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/cache.h>
31 #include <asm/cacheflush.h>
32 #include <asm/cpu-type.h>
33 #include <asm/pgtable.h>
36 #include <asm/setup.h>
39 * TLB load/store/modify handlers.
41 * Only the fastpath gets synthesized at runtime, the slowpath for
42 * do_page_fault remains normal asm.
44 extern void tlb_do_page_fault_0(void);
45 extern void tlb_do_page_fault_1(void);
47 struct work_registers {
56 } ____cacheline_aligned_in_smp;
58 static struct tlb_reg_save handler_reg_save[NR_CPUS];
60 static inline int r45k_bvahwbug(void)
62 /* XXX: We should probe for the presence of this bug, but we don't. */
66 static inline int r4k_250MHZhwbug(void)
68 /* XXX: We should probe for the presence of this bug, but we don't. */
72 static inline int __maybe_unused bcm1250_m3_war(void)
74 return BCM1250_M3_WAR;
77 static inline int __maybe_unused r10000_llsc_war(void)
79 return R10000_LLSC_WAR;
82 static int use_bbit_insns(void)
84 switch (current_cpu_type()) {
85 case CPU_CAVIUM_OCTEON:
86 case CPU_CAVIUM_OCTEON_PLUS:
87 case CPU_CAVIUM_OCTEON2:
88 case CPU_CAVIUM_OCTEON3:
95 static int use_lwx_insns(void)
97 switch (current_cpu_type()) {
98 case CPU_CAVIUM_OCTEON2:
99 case CPU_CAVIUM_OCTEON3:
105 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
106 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
107 static bool scratchpad_available(void)
111 static int scratchpad_offset(int i)
114 * CVMSEG starts at address -32768 and extends for
115 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
117 i += 1; /* Kernel use starts at the top and works down. */
118 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
121 static bool scratchpad_available(void)
125 static int scratchpad_offset(int i)
128 /* Really unreachable, but evidently some GCC want this. */
133 * Found by experiment: At least some revisions of the 4kc throw under
134 * some circumstances a machine check exception, triggered by invalid
135 * values in the index register. Delaying the tlbp instruction until
136 * after the next branch, plus adding an additional nop in front of
137 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
138 * why; it's not an issue caused by the core RTL.
141 static int m4kc_tlbp_war(void)
143 return (current_cpu_data.processor_id & 0xffff00) ==
144 (PRID_COMP_MIPS | PRID_IMP_4KC);
147 /* Handle labels (which must be positive integers). */
149 label_second_part = 1,
154 label_split = label_tlbw_hazard_0 + 8,
155 label_tlbl_goaround1,
156 label_tlbl_goaround2,
160 label_smp_pgtable_change,
161 label_r3000_write_probe_fail,
162 label_large_segbits_fault,
163 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
164 label_tlb_huge_update,
168 UASM_L_LA(_second_part)
171 UASM_L_LA(_vmalloc_done)
172 /* _tlbw_hazard_x is handled differently. */
174 UASM_L_LA(_tlbl_goaround1)
175 UASM_L_LA(_tlbl_goaround2)
176 UASM_L_LA(_nopage_tlbl)
177 UASM_L_LA(_nopage_tlbs)
178 UASM_L_LA(_nopage_tlbm)
179 UASM_L_LA(_smp_pgtable_change)
180 UASM_L_LA(_r3000_write_probe_fail)
181 UASM_L_LA(_large_segbits_fault)
182 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
183 UASM_L_LA(_tlb_huge_update)
186 static int hazard_instance;
188 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
192 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
199 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
203 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
211 * pgtable bits are assigned dynamically depending on processor feature
212 * and statically based on kernel configuration. This spits out the actual
213 * values the kernel is using. Required to make sense from disassembled
214 * TLB exception handlers.
216 static void output_pgtable_bits_defines(void)
218 #define pr_define(fmt, ...) \
219 pr_debug("#define " fmt, ##__VA_ARGS__)
221 pr_debug("#include <asm/asm.h>\n");
222 pr_debug("#include <asm/regdef.h>\n");
225 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
226 pr_define("_PAGE_READ_SHIFT %d\n", _PAGE_READ_SHIFT);
227 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
228 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
229 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
230 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
231 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
232 pr_define("_PAGE_SPLITTING_SHIFT %d\n", _PAGE_SPLITTING_SHIFT);
235 #ifdef _PAGE_NO_EXEC_SHIFT
236 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
238 #ifdef _PAGE_NO_READ_SHIFT
239 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
242 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
243 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
244 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
245 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
249 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
253 pr_debug("LEAF(%s)\n", symbol);
255 pr_debug("\t.set push\n");
256 pr_debug("\t.set noreorder\n");
258 for (i = 0; i < count; i++)
259 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
261 pr_debug("\t.set\tpop\n");
263 pr_debug("\tEND(%s)\n", symbol);
266 /* The only general purpose registers allowed in TLB handlers. */
270 /* Some CP0 registers */
271 #define C0_INDEX 0, 0
272 #define C0_ENTRYLO0 2, 0
273 #define C0_TCBIND 2, 2
274 #define C0_ENTRYLO1 3, 0
275 #define C0_CONTEXT 4, 0
276 #define C0_PAGEMASK 5, 0
277 #define C0_BADVADDR 8, 0
278 #define C0_ENTRYHI 10, 0
280 #define C0_XCONTEXT 20, 0
283 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
285 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
288 /* The worst case length of the handler is around 18 instructions for
289 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
290 * Maximum space available is 32 instructions for R3000 and 64
291 * instructions for R4000.
293 * We deliberately chose a buffer size of 128, so we won't scribble
294 * over anything important on overflow before we panic.
296 static u32 tlb_handler[128];
298 /* simply assume worst case size for labels and relocs */
299 static struct uasm_label labels[128];
300 static struct uasm_reloc relocs[128];
302 static int check_for_high_segbits;
304 static unsigned int kscratch_used_mask;
306 static inline int __maybe_unused c0_kscratch(void)
308 switch (current_cpu_type()) {
317 static int allocate_kscratch(void)
320 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
327 r--; /* make it zero based */
329 kscratch_used_mask |= (1 << r);
334 static int scratch_reg;
336 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
338 static struct work_registers build_get_work_registers(u32 **p)
340 struct work_registers r;
342 if (scratch_reg >= 0) {
343 /* Save in CPU local C0_KScratch? */
344 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
351 if (num_possible_cpus() > 1) {
352 /* Get smp_processor_id */
353 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
354 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
356 /* handler_reg_save index in K0 */
357 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
359 UASM_i_LA(p, K1, (long)&handler_reg_save);
360 UASM_i_ADDU(p, K0, K0, K1);
362 UASM_i_LA(p, K0, (long)&handler_reg_save);
364 /* K0 now points to save area, save $1 and $2 */
365 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
366 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
374 static void build_restore_work_registers(u32 **p)
376 if (scratch_reg >= 0) {
377 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
380 /* K0 already points to save area, restore $1 and $2 */
381 UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
382 UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
385 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
388 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
389 * we cannot do r3000 under these circumstances.
391 * Declare pgd_current here instead of including mmu_context.h to avoid type
392 * conflicts for tlbmiss_handler_setup_pgd
394 extern unsigned long pgd_current[];
397 * The R3000 TLB handler is simple.
399 static void build_r3000_tlb_refill_handler(void)
401 long pgdc = (long)pgd_current;
404 memset(tlb_handler, 0, sizeof(tlb_handler));
407 uasm_i_mfc0(&p, K0, C0_BADVADDR);
408 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
409 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
410 uasm_i_srl(&p, K0, K0, 22); /* load delay */
411 uasm_i_sll(&p, K0, K0, 2);
412 uasm_i_addu(&p, K1, K1, K0);
413 uasm_i_mfc0(&p, K0, C0_CONTEXT);
414 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
415 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
416 uasm_i_addu(&p, K1, K1, K0);
417 uasm_i_lw(&p, K0, 0, K1);
418 uasm_i_nop(&p); /* load delay */
419 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
420 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
421 uasm_i_tlbwr(&p); /* cp0 delay */
423 uasm_i_rfe(&p); /* branch delay */
425 if (p > tlb_handler + 32)
426 panic("TLB refill handler space exceeded");
428 pr_debug("Wrote TLB refill handler (%u instructions).\n",
429 (unsigned int)(p - tlb_handler));
431 memcpy((void *)ebase, tlb_handler, 0x80);
433 dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
435 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
438 * The R4000 TLB handler is much more complicated. We have two
439 * consecutive handler areas with 32 instructions space each.
440 * Since they aren't used at the same time, we can overflow in the
441 * other one.To keep things simple, we first assume linear space,
442 * then we relocate it to the final handler layout as needed.
444 static u32 final_handler[64];
449 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
450 * 2. A timing hazard exists for the TLBP instruction.
452 * stalling_instruction
455 * The JTLB is being read for the TLBP throughout the stall generated by the
456 * previous instruction. This is not really correct as the stalling instruction
457 * can modify the address used to access the JTLB. The failure symptom is that
458 * the TLBP instruction will use an address created for the stalling instruction
459 * and not the address held in C0_ENHI and thus report the wrong results.
461 * The software work-around is to not allow the instruction preceding the TLBP
462 * to stall - make it an NOP or some other instruction guaranteed not to stall.
464 * Errata 2 will not be fixed. This errata is also on the R5000.
466 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
468 static void __maybe_unused build_tlb_probe_entry(u32 **p)
470 switch (current_cpu_type()) {
471 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
487 * Write random or indexed TLB entry, and care about the hazards from
488 * the preceding mtc0 and for the following eret.
490 enum tlb_write_entry { tlb_random, tlb_indexed };
492 static void build_tlb_write_entry(u32 **p, struct uasm_label **l,
493 struct uasm_reloc **r,
494 enum tlb_write_entry wmode)
496 void(*tlbw)(u32 **) = NULL;
499 case tlb_random: tlbw = uasm_i_tlbwr; break;
500 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
503 if (cpu_has_mips_r2) {
505 * The architecture spec says an ehb is required here,
506 * but a number of cores do not have the hazard and
507 * using an ehb causes an expensive pipeline stall.
509 switch (current_cpu_type()) {
525 switch (current_cpu_type()) {
533 * This branch uses up a mtc0 hazard nop slot and saves
534 * two nops after the tlbw instruction.
536 uasm_bgezl_hazard(p, r, hazard_instance);
538 uasm_bgezl_label(l, p, hazard_instance);
552 uasm_i_nop(p); /* QED specifies 2 nops hazard */
553 uasm_i_nop(p); /* QED specifies 2 nops hazard */
625 panic("No TLB refill handler yet (CPU type: %d)",
626 current_cpu_data.cputype);
631 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
635 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
637 #ifdef CONFIG_64BIT_PHYS_ADDR
638 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
640 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
645 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
647 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
648 unsigned int tmp, enum label_id lid,
651 if (restore_scratch) {
652 /* Reset default page size */
653 if (PM_DEFAULT_MASK >> 16) {
654 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
655 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
656 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
657 uasm_il_b(p, r, lid);
658 } else if (PM_DEFAULT_MASK) {
659 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
660 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
661 uasm_il_b(p, r, lid);
663 uasm_i_mtc0(p, 0, C0_PAGEMASK);
664 uasm_il_b(p, r, lid);
666 if (scratch_reg >= 0)
667 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
669 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
671 /* Reset default page size */
672 if (PM_DEFAULT_MASK >> 16) {
673 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
674 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
675 uasm_il_b(p, r, lid);
676 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
677 } else if (PM_DEFAULT_MASK) {
678 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
679 uasm_il_b(p, r, lid);
680 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
682 uasm_il_b(p, r, lid);
683 uasm_i_mtc0(p, 0, C0_PAGEMASK);
688 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
689 struct uasm_reloc **r,
691 enum tlb_write_entry wmode,
694 /* Set huge page tlb entry size */
695 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
696 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
697 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
699 build_tlb_write_entry(p, l, r, wmode);
701 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
705 * Check if Huge PTE is present, if so then jump to LABEL.
708 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
709 unsigned int pmd, int lid)
711 UASM_i_LW(p, tmp, 0, pmd);
712 if (use_bbit_insns()) {
713 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
715 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
716 uasm_il_bnez(p, r, tmp, lid);
720 static void build_huge_update_entries(u32 **p, unsigned int pte,
726 * A huge PTE describes an area the size of the
727 * configured huge page size. This is twice the
728 * of the large TLB entry size we intend to use.
729 * A TLB entry half the size of the configured
730 * huge page size is configured into entrylo0
731 * and entrylo1 to cover the contiguous huge PTE
734 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
736 /* We can clobber tmp. It isn't used after this.*/
738 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
740 build_convert_pte_to_entrylo(p, pte);
741 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
742 /* convert to entrylo1 */
744 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
746 UASM_i_ADDU(p, pte, pte, tmp);
748 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
751 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
752 struct uasm_label **l,
757 UASM_i_SC(p, pte, 0, ptr);
758 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
759 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
761 UASM_i_SW(p, pte, 0, ptr);
763 build_huge_update_entries(p, pte, ptr);
764 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
766 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
770 * TMP and PTR are scratch.
771 * TMP will be clobbered, PTR will hold the pmd entry.
774 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
775 unsigned int tmp, unsigned int ptr)
777 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
778 long pgdc = (long)pgd_current;
781 * The vmalloc handling is not in the hotpath.
783 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
785 if (check_for_high_segbits) {
787 * The kernel currently implicitely assumes that the
788 * MIPS SEGBITS parameter for the processor is
789 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
790 * allocate virtual addresses outside the maximum
791 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
792 * that doesn't prevent user code from accessing the
793 * higher xuseg addresses. Here, we make sure that
794 * everything but the lower xuseg addresses goes down
795 * the module_alloc/vmalloc path.
797 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
798 uasm_il_bnez(p, r, ptr, label_vmalloc);
800 uasm_il_bltz(p, r, tmp, label_vmalloc);
802 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
805 /* pgd is in pgd_reg */
806 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
808 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
810 * &pgd << 11 stored in CONTEXT [23..63].
812 UASM_i_MFC0(p, ptr, C0_CONTEXT);
814 /* Clear lower 23 bits of context. */
815 uasm_i_dins(p, ptr, 0, 0, 23);
817 /* 1 0 1 0 1 << 6 xkphys cached */
818 uasm_i_ori(p, ptr, ptr, 0x540);
819 uasm_i_drotr(p, ptr, ptr, 11);
820 #elif defined(CONFIG_SMP)
821 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
822 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
823 UASM_i_LA_mostly(p, tmp, pgdc);
824 uasm_i_daddu(p, ptr, ptr, tmp);
825 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
826 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
828 UASM_i_LA_mostly(p, ptr, pgdc);
829 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
833 uasm_l_vmalloc_done(l, *p);
835 /* get pgd offset in bytes */
836 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
838 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
839 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
840 #ifndef __PAGETABLE_PMD_FOLDED
841 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
842 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
843 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
844 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
845 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
850 * BVADDR is the faulting address, PTR is scratch.
851 * PTR will hold the pgd for vmalloc.
854 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
855 unsigned int bvaddr, unsigned int ptr,
856 enum vmalloc64_mode mode)
858 long swpd = (long)swapper_pg_dir;
859 int single_insn_swpd;
860 int did_vmalloc_branch = 0;
862 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
864 uasm_l_vmalloc(l, *p);
866 if (mode != not_refill && check_for_high_segbits) {
867 if (single_insn_swpd) {
868 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
869 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
870 did_vmalloc_branch = 1;
873 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
876 if (!did_vmalloc_branch) {
877 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
878 uasm_il_b(p, r, label_vmalloc_done);
879 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
881 UASM_i_LA_mostly(p, ptr, swpd);
882 uasm_il_b(p, r, label_vmalloc_done);
883 if (uasm_in_compat_space_p(swpd))
884 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
886 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
889 if (mode != not_refill && check_for_high_segbits) {
890 uasm_l_large_segbits_fault(l, *p);
892 * We get here if we are an xsseg address, or if we are
893 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
895 * Ignoring xsseg (assume disabled so would generate
896 * (address errors?), the only remaining possibility
897 * is the upper xuseg addresses. On processors with
898 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
899 * addresses would have taken an address error. We try
900 * to mimic that here by taking a load/istream page
903 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
906 if (mode == refill_scratch) {
907 if (scratch_reg >= 0)
908 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
910 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
917 #else /* !CONFIG_64BIT */
920 * TMP and PTR are scratch.
921 * TMP will be clobbered, PTR will hold the pgd entry.
923 static void __maybe_unused
924 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
927 /* pgd is in pgd_reg */
928 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
929 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
931 long pgdc = (long)pgd_current;
933 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
935 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
936 UASM_i_LA_mostly(p, tmp, pgdc);
937 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
938 uasm_i_addu(p, ptr, tmp, ptr);
940 UASM_i_LA_mostly(p, ptr, pgdc);
942 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
943 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
945 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
946 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
947 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
950 #endif /* !CONFIG_64BIT */
952 static void build_adjust_context(u32 **p, unsigned int ctx)
954 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
955 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
957 switch (current_cpu_type()) {
974 UASM_i_SRL(p, ctx, ctx, shift);
975 uasm_i_andi(p, ctx, ctx, mask);
978 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
981 * Bug workaround for the Nevada. It seems as if under certain
982 * circumstances the move from cp0_context might produce a
983 * bogus result when the mfc0 instruction and its consumer are
984 * in a different cacheline or a load instruction, probably any
985 * memory reference, is between them.
987 switch (current_cpu_type()) {
989 UASM_i_LW(p, ptr, 0, ptr);
990 GET_CONTEXT(p, tmp); /* get context reg */
994 GET_CONTEXT(p, tmp); /* get context reg */
995 UASM_i_LW(p, ptr, 0, ptr);
999 build_adjust_context(p, tmp);
1000 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1003 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1006 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1007 * Kernel is a special case. Only a few CPUs use it.
1009 #ifdef CONFIG_64BIT_PHYS_ADDR
1010 if (cpu_has_64bits) {
1011 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1012 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1014 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1015 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1016 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1018 uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1019 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1020 uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1022 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1024 int pte_off_even = sizeof(pte_t) / 2;
1025 int pte_off_odd = pte_off_even + sizeof(pte_t);
1027 /* The pte entries are pre-shifted */
1028 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1029 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1030 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1031 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1034 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1035 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1036 if (r45k_bvahwbug())
1037 build_tlb_probe_entry(p);
1039 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1040 if (r4k_250MHZhwbug())
1041 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1042 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1043 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1045 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1046 if (r4k_250MHZhwbug())
1047 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1048 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1049 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1050 if (r45k_bvahwbug())
1051 uasm_i_mfc0(p, tmp, C0_INDEX);
1053 if (r4k_250MHZhwbug())
1054 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1055 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1059 struct mips_huge_tlb_info {
1061 int restore_scratch;
1064 static struct mips_huge_tlb_info
1065 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1066 struct uasm_reloc **r, unsigned int tmp,
1067 unsigned int ptr, int c0_scratch_reg)
1069 struct mips_huge_tlb_info rv;
1070 unsigned int even, odd;
1071 int vmalloc_branch_delay_filled = 0;
1072 const int scratch = 1; /* Our extra working register */
1074 rv.huge_pte = scratch;
1075 rv.restore_scratch = 0;
1077 if (check_for_high_segbits) {
1078 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1081 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1083 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1085 if (c0_scratch_reg >= 0)
1086 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1088 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1090 uasm_i_dsrl_safe(p, scratch, tmp,
1091 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1092 uasm_il_bnez(p, r, scratch, label_vmalloc);
1094 if (pgd_reg == -1) {
1095 vmalloc_branch_delay_filled = 1;
1096 /* Clear lower 23 bits of context. */
1097 uasm_i_dins(p, ptr, 0, 0, 23);
1101 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1103 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1105 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1107 if (c0_scratch_reg >= 0)
1108 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1110 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1113 /* Clear lower 23 bits of context. */
1114 uasm_i_dins(p, ptr, 0, 0, 23);
1116 uasm_il_bltz(p, r, tmp, label_vmalloc);
1119 if (pgd_reg == -1) {
1120 vmalloc_branch_delay_filled = 1;
1121 /* 1 0 1 0 1 << 6 xkphys cached */
1122 uasm_i_ori(p, ptr, ptr, 0x540);
1123 uasm_i_drotr(p, ptr, ptr, 11);
1126 #ifdef __PAGETABLE_PMD_FOLDED
1127 #define LOC_PTEP scratch
1129 #define LOC_PTEP ptr
1132 if (!vmalloc_branch_delay_filled)
1133 /* get pgd offset in bytes */
1134 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1136 uasm_l_vmalloc_done(l, *p);
1140 * fall-through case = badvaddr *pgd_current
1141 * vmalloc case = badvaddr swapper_pg_dir
1144 if (vmalloc_branch_delay_filled)
1145 /* get pgd offset in bytes */
1146 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1148 #ifdef __PAGETABLE_PMD_FOLDED
1149 GET_CONTEXT(p, tmp); /* get context reg */
1151 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1153 if (use_lwx_insns()) {
1154 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1156 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1157 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1160 #ifndef __PAGETABLE_PMD_FOLDED
1161 /* get pmd offset in bytes */
1162 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1163 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1164 GET_CONTEXT(p, tmp); /* get context reg */
1166 if (use_lwx_insns()) {
1167 UASM_i_LWX(p, scratch, scratch, ptr);
1169 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1170 UASM_i_LW(p, scratch, 0, ptr);
1173 /* Adjust the context during the load latency. */
1174 build_adjust_context(p, tmp);
1176 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1177 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1179 * The in the LWX case we don't want to do the load in the
1180 * delay slot. It cannot issue in the same cycle and may be
1181 * speculative and unneeded.
1183 if (use_lwx_insns())
1185 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1188 /* build_update_entries */
1189 if (use_lwx_insns()) {
1192 UASM_i_LWX(p, even, scratch, tmp);
1193 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1194 UASM_i_LWX(p, odd, scratch, tmp);
1196 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1199 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1200 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1203 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1204 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1205 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1207 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1208 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1209 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1211 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1213 if (c0_scratch_reg >= 0) {
1214 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1215 build_tlb_write_entry(p, l, r, tlb_random);
1216 uasm_l_leave(l, *p);
1217 rv.restore_scratch = 1;
1218 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1219 build_tlb_write_entry(p, l, r, tlb_random);
1220 uasm_l_leave(l, *p);
1221 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1223 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1224 build_tlb_write_entry(p, l, r, tlb_random);
1225 uasm_l_leave(l, *p);
1226 rv.restore_scratch = 1;
1229 uasm_i_eret(p); /* return from trap */
1235 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1236 * because EXL == 0. If we wrap, we can also use the 32 instruction
1237 * slots before the XTLB refill exception handler which belong to the
1238 * unused TLB refill exception.
1240 #define MIPS64_REFILL_INSNS 32
1242 static void build_r4000_tlb_refill_handler(void)
1244 u32 *p = tlb_handler;
1245 struct uasm_label *l = labels;
1246 struct uasm_reloc *r = relocs;
1248 unsigned int final_len;
1249 struct mips_huge_tlb_info htlb_info __maybe_unused;
1250 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1252 memset(tlb_handler, 0, sizeof(tlb_handler));
1253 memset(labels, 0, sizeof(labels));
1254 memset(relocs, 0, sizeof(relocs));
1255 memset(final_handler, 0, sizeof(final_handler));
1257 if ((scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1258 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1260 vmalloc_mode = refill_scratch;
1262 htlb_info.huge_pte = K0;
1263 htlb_info.restore_scratch = 0;
1264 vmalloc_mode = refill_noscratch;
1266 * create the plain linear handler
1268 if (bcm1250_m3_war()) {
1269 unsigned int segbits = 44;
1271 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1272 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1273 uasm_i_xor(&p, K0, K0, K1);
1274 uasm_i_dsrl_safe(&p, K1, K0, 62);
1275 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1276 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1277 uasm_i_or(&p, K0, K0, K1);
1278 uasm_il_bnez(&p, &r, K0, label_leave);
1279 /* No need for uasm_i_nop */
1283 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1285 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1288 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1289 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1292 build_get_ptep(&p, K0, K1);
1293 build_update_entries(&p, K0, K1);
1294 build_tlb_write_entry(&p, &l, &r, tlb_random);
1295 uasm_l_leave(&l, p);
1296 uasm_i_eret(&p); /* return from trap */
1298 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1299 uasm_l_tlb_huge_update(&l, p);
1300 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1301 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1302 htlb_info.restore_scratch);
1306 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1310 * Overflow check: For the 64bit handler, we need at least one
1311 * free instruction slot for the wrap-around branch. In worst
1312 * case, if the intended insertion point is a delay slot, we
1313 * need three, with the second nop'ed and the third being
1316 switch (boot_cpu_type()) {
1318 if (sizeof(long) == 4) {
1320 /* Loongson2 ebase is different than r4k, we have more space */
1321 if ((p - tlb_handler) > 64)
1322 panic("TLB refill handler space exceeded");
1324 * Now fold the handler in the TLB refill handler space.
1327 /* Simplest case, just copy the handler. */
1328 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1329 final_len = p - tlb_handler;
1332 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1333 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1334 && uasm_insn_has_bdelay(relocs,
1335 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1336 panic("TLB refill handler space exceeded");
1338 * Now fold the handler in the TLB refill handler space.
1340 f = final_handler + MIPS64_REFILL_INSNS;
1341 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1342 /* Just copy the handler. */
1343 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1344 final_len = p - tlb_handler;
1346 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1347 const enum label_id ls = label_tlb_huge_update;
1349 const enum label_id ls = label_vmalloc;
1355 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1357 BUG_ON(i == ARRAY_SIZE(labels));
1358 split = labels[i].addr;
1361 * See if we have overflown one way or the other.
1363 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1364 split < p - MIPS64_REFILL_INSNS)
1369 * Split two instructions before the end. One
1370 * for the branch and one for the instruction
1371 * in the delay slot.
1373 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1376 * If the branch would fall in a delay slot,
1377 * we must back up an additional instruction
1378 * so that it is no longer in a delay slot.
1380 if (uasm_insn_has_bdelay(relocs, split - 1))
1383 /* Copy first part of the handler. */
1384 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1385 f += split - tlb_handler;
1388 /* Insert branch. */
1389 uasm_l_split(&l, final_handler);
1390 uasm_il_b(&f, &r, label_split);
1391 if (uasm_insn_has_bdelay(relocs, split))
1394 uasm_copy_handler(relocs, labels,
1395 split, split + 1, f);
1396 uasm_move_labels(labels, f, f + 1, -1);
1402 /* Copy the rest of the handler. */
1403 uasm_copy_handler(relocs, labels, split, p, final_handler);
1404 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1411 uasm_resolve_relocs(relocs, labels);
1412 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1415 memcpy((void *)ebase, final_handler, 0x100);
1417 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1420 extern u32 handle_tlbl[], handle_tlbl_end[];
1421 extern u32 handle_tlbs[], handle_tlbs_end[];
1422 extern u32 handle_tlbm[], handle_tlbm_end[];
1423 extern u32 tlbmiss_handler_setup_pgd[], tlbmiss_handler_setup_pgd_end[];
1425 static void build_setup_pgd(void)
1428 const int __maybe_unused a1 = 5;
1429 const int __maybe_unused a2 = 6;
1430 u32 *p = tlbmiss_handler_setup_pgd;
1431 const int tlbmiss_handler_setup_pgd_size =
1432 tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd;
1433 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1434 long pgdc = (long)pgd_current;
1437 memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1438 sizeof(tlbmiss_handler_setup_pgd[0]));
1439 memset(labels, 0, sizeof(labels));
1440 memset(relocs, 0, sizeof(relocs));
1441 pgd_reg = allocate_kscratch();
1442 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1443 if (pgd_reg == -1) {
1444 struct uasm_label *l = labels;
1445 struct uasm_reloc *r = relocs;
1447 /* PGD << 11 in c0_Context */
1449 * If it is a ckseg0 address, convert to a physical
1450 * address. Shifting right by 29 and adding 4 will
1451 * result in zero for these addresses.
1454 UASM_i_SRA(&p, a1, a0, 29);
1455 UASM_i_ADDIU(&p, a1, a1, 4);
1456 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1458 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1459 uasm_l_tlbl_goaround1(&l, p);
1460 UASM_i_SLL(&p, a0, a0, 11);
1462 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1464 /* PGD in c0_KScratch */
1466 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1470 /* Save PGD to pgd_current[smp_processor_id()] */
1471 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1472 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1473 UASM_i_LA_mostly(&p, a2, pgdc);
1474 UASM_i_ADDU(&p, a2, a2, a1);
1475 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1477 UASM_i_LA_mostly(&p, a2, pgdc);
1478 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1482 /* if pgd_reg is allocated, save PGD also to scratch register */
1484 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1488 if (p >= tlbmiss_handler_setup_pgd_end)
1489 panic("tlbmiss_handler_setup_pgd space exceeded");
1491 uasm_resolve_relocs(relocs, labels);
1492 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1493 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1495 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1496 tlbmiss_handler_setup_pgd_size);
1500 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1503 # ifdef CONFIG_64BIT_PHYS_ADDR
1505 uasm_i_lld(p, pte, 0, ptr);
1508 UASM_i_LL(p, pte, 0, ptr);
1510 # ifdef CONFIG_64BIT_PHYS_ADDR
1512 uasm_i_ld(p, pte, 0, ptr);
1515 UASM_i_LW(p, pte, 0, ptr);
1520 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1523 #ifdef CONFIG_64BIT_PHYS_ADDR
1524 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1527 uasm_i_ori(p, pte, pte, mode);
1529 # ifdef CONFIG_64BIT_PHYS_ADDR
1531 uasm_i_scd(p, pte, 0, ptr);
1534 UASM_i_SC(p, pte, 0, ptr);
1536 if (r10000_llsc_war())
1537 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1539 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1541 # ifdef CONFIG_64BIT_PHYS_ADDR
1542 if (!cpu_has_64bits) {
1543 /* no uasm_i_nop needed */
1544 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1545 uasm_i_ori(p, pte, pte, hwmode);
1546 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1547 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1548 /* no uasm_i_nop needed */
1549 uasm_i_lw(p, pte, 0, ptr);
1556 # ifdef CONFIG_64BIT_PHYS_ADDR
1558 uasm_i_sd(p, pte, 0, ptr);
1561 UASM_i_SW(p, pte, 0, ptr);
1563 # ifdef CONFIG_64BIT_PHYS_ADDR
1564 if (!cpu_has_64bits) {
1565 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1566 uasm_i_ori(p, pte, pte, hwmode);
1567 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1568 uasm_i_lw(p, pte, 0, ptr);
1575 * Check if PTE is present, if not then jump to LABEL. PTR points to
1576 * the page table where this PTE is located, PTE will be re-loaded
1577 * with it's original value.
1580 build_pte_present(u32 **p, struct uasm_reloc **r,
1581 int pte, int ptr, int scratch, enum label_id lid)
1583 int t = scratch >= 0 ? scratch : pte;
1586 if (use_bbit_insns()) {
1587 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1590 uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1591 uasm_il_beqz(p, r, t, lid);
1593 /* You lose the SMP race :-(*/
1594 iPTE_LW(p, pte, ptr);
1597 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1598 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1599 uasm_il_bnez(p, r, t, lid);
1601 /* You lose the SMP race :-(*/
1602 iPTE_LW(p, pte, ptr);
1606 /* Make PTE valid, store result in PTR. */
1608 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1611 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1613 iPTE_SW(p, r, pte, ptr, mode);
1617 * Check if PTE can be written to, if not branch to LABEL. Regardless
1618 * restore PTE with value from PTR when done.
1621 build_pte_writable(u32 **p, struct uasm_reloc **r,
1622 unsigned int pte, unsigned int ptr, int scratch,
1625 int t = scratch >= 0 ? scratch : pte;
1627 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1628 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1629 uasm_il_bnez(p, r, t, lid);
1631 /* You lose the SMP race :-(*/
1632 iPTE_LW(p, pte, ptr);
1637 /* Make PTE writable, update software status bits as well, then store
1641 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1644 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1647 iPTE_SW(p, r, pte, ptr, mode);
1651 * Check if PTE can be modified, if not branch to LABEL. Regardless
1652 * restore PTE with value from PTR when done.
1655 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1656 unsigned int pte, unsigned int ptr, int scratch,
1659 if (use_bbit_insns()) {
1660 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1663 int t = scratