2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Synthesize TLB refill handlers at runtime.
8 * Copyright (C) 2004, 2005, 2006, 2008 Thiemo Seufer
9 * Copyright (C) 2005, 2007, 2008, 2009 Maciej W. Rozycki
10 * Copyright (C) 2006 Ralf Baechle (ralf@linux-mips.org)
11 * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12 * Copyright (C) 2011 MIPS Technologies, Inc.
14 * ... and the days got worse and worse and now you see
15 * I've gone completly out of my mind.
17 * They're coming to take me a away haha
18 * they're coming to take me a away hoho hihi haha
19 * to the funny farm where code is beautiful all the time ...
21 * (Condolences to Napoleon XIV)
24 #include <linux/bug.h>
25 #include <linux/kernel.h>
26 #include <linux/types.h>
27 #include <linux/smp.h>
28 #include <linux/string.h>
29 #include <linux/init.h>
30 #include <linux/cache.h>
32 #include <asm/cacheflush.h>
33 #include <asm/cpu-type.h>
34 #include <asm/pgtable.h>
37 #include <asm/setup.h>
40 * TLB load/store/modify handlers.
42 * Only the fastpath gets synthesized at runtime, the slowpath for
43 * do_page_fault remains normal asm.
45 extern void tlb_do_page_fault_0(void);
46 extern void tlb_do_page_fault_1(void);
48 struct work_registers {
57 } ____cacheline_aligned_in_smp;
59 static struct tlb_reg_save handler_reg_save[NR_CPUS];
61 static inline int r45k_bvahwbug(void)
63 /* XXX: We should probe for the presence of this bug, but we don't. */
67 static inline int r4k_250MHZhwbug(void)
69 /* XXX: We should probe for the presence of this bug, but we don't. */
73 static inline int __maybe_unused bcm1250_m3_war(void)
75 return BCM1250_M3_WAR;
78 static inline int __maybe_unused r10000_llsc_war(void)
80 return R10000_LLSC_WAR;
83 static int use_bbit_insns(void)
85 switch (current_cpu_type()) {
86 case CPU_CAVIUM_OCTEON:
87 case CPU_CAVIUM_OCTEON_PLUS:
88 case CPU_CAVIUM_OCTEON2:
89 case CPU_CAVIUM_OCTEON3:
96 static int use_lwx_insns(void)
98 switch (current_cpu_type()) {
99 case CPU_CAVIUM_OCTEON2:
100 case CPU_CAVIUM_OCTEON3:
106 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
107 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
108 static bool scratchpad_available(void)
112 static int scratchpad_offset(int i)
115 * CVMSEG starts at address -32768 and extends for
116 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
118 i += 1; /* Kernel use starts at the top and works down. */
119 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
122 static bool scratchpad_available(void)
126 static int scratchpad_offset(int i)
129 /* Really unreachable, but evidently some GCC want this. */
134 * Found by experiment: At least some revisions of the 4kc throw under
135 * some circumstances a machine check exception, triggered by invalid
136 * values in the index register. Delaying the tlbp instruction until
137 * after the next branch, plus adding an additional nop in front of
138 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
139 * why; it's not an issue caused by the core RTL.
142 static int m4kc_tlbp_war(void)
144 return (current_cpu_data.processor_id & 0xffff00) ==
145 (PRID_COMP_MIPS | PRID_IMP_4KC);
148 /* Handle labels (which must be positive integers). */
150 label_second_part = 1,
155 label_split = label_tlbw_hazard_0 + 8,
156 label_tlbl_goaround1,
157 label_tlbl_goaround2,
161 label_smp_pgtable_change,
162 label_r3000_write_probe_fail,
163 label_large_segbits_fault,
164 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
165 label_tlb_huge_update,
169 UASM_L_LA(_second_part)
172 UASM_L_LA(_vmalloc_done)
173 /* _tlbw_hazard_x is handled differently. */
175 UASM_L_LA(_tlbl_goaround1)
176 UASM_L_LA(_tlbl_goaround2)
177 UASM_L_LA(_nopage_tlbl)
178 UASM_L_LA(_nopage_tlbs)
179 UASM_L_LA(_nopage_tlbm)
180 UASM_L_LA(_smp_pgtable_change)
181 UASM_L_LA(_r3000_write_probe_fail)
182 UASM_L_LA(_large_segbits_fault)
183 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
184 UASM_L_LA(_tlb_huge_update)
187 static int hazard_instance;
189 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
193 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
200 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
204 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
212 * pgtable bits are assigned dynamically depending on processor feature
213 * and statically based on kernel configuration. This spits out the actual
214 * values the kernel is using. Required to make sense from disassembled
215 * TLB exception handlers.
217 static void output_pgtable_bits_defines(void)
219 #define pr_define(fmt, ...) \
220 pr_debug("#define " fmt, ##__VA_ARGS__)
222 pr_debug("#include <asm/asm.h>\n");
223 pr_debug("#include <asm/regdef.h>\n");
226 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
227 pr_define("_PAGE_READ_SHIFT %d\n", _PAGE_READ_SHIFT);
228 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
229 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
230 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
231 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
232 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
233 pr_define("_PAGE_SPLITTING_SHIFT %d\n", _PAGE_SPLITTING_SHIFT);
236 #ifdef _PAGE_NO_EXEC_SHIFT
237 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
239 #ifdef _PAGE_NO_READ_SHIFT
240 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
243 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
244 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
245 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
246 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
250 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
254 pr_debug("LEAF(%s)\n", symbol);
256 pr_debug("\t.set push\n");
257 pr_debug("\t.set noreorder\n");
259 for (i = 0; i < count; i++)
260 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
262 pr_debug("\t.set\tpop\n");
264 pr_debug("\tEND(%s)\n", symbol);
267 /* The only general purpose registers allowed in TLB handlers. */
271 /* Some CP0 registers */
272 #define C0_INDEX 0, 0
273 #define C0_ENTRYLO0 2, 0
274 #define C0_TCBIND 2, 2
275 #define C0_ENTRYLO1 3, 0
276 #define C0_CONTEXT 4, 0
277 #define C0_PAGEMASK 5, 0
278 #define C0_BADVADDR 8, 0
279 #define C0_ENTRYHI 10, 0
281 #define C0_XCONTEXT 20, 0
284 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
286 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
289 /* The worst case length of the handler is around 18 instructions for
290 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
291 * Maximum space available is 32 instructions for R3000 and 64
292 * instructions for R4000.
