sllx REG2, 4, REG2; \
+ add REG1, REG2, REG2; \
+ KTSB_LOAD_QUAD(REG2, REG3); \
+ cmp REG3, TAG; \
+ be,a,pt %xcc, OK_LABEL; \
+ mov REG4, REG1;
+
+#ifndef CONFIG_DEBUG_PAGEALLOC
+ /* This version uses a trick, the TAG is already (VADDR >> 22) so
+ * we can make use of that for the index computation.
+ */
+#define KERN_TSB4M_LOOKUP_TL1(TAG, REG1, REG2, REG3, REG4, OK_LABEL) \
+ sethi %hi(swapper_4m_tsb), REG1; \
+ or REG1, %lo(swapper_4m_tsb), REG1; \
+ and TAG, (KERNEL_TSB_NENTRIES - 1), REG2; \
sllx REG2, 4, REG2; \
add REG1, REG2, REG2; \
- ldda [REG2] ASI_NUCLEUS_QUAD_LDD, REG3; \
+ KTSB_LOAD_QUAD(REG2, REG3); \
cmp REG3, TAG; \
be,a,pt %xcc, OK_LABEL; \
mov REG4, REG1;
+#endif
#endif /* !(_SPARC64_TSB_H) */