294 * We deliberately chose a buffer size of 128, so we won't scribble
295 * over anything important on overflow before we panic.
297 static u32 tlb_handler[128];
299 /* simply assume worst case size for labels and relocs */
300 static struct uasm_label labels[128];
301 static struct uasm_reloc relocs[128];
303 static int check_for_high_segbits;
305 static unsigned int kscratch_used_mask;
307 static inline int __maybe_unused c0_kscratch(void)
309 switch (current_cpu_type()) {
318 static int allocate_kscratch(void)
321 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
328 r--; /* make it zero based */
330 kscratch_used_mask |= (1 << r);
335 static int scratch_reg;
337 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
339 static struct work_registers build_get_work_registers(u32 **p)
341 struct work_registers r;
343 if (scratch_reg >= 0) {
344 /* Save in CPU local C0_KScratch? */
345 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
352 if (num_possible_cpus() > 1) {
353 /* Get smp_processor_id */
354 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
355 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
357 /* handler_reg_save index in K0 */
358 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
360 UASM_i_LA(p, K1, (long)&handler_reg_save);
361 UASM_i_ADDU(p, K0, K0, K1);
363 UASM_i_LA(p, K0, (long)&handler_reg_save);
365 /* K0 now points to save area, save $1 and $2 */
366 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
367 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
375 static void build_restore_work_registers(u32 **p)
377 if (scratch_reg >= 0) {
378 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
381 /* K0 already points to save area, restore $1 and $2 */
382 UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
383 UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
386 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
389 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
390 * we cannot do r3000 under these circumstances.
392 * Declare pgd_current here instead of including mmu_context.h to avoid type
393 * conflicts for tlbmiss_handler_setup_pgd
395 extern unsigned long pgd_current[];
398 * The R3000 TLB handler is simple.
400 static void build_r3000_tlb_refill_handler(void)
402 long pgdc = (long)pgd_current;
405 memset(tlb_handler, 0, sizeof(tlb_handler));
408 uasm_i_mfc0(&p, K0, C0_BADVADDR);
409 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
410 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
411 uasm_i_srl(&p, K0, K0, 22); /* load delay */
412 uasm_i_sll(&p, K0, K0, 2);
413 uasm_i_addu(&p, K1, K1, K0);
414 uasm_i_mfc0(&p, K0, C0_CONTEXT);
415 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
416 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
417 uasm_i_addu(&p, K1, K1, K0);
418 uasm_i_lw(&p, K0, 0, K1);
419 uasm_i_nop(&p); /* load delay */
420 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
421 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
422 uasm_i_tlbwr(&p); /* cp0 delay */
424 uasm_i_rfe(&p); /* branch delay */
426 if (p > tlb_handler + 32)
427 panic("TLB refill handler space exceeded");
429 pr_debug("Wrote TLB refill handler (%u instructions).\n",
430 (unsigned int)(p - tlb_handler));
432 memcpy((void *)ebase, tlb_handler, 0x80);
434 dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
436 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
439 * The R4000 TLB handler is much more complicated. We have two
440 * consecutive handler areas with 32 instructions space each.
441 * Since they aren't used at the same time, we can overflow in the
442 * other one.To keep things simple, we first assume linear space,
443 * then we relocate it to the final handler layout as needed.
445 static u32 final_handler[64];
450 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
451 * 2. A timing hazard exists for the TLBP instruction.
453 * stalling_instruction
456 * The JTLB is being read for the TLBP throughout the stall generated by the
457 * previous instruction. This is not really correct as the stalling instruction
458 * can modify the address used to access the JTLB. The failure symptom is that
459 * the TLBP instruction will use an address created for the stalling instruction
460 * and not the address held in C0_ENHI and thus report the wrong results.
462 * The software work-around is to not allow the instruction preceding the TLBP
463 * to stall - make it an NOP or some other instruction guaranteed not to stall.
465 * Errata 2 will not be fixed. This errata is also on the R5000.
467 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
469 static void __maybe_unused build_tlb_probe_entry(u32 **p)
471 switch (current_cpu_type()) {
472 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
488 * Write random or indexed TLB entry, and care about the hazards from
489 * the preceding mtc0 and for the following eret.
491 enum tlb_write_entry { tlb_random, tlb_indexed };
493 static void build_tlb_write_entry(u32 **p, struct uasm_label **l,
494 struct uasm_reloc **r,
495 enum tlb_write_entry wmode)
497 void(*tlbw)(u32 **) = NULL;
500 case tlb_random: tlbw = uasm_i_tlbwr; break;
501 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
504 if (cpu_has_mips_r2) {
506 * The architecture spec says an ehb is required here,
507 * but a number of cores do not have the hazard and
508 * using an ehb causes an expensive pipeline stall.
510 switch (current_cpu_type()) {
524 switch (current_cpu_type()) {
532 * This branch uses up a mtc0 hazard nop slot and saves
533 * two nops after the tlbw instruction.
535 uasm_bgezl_hazard(p, r, hazard_instance);
537 uasm_bgezl_label(l, p, hazard_instance);
551 uasm_i_nop(p); /* QED specifies 2 nops hazard */
552 uasm_i_nop(p); /* QED specifies 2 nops hazard */
624 panic("No TLB refill handler yet (CPU type: %d)",
625 current_cpu_data.cputype);
630 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
634 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
636 #ifdef CONFIG_64BIT_PHYS_ADDR
637 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
639 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
644 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
646 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
647 unsigned int tmp, enum label_id lid,
650 if (restore_scratch) {
651 /* Reset default page size */
652 if (PM_DEFAULT_MASK >> 16) {
653 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
654 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
655 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
656 uasm_il_b(p, r, lid);
657 } else if (PM_DEFAULT_MASK) {
658 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
659 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
660 uasm_il_b(p, r, lid);
662 uasm_i_mtc0(p, 0, C0_PAGEMASK);
663 uasm_il_b(p, r, lid);
665 if (scratch_reg >= 0)
666 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
668 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
670 /* Reset default page size */
671 if (PM_DEFAULT_MASK >> 16) {
672 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
673 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
674 uasm_il_b(p, r, lid);
675 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
676 } else if (PM_DEFAULT_MASK) {
677 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
678 uasm_il_b(p, r, lid);
679 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
681 uasm_il_b(p, r, lid);
682 uasm_i_mtc0(p, 0, C0_PAGEMASK);
687 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
688 struct uasm_reloc **r,
690 enum tlb_write_entry wmode,
693 /* Set huge page tlb entry size */
694 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
695 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
696 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
698 build_tlb_write_entry(p, l, r, wmode);
700 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
704 * Check if Huge PTE is present, if so then jump to LABEL.
707 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
708 unsigned int pmd, int lid)
710 UASM_i_LW(p, tmp, 0, pmd);
711 if (use_bbit_insns()) {
712 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
714 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
715 uasm_il_bnez(p, r, tmp, lid);
719 static void build_huge_update_entries(u32 **p, unsigned int pte,
725 * A huge PTE describes an area the size of the
726 * configured huge page size. This is twice the
727 * of the large TLB entry size we intend to use.
728 * A TLB entry half the size of the configured
729 * huge page size is configured into entrylo0
730 * and entrylo1 to cover the contiguous huge PTE
733 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
735 /* We can clobber tmp. It isn't used after this.*/
737 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
739 build_convert_pte_to_entrylo(p, pte);
740 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
741 /* convert to entrylo1 */
743 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
745 UASM_i_ADDU(p, pte, pte, tmp);
747 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
750 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
751 struct uasm_label **l,
756 UASM_i_SC(p, pte, 0, ptr);
757 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
758 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
760 UASM_i_SW(p, pte, 0, ptr);
762 build_huge_update_entries(p, pte, ptr);
763 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
765 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
769 * TMP and PTR are scratch.
770 * TMP will be clobbered, PTR will hold the pmd entry.
773 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
774 unsigned int tmp, unsigned int ptr)
776 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
777 long pgdc = (long)pgd_current;
780 * The vmalloc handling is not in the hotpath.
782 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
784 if (check_for_high_segbits) {
786 * The kernel currently implicitely assumes that the
787 * MIPS SEGBITS parameter for the processor is
788 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
789 * allocate virtual addresses outside the maximum
790 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
791 * that doesn't prevent user code from accessing the
792 * higher xuseg addresses. Here, we make sure that
793 * everything but the lower xuseg addresses goes down
794 * the module_alloc/vmalloc path.
796 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
797 uasm_il_bnez(p, r, ptr, label_vmalloc);
799 uasm_il_bltz(p, r, tmp, label_vmalloc);
801 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
804 /* pgd is in pgd_reg */
805 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
807 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
809 * &pgd << 11 stored in CONTEXT [23..63].
811 UASM_i_MFC0(p, ptr, C0_CONTEXT);
813 /* Clear lower 23 bits of context. */
814 uasm_i_dins(p, ptr, 0, 0, 23);
816 /* 1 0 1 0 1 << 6 xkphys cached */
817 uasm_i_ori(p, ptr, ptr, 0x540);
818 uasm_i_drotr(p, ptr, ptr, 11);
819 #elif defined(CONFIG_SMP)
820 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
821 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
822 UASM_i_LA_mostly(p, tmp, pgdc);
823 uasm_i_daddu(p, ptr, ptr, tmp);
824 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
825 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
827 UASM_i_LA_mostly(p, ptr, pgdc);
828 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
832 uasm_l_vmalloc_done(l, *p);
834 /* get pgd offset in bytes */
835 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
837 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
838 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
839 #ifndef __PAGETABLE_PMD_FOLDED
840 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
841 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
842 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
843 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
844 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
849 * BVADDR is the faulting address, PTR is scratch.
850 * PTR will hold the pgd for vmalloc.
853 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
854 unsigned int bvaddr, unsigned int ptr,
855 enum vmalloc64_mode mode)
857 long swpd = (long)swapper_pg_dir;
858 int single_insn_swpd;
859 int did_vmalloc_branch = 0;
861 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
863 uasm_l_vmalloc(l, *p);
865 if (mode != not_refill && check_for_high_segbits) {
866 if (single_insn_swpd) {
867 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
868 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
869 did_vmalloc_branch = 1;
872 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
875 if (!did_vmalloc_branch) {
876 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
877 uasm_il_b(p, r, label_vmalloc_done);
878 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
880 UASM_i_LA_mostly(p, ptr, swpd);
881 uasm_il_b(p, r, label_vmalloc_done);
882 if (uasm_in_compat_space_p(swpd))
883 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
885 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
888 if (mode != not_refill && check_for_high_segbits) {
889 uasm_l_large_segbits_fault(l, *p);
891 * We get here if we are an xsseg address, or if we are
892 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
894 * Ignoring xsseg (assume disabled so would generate
895 * (address errors?), the only remaining possibility
896 * is the upper xuseg addresses. On processors with
897 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
898 * addresses would have taken an address error. We try
899 * to mimic that here by taking a load/istream page
902 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
905 if (mode == refill_scratch) {
906 if (scratch_reg >= 0)
907 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
909 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
916 #else /* !CONFIG_64BIT */
919 * TMP and PTR are scratch.
920 * TMP will be clobbered, PTR will hold the pgd entry.
922 static void __maybe_unused
923 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
926 /* pgd is in pgd_reg */
927 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
928 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
930 long pgdc = (long)pgd_current;
932 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
934 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
935 UASM_i_LA_mostly(p, tmp, pgdc);
936 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
937 uasm_i_addu(p, ptr, tmp, ptr);
939 UASM_i_LA_mostly(p, ptr, pgdc);
941 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
942 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
944 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
945 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
946 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
949 #endif /* !CONFIG_64BIT */
951 static void build_adjust_context(u32 **p, unsigned int ctx)
953 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
954 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
956 switch (current_cpu_type()) {
973 UASM_i_SRL(p, ctx, ctx, shift);
974 uasm_i_andi(p, ctx, ctx, mask);
977 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
980 * Bug workaround for the Nevada. It seems as if under certain
981 * circumstances the move from cp0_context might produce a
982 * bogus result when the mfc0 instruction and its consumer are
983 * in a different cacheline or a load instruction, probably any
984 * memory reference, is between them.
986 switch (current_cpu_type()) {
988 UASM_i_LW(p, ptr, 0, ptr);
989 GET_CONTEXT(p, tmp); /* get context reg */
993 GET_CONTEXT(p, tmp); /* get context reg */
994 UASM_i_LW(p, ptr, 0, ptr);
998 build_adjust_context(p, tmp);
999 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1002 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1005 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1006 * Kernel is a special case. Only a few CPUs use it.
1008 #ifdef CONFIG_64BIT_PHYS_ADDR
1009 if (cpu_has_64bits) {
1010 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1011 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1013 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1014 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1015 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1017 uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1018 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1019 uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1021 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1023 int pte_off_even = sizeof(pte_t) / 2;
1024 int pte_off_odd = pte_off_even + sizeof(pte_t);
1026 /* The pte entries are pre-shifted */
1027 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1028 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1029 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1030 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1033 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1034 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1035 if (r45k_bvahwbug())
1036 build_tlb_probe_entry(p);
1038 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1039 if (r4k_250MHZhwbug())
1040 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1041 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1042 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1044 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1045 if (r4k_250MHZhwbug())
1046 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1047 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1048 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1049 if (r45k_bvahwbug())
1050 uasm_i_mfc0(p, tmp, C0_INDEX);
1052 if (r4k_250MHZhwbug())
1053 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1054 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1058 struct mips_huge_tlb_info {
1060 int restore_scratch;
1063 static struct mips_huge_tlb_info
1064 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1065 struct uasm_reloc **r, unsigned int tmp,
1066 unsigned int ptr, int c0_scratch_reg)
1068 struct mips_huge_tlb_info rv;
1069 unsigned int even, odd;
1070 int vmalloc_branch_delay_filled = 0;
1071 const int scratch = 1; /* Our extra working register */
1073 rv.huge_pte = scratch;
1074 rv.restore_scratch = 0;
1076 if (check_for_high_segbits) {
1077 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1080 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1082 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1084 if (c0_scratch_reg >= 0)
1085 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1087 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1089 uasm_i_dsrl_safe(p, scratch, tmp,
1090 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1091 uasm_il_bnez(p, r, scratch, label_vmalloc);
1093 if (pgd_reg == -1) {
1094 vmalloc_branch_delay_filled = 1;
1095 /* Clear lower 23 bits of context. */
1096 uasm_i_dins(p, ptr, 0, 0, 23);
1100 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1102 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1104 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1106 if (c0_scratch_reg >= 0)
1107 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1109 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1112 /* Clear lower 23 bits of context. */
1113 uasm_i_dins(p, ptr, 0, 0, 23);
1115 uasm_il_bltz(p, r, tmp, label_vmalloc);
1118 if (pgd_reg == -1) {
1119 vmalloc_branch_delay_filled = 1;
1120 /* 1 0 1 0 1 << 6 xkphys cached */
1121 uasm_i_ori(p, ptr, ptr, 0x540);
1122 uasm_i_drotr(p, ptr, ptr, 11);
1125 #ifdef __PAGETABLE_PMD_FOLDED
1126 #define LOC_PTEP scratch
1128 #define LOC_PTEP ptr
1131 if (!vmalloc_branch_delay_filled)
1132 /* get pgd offset in bytes */
1133 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1135 uasm_l_vmalloc_done(l, *p);
1139 * fall-through case = badvaddr *pgd_current
1140 * vmalloc case = badvaddr swapper_pg_dir
1143 if (vmalloc_branch_delay_filled)
1144 /* get pgd offset in bytes */
1145 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1147 #ifdef __PAGETABLE_PMD_FOLDED
1148 GET_CONTEXT(p, tmp); /* get context reg */
1150 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1152 if (use_lwx_insns()) {
1153 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1155 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1156 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1159 #ifndef __PAGETABLE_PMD_FOLDED
1160 /* get pmd offset in bytes */
1161 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1162 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1163 GET_CONTEXT(p, tmp); /* get context reg */
1165 if (use_lwx_insns()) {
1166 UASM_i_LWX(p, scratch, scratch, ptr);
1168 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1169 UASM_i_LW(p, scratch, 0, ptr);
1172 /* Adjust the context during the load latency. */
1173 build_adjust_context(p, tmp);
1175 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1176 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1178 * The in the LWX case we don't want to do the load in the
1179 * delay slot. It cannot issue in the same cycle and may be
1180 * speculative and unneeded.
1182 if (use_lwx_insns())
1184 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1187 /* build_update_entries */
1188 if (use_lwx_insns()) {
1191 UASM_i_LWX(p, even, scratch, tmp);
1192 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1193 UASM_i_LWX(p, odd, scratch, tmp);
1195 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1198 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1199 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1202 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1203 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1204 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1206 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1207 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1208 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1210 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1212 if (c0_scratch_reg >= 0) {
1213 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1214 build_tlb_write_entry(p, l, r, tlb_random);
1215 uasm_l_leave(l, *p);
1216 rv.restore_scratch = 1;
1217 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1218 build_tlb_write_entry(p, l, r, tlb_random);
1219 uasm_l_leave(l, *p);
1220 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1222 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1223 build_tlb_write_entry(p, l, r, tlb_random);
1224 uasm_l_leave(l, *p);
1225 rv.restore_scratch = 1;
1228 uasm_i_eret(p); /* return from trap */
1234 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1235 * because EXL == 0. If we wrap, we can also use the 32 instruction
1236 * slots before the XTLB refill exception handler which belong to the
1237 * unused TLB refill exception.
1239 #define MIPS64_REFILL_INSNS 32
1241 static void build_r4000_tlb_refill_handler(void)
1243 u32 *p = tlb_handler;
1244 struct uasm_label *l = labels;
1245 struct uasm_reloc *r = relocs;
1247 unsigned int final_len;
1248 struct mips_huge_tlb_info htlb_info __maybe_unused;
1249 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1251 memset(tlb_handler, 0, sizeof(tlb_handler));
1252 memset(labels, 0, sizeof(labels));
1253 memset(relocs, 0, sizeof(relocs));
1254 memset(final_handler, 0, sizeof(final_handler));
1256 if ((scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1257 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1259 vmalloc_mode = refill_scratch;
1261 htlb_info.huge_pte = K0;
1262 htlb_info.restore_scratch = 0;
1263 vmalloc_mode = refill_noscratch;
1265 * create the plain linear handler
1267 if (bcm1250_m3_war()) {
1268 unsigned int segbits = 44;
1270 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1271 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1272 uasm_i_xor(&p, K0, K0, K1);
1273 uasm_i_dsrl_safe(&p, K1, K0, 62);
1274 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1275 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1276 uasm_i_or(&p, K0, K0, K1);
1277 uasm_il_bnez(&p, &r, K0, label_leave);
1278 /* No need for uasm_i_nop */
1282 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1284 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1287 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1288 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1291 build_get_ptep(&p, K0, K1);
1292 build_update_entries(&p, K0, K1);
1293 build_tlb_write_entry(&p, &l, &r, tlb_random);
1294 uasm_l_leave(&l, p);
1295 uasm_i_eret(&p); /* return from trap */
1297 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1298 uasm_l_tlb_huge_update(&l, p);
1299 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1300 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1301 htlb_info.restore_scratch);
1305 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1309 * Overflow check: For the 64bit handler, we need at least one
1310 * free instruction slot for the wrap-around branch. In worst
1311 * case, if the intended insertion point is a delay slot, we
1312 * need three, with the second nop'ed and the third being
1315 switch (boot_cpu_type()) {
1317 if (sizeof(long) == 4) {
1319 /* Loongson2 ebase is different than r4k, we have more space */
1320 if ((p - tlb_handler) > 64)
1321 panic("TLB refill handler space exceeded");
1323 * Now fold the handler in the TLB refill handler space.
1326 /* Simplest case, just copy the handler. */
1327 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1328 final_len = p - tlb_handler;
1331 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1332 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1333 && uasm_insn_has_bdelay(relocs,
1334 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1335 panic("TLB refill handler space exceeded");
1337 * Now fold the handler in the TLB refill handler space.
1339 f = final_handler + MIPS64_REFILL_INSNS;
1340 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1341 /* Just copy the handler. */
1342 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1343 final_len = p - tlb_handler;
1345 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1346 const enum label_id ls = label_tlb_huge_update;
1348 const enum label_id ls = label_vmalloc;
1354 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1356 BUG_ON(i == ARRAY_SIZE(labels));
1357 split = labels[i].addr;
1360 * See if we have overflown one way or the other.
1362 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1363 split < p - MIPS64_REFILL_INSNS)
1368 * Split two instructions before the end. One
1369 * for the branch and one for the instruction
1370 * in the delay slot.
1372 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1375 * If the branch would fall in a delay slot,
1376 * we must back up an additional instruction
1377 * so that it is no longer in a delay slot.
1379 if (uasm_insn_has_bdelay(relocs, split - 1))
1382 /* Copy first part of the handler. */
1383 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1384 f += split - tlb_handler;
1387 /* Insert branch. */
1388 uasm_l_split(&l, final_handler);
1389 uasm_il_b(&f, &r, label_split);
1390 if (uasm_insn_has_bdelay(relocs, split))
1393 uasm_copy_handler(relocs, labels,
1394 split, split + 1, f);
1395 uasm_move_labels(labels, f, f + 1, -1);
1401 /* Copy the rest of the handler. */
1402 uasm_copy_handler(relocs, labels, split, p, final_handler);
1403 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1410 uasm_resolve_relocs(relocs, labels);
1411 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1414 memcpy((void *)ebase, final_handler, 0x100);
1416 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1419 extern u32 handle_tlbl[], handle_tlbl_end[];
1420 extern u32 handle_tlbs[], handle_tlbs_end[];
1421 extern u32 handle_tlbm[], handle_tlbm_end[];
1422 extern u32 tlbmiss_handler_setup_pgd[], tlbmiss_handler_setup_pgd_end[];
1424 static void build_setup_pgd(void)
1427 const int __maybe_unused a1 = 5;
1428 const int __maybe_unused a2 = 6;
1429 u32 *p = tlbmiss_handler_setup_pgd;
1430 const int tlbmiss_handler_setup_pgd_size =
1431 tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd;
1432 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1433 long pgdc = (long)pgd_current;
1436 memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1437 sizeof(tlbmiss_handler_setup_pgd[0]));
1438 memset(labels, 0, sizeof(labels));
1439 memset(relocs, 0, sizeof(relocs));
1440 pgd_reg = allocate_kscratch();
1441 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1442 if (pgd_reg == -1) {
1443 struct uasm_label *l = labels;
1444 struct uasm_reloc *r = relocs;
1446 /* PGD << 11 in c0_Context */
1448 * If it is a ckseg0 address, convert to a physical
1449 * address. Shifting right by 29 and adding 4 will
1450 * result in zero for these addresses.
1453 UASM_i_SRA(&p, a1, a0, 29);
1454 UASM_i_ADDIU(&p, a1, a1, 4);
1455 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1457 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1458 uasm_l_tlbl_goaround1(&l, p);
1459 UASM_i_SLL(&p, a0, a0, 11);
1461 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1463 /* PGD in c0_KScratch */
1465 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1469 /* Save PGD to pgd_current[smp_processor_id()] */
1470 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1471 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1472 UASM_i_LA_mostly(&p, a2, pgdc);
1473 UASM_i_ADDU(&p, a2, a2, a1);
1474 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1476 UASM_i_LA_mostly(&p, a2, pgdc);
1477 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1481 /* if pgd_reg is allocated, save PGD also to scratch register */
1483 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1487 if (p >= tlbmiss_handler_setup_pgd_end)
1488 panic("tlbmiss_handler_setup_pgd space exceeded");
1490 uasm_resolve_relocs(relocs, labels);
1491 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1492 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1494 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1495 tlbmiss_handler_setup_pgd_size);
1499 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1502 # ifdef CONFIG_64BIT_PHYS_ADDR
1504 uasm_i_lld(p, pte, 0, ptr);
1507 UASM_i_LL(p, pte, 0, ptr);
1509 # ifdef CONFIG_64BIT_PHYS_ADDR
1511 uasm_i_ld(p, pte, 0, ptr);
1514 UASM_i_LW(p, pte, 0, ptr);
1519 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1522 #ifdef CONFIG_64BIT_PHYS_ADDR
1523 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1526 uasm_i_ori(p, pte, pte, mode);
1528 # ifdef CONFIG_64BIT_PHYS_ADDR
1530 uasm_i_scd(p, pte, 0, ptr);
1533 UASM_i_SC(p, pte, 0, ptr);
1535 if (r10000_llsc_war())
1536 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1538 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1540 # ifdef CONFIG_64BIT_PHYS_ADDR
1541 if (!cpu_has_64bits) {
1542 /* no uasm_i_nop needed */
1543 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1544 uasm_i_ori(p, pte, pte, hwmode);
1545 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1546 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1547 /* no uasm_i_nop needed */
1548 uasm_i_lw(p, pte, 0, ptr);
1555 # ifdef CONFIG_64BIT_PHYS_ADDR
1557 uasm_i_sd(p, pte, 0, ptr);
1560 UASM_i_SW(p, pte, 0, ptr);
1562 # ifdef CONFIG_64BIT_PHYS_ADDR
1563 if (!cpu_has_64bits) {
1564 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1565 uasm_i_ori(p, pte, pte, hwmode);
1566 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1567 uasm_i_lw(p, pte, 0, ptr);
1574 * Check if PTE is present, if not then jump to LABEL. PTR points to
1575 * the page table where this PTE is located, PTE will be re-loaded
1576 * with it's original value.
1579 build_pte_present(u32 **p, struct uasm_reloc **r,
1580 int pte, int ptr, int scratch, enum label_id lid)
1582 int t = scratch >= 0 ? scratch : pte;
1585 if (use_bbit_insns()) {
1586 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1589 uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1590 uasm_il_beqz(p, r, t, lid);
1592 /* You lose the SMP race :-(*/
1593 iPTE_LW(p, pte, ptr);
1596 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1597 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1598 uasm_il_bnez(p, r, t, lid);
1600 /* You lose the SMP race :-(*/
1601 iPTE_LW(p, pte, ptr);
1605 /* Make PTE valid, store result in PTR. */
1607 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1610 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1612 iPTE_SW(p, r, pte, ptr, mode);
1616 * Check if PTE can be written to, if not branch to LABEL. Regardless
1617 * restore PTE with value from PTR when done.
1620 build_pte_writable(u32 **p, struct uasm_reloc **r,
1621 unsigned int pte, unsigned int ptr, int scratch,
1624 int t = scratch >= 0 ? scratch : pte;
1626 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1627 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1628 uasm_il_bnez(p, r, t, lid);
1630 /* You lose the SMP race :-(*/
1631 iPTE_LW(p, pte, ptr);
1636 /* Make PTE writable, update software status bits as well, then store
1640 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1643 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1646 iPTE_SW(p, r, pte, ptr, mode);
1650 * Check if PTE can be modified, if not branch to LABEL. Regardless
1651 * restore PTE with value from PTR when done.
1654 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1655 unsigned int pte, unsigned int ptr, int scratch,
1658 if (use_bbit_insns()) {
1659 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1662 int t = scratch >= 0 ? scratch : pte;
1663 uasm_i_andi(p, t, pte, _PAGE_WRITE);
1664 uasm_il_beqz(p, r, t, lid);
1666 /* You lose the SMP race :-(*/
1667 iPTE_LW(p, pte, ptr);
1671 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1675 * R3000 style TLB load/store/modify handlers.
1679 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1683 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1685 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1686 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1689 uasm_i_rfe(p); /* branch delay */
1693 * This places the pte into ENTRYLO0 and writes it with tlbwi
1694 * or tlbwr as appropriate. This is because the index register
1695 * may have the probe fail bit set as a result of a trap on a
1696 * kseg2 access, i.e. without refill. Then it returns.
1699 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1700 struct uasm_reloc **r, unsigned int pte,
1703 uasm_i_mfc0(p, tmp, C0_INDEX);
1704 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1705 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1706 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1707 uasm_i_tlbwi(p); /* cp0 delay */
1709 uasm_i_rfe(p); /* branch delay */
1710 uasm_l_r3000_write_probe_fail(l, *p);
1711 uasm_i_tlbwr(p); /* cp0 delay */
1713 uasm_i_rfe(p); /* branch delay */
1717 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1720 long pgdc = (long)pgd_current;
1722 uasm_i_mfc0(p, pte, C0_BADVADDR);
1723 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1724 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1725 uasm_i_srl(p, pte, pte, 22); /* load delay */
1726 uasm_i_sll(p, pte, pte, 2);
1727 uasm_i_addu(p, ptr, ptr, pte);
1728 uasm_i_mfc0(p, pte, C0_CONTEXT);
1729 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1730 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1731 uasm_i_addu(p, ptr, ptr, pte);
1732 uasm_i_lw(p, pte, 0, ptr);
1733 uasm_i_tlbp(p); /* load delay */
1736 static void build_r3000_tlb_load_handler(void)
1738 u32 *p = handle_tlbl;
1739 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1740 struct uasm_label *l = labels;
1741 struct uasm_reloc *r = relocs;
1743 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1744 memset(labels, 0, sizeof(labels));
1745 memset(relocs, 0, sizeof(relocs));
1747 build_r3000_tlbchange_handler_head(&p, K0, K1);
1748 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1749 uasm_i_nop(&p); /* load delay */
1750 build_make_valid(&p, &r, K0, K1);
1751 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1753 uasm_l_nopage_tlbl(&l, p);
1754 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1757 if (p >= handle_tlbl_end)
1758 panic("TLB load handler fastpath space exceeded");
1760 uasm_resolve_relocs(relocs, labels);
1761 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1762 (unsigned int)(p - handle_tlbl));
1764 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1767 static void build_r3000_tlb_store_handler(void)
1769 u32 *p = handle_tlbs;
1770 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1771 struct uasm_label *l = labels;
1772 struct uasm_reloc *r = relocs;
1774 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1775 memset(labels, 0, sizeof(labels));
1776 memset(relocs, 0, sizeof(relocs));
1778 build_r3000_tlbchange_handler_head(&p, K0, K1);
1779 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1780 uasm_i_nop(&p); /* load delay */
1781 build_make_write(&p, &r, K0, K1);
1782 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1784 uasm_l_nopage_tlbs(&l, p);
1785 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1788 if (p >= handle_tlbs_end)
1789 panic("TLB store handler fastpath space exceeded");
1791 uasm_resolve_relocs(relocs, labels);
1792 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1793 (unsigned int)(p - handle_tlbs));
1795 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1798 static void build_r3000_tlb_modify_handler(void)
1800 u32 *p = handle_tlbm;
1801 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1802 struct uasm_label *l = labels;
1803 struct uasm_reloc *r = relocs;
1805 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1806 memset(labels, 0, sizeof(labels));
1807 memset(relocs, 0, sizeof(relocs));
1809 build_r3000_tlbchange_handler_head(&p, K0, K1);
1810 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1811 uasm_i_nop(&p); /* load delay */
1812 build_make_write(&p, &r, K0, K1);
1813 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1815 uasm_l_nopage_tlbm(&l, p);
1816 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1819 if (p >= handle_tlbm_end)
1820 panic("TLB modify handler fastpath space exceeded");
1822 uasm_resolve_relocs(relocs, labels);
1823 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1824 (unsigned int)(p - handle_tlbm));
1826 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1828 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1831 * R4000 style TLB load/store/modify handlers.
1833 static struct work_registers
1834 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1835 struct uasm_reloc **r)
1837 struct work_registers wr = build_get_work_registers(p);
1840 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1842 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1845 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1847 * For huge tlb entries, pmd doesn't contain an address but
1848 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1849 * see if we need to jump to huge tlb processing.
1851 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1854 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1855 UASM_i_LW(p, wr.r2, 0, wr.r2);
1856 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1857 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1858 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1861 uasm_l_smp_pgtable_change(l, *p);
1863 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1864 if (!m4kc_tlbp_war())
1865 build_tlb_probe_entry(p);
1870 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1871 struct uasm_reloc **r, unsigned int tmp,
1874 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1875 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1876 build_update_entries(p, tmp, ptr);
1877 build_tlb_write_entry(p, l, r, tlb_indexed);
1878 uasm_l_leave(l, *p);
1879 build_restore_work_registers(p);
1880 uasm_i_eret(p); /* return from trap */
1883 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1887 static void build_r4000_tlb_load_handler(void)
1889 u32 *p = handle_tlbl;
1890 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1891 struct uasm_label *l = labels;
1892 struct uasm_reloc *r = relocs;
1893 struct work_registers wr;
1895 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1896 memset(labels, 0, sizeof(labels));
1897 memset(relocs, 0, sizeof(relocs));
1899 if (bcm1250_m3_war()) {
1900 unsigned int segbits = 44;
1902 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1903 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1904 uasm_i_xor(&p, K0, K0, K1);
1905 uasm_i_dsrl_safe(&p, K1, K0, 62);
1906 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1907 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1908 uasm_i_or(&p, K0, K0, K1);
1909 uasm_il_bnez(&p, &r, K0, label_leave);
1910 /* No need for uasm_i_nop */
1913 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1914 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1915 if (m4kc_tlbp_war())
1916 build_tlb_probe_entry(&p);
1920 * If the page is not _PAGE_VALID, RI or XI could not
1921 * have triggered it. Skip the expensive test..
1923 if (use_bbit_insns()) {
1924 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1925 label_tlbl_goaround1);
1927 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1928 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1934 switch (current_cpu_type()) {
1936 if (cpu_has_mips_r2) {
1939 case CPU_CAVIUM_OCTEON:
1940 case CPU_CAVIUM_OCTEON_PLUS:
1941 case CPU_CAVIUM_OCTEON2:
1946 /* Examine entrylo 0 or 1 based on ptr. */
1947 if (use_bbit_insns()) {
1948 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1950 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1951 uasm_i_beqz(&p, wr.r3, 8);
1953 /* load it in the delay slot*/
1954 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1955 /* load it if ptr is odd */
1956 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1958 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1959 * XI must have triggered it.
1961 if (use_bbit_insns()) {
1962 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1964 uasm_l_tlbl_goaround1(&l, p);
1966 uasm_i_andi(&p, wr.r3, wr.r3, 2);
1967 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1970 uasm_l_tlbl_goaround1(&l, p);
1972 build_make_valid(&p, &r, wr.r1, wr.r2);
1973 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1975 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1977 * This is the entry point when build_r4000_tlbchange_handler_head
1978 * spots a huge page.
1980 uasm_l_tlb_huge_update(&l, p);
1981 iPTE_LW(&p, wr.r1, wr.r2);
1982 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1983 build_tlb_probe_entry(&p);
1987 * If the page is not _PAGE_VALID, RI or XI could not
1988 * have triggered it. Skip the expensive test..
1990 if (use_bbit_insns()) {
1991 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1992 label_tlbl_goaround2);
1994 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1995 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2001 switch (current_cpu_type()) {
2003 if (cpu_has_mips_r2) {
2006 case CPU_CAVIUM_OCTEON:
2007 case CPU_CAVIUM_OCTEON_PLUS:
2008 case CPU_CAVIUM_OCTEON2:
2013 /* Examine entrylo 0 or 1 based on ptr. */
2014 if (use_bbit_insns()) {
2015 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2017 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2018 uasm_i_beqz(&p, wr.r3, 8);
2020 /* load it in the delay slot*/
2021 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2022 /* load it if ptr is odd */
2023 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2025 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2026 * XI must have triggered it.
2028 if (use_bbit_insns()) {
2029 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2031 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2032 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2034 if (PM_DEFAULT_MASK == 0)
2037 * We clobbered C0_PAGEMASK, restore it. On the other branch
2038 * it is restored in build_huge_tlb_write_entry.
2040 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2042 uasm_l_tlbl_goaround2(&l, p);
2044 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2045 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2048 uasm_l_nopage_tlbl(&l, p);
2049 build_restore_work_registers(&p);
2050 #ifdef CONFIG_CPU_MICROMIPS
2051 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2052 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2053 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2057 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2060 if (p >= handle_tlbl_end)
2061 panic("TLB load handler fastpath space exceeded");
2063 uasm_resolve_relocs(relocs, labels);
2064 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2065 (unsigned int)(p - handle_tlbl));
2067 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2070 static void build_r4000_tlb_store_handler(void)
2072 u32 *p = handle_tlbs;
2073 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2074 struct uasm_label *l = labels;
2075 struct uasm_reloc *r = relocs;
2076 struct work_registers wr;
2078 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2079 memset(labels, 0, sizeof(labels));
2080 memset(relocs, 0, sizeof(relocs));
2082 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2083 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2084 if (m4kc_tlbp_war())
2085 build_tlb_probe_entry(&p);
2086 build_make_write(&p, &r, wr.r1, wr.r2);
2087 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2089 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2091 * This is the entry point when
2092 * build_r4000_tlbchange_handler_head spots a huge page.
2094 uasm_l_tlb_huge_update(&l, p);
2095 iPTE_LW(&p, wr.r1, wr.r2);
2096 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2097 build_tlb_probe_entry(&p);
2098 uasm_i_ori(&p, wr.r1, wr.r1,
2099 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2100 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2103 uasm_l_nopage_tlbs(&l, p);
2104 build_restore_work_registers(&p);
2105 #ifdef CONFIG_CPU_MICROMIPS
2106 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2107 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2108 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2112 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2115 if (p >= handle_tlbs_end)
2116 panic("TLB store handler fastpath space exceeded");
2118 uasm_resolve_relocs(relocs, labels);
2119 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2120 (unsigned int)(p - handle_tlbs));
2122 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2125 static void build_r4000_tlb_modify_handler(void)
2127 u32 *p = handle_tlbm;
2128 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2129 struct uasm_label *l = labels;
2130 struct uasm_reloc *r = relocs;
2131 struct work_registers wr;
2133 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2134 memset(labels, 0, sizeof(labels));
2135 memset(relocs, 0, sizeof(relocs));
2137 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2138 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2139 if (m4kc_tlbp_war())
2140 build_tlb_probe_entry(&p);
2141 /* Present and writable bits set, set accessed and dirty bits. */
2142 build_make_write(&p, &r, wr.r1, wr.r2);
2143 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2145 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2147 * This is the entry point when
2148 * build_r4000_tlbchange_handler_head spots a huge page.
2150 uasm_l_tlb_huge_update(&l, p);
2151 iPTE_LW(&p, wr.r1, wr.r2);
2152 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2153 build_tlb_probe_entry(&p);
2154 uasm_i_ori(&p, wr.r1, wr.r1,
2155 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2156 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2159 uasm_l_nopage_tlbm(&l, p);
2160 build_restore_work_registers(&p);
2161 #ifdef CONFIG_CPU_MICROMIPS
2162 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2163 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2164 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2168 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2171 if (p >= handle_tlbm_end)
2172 panic("TLB modify handler fastpath space exceeded");
2174 uasm_resolve_relocs(relocs, labels);
2175 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2176 (unsigned int)(p - handle_tlbm));
2178 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2181 static void flush_tlb_handlers(void)
2183 local_flush_icache_range((unsigned long)handle_tlbl,
2184 (unsigned long)handle_tlbl_end);
2185 local_flush_icache_range((unsigned long)handle_tlbs,
2186 (unsigned long)handle_tlbs_end);
2187 local_flush_icache_range((unsigned long)handle_tlbm,
2188 (unsigned long)handle_tlbm_end);
2189 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2190 (unsigned long)tlbmiss_handler_setup_pgd_end);
2193 void build_tlb_refill_handler(void)
2196 * The refill handler is generated per-CPU, multi-node systems
2197 * may have local storage for it. The other handlers are only
2200 static int run_once = 0;
2202 output_pgtable_bits_defines();
2205 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2208 switch (current_cpu_type()) {
2216 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2217 if (cpu_has_local_ebase)
2218 build_r3000_tlb_refill_handler();
2220 if (!cpu_has_local_ebase)
2221 build_r3000_tlb_refill_handler();
2223 build_r3000_tlb_load_handler();
2224 build_r3000_tlb_store_handler();
2225 build_r3000_tlb_modify_handler();
2226 flush_tlb_handlers();
2230 panic("No R3000 TLB refill handler");
2236 panic("No R6000 TLB refill handler yet");
2240 panic("No R8000 TLB refill handler yet");
2245 scratch_reg = allocate_kscratch();
2247 build_r4000_tlb_load_handler();
2248 build_r4000_tlb_store_handler();
2249 build_r4000_tlb_modify_handler();
2250 if (!cpu_has_local_ebase)
2251 build_r4000_tlb_refill_handler();
2252 flush_tlb_handlers();
2255 if (cpu_has_local_ebase)
2256 build_r4000_tlb_refill_handler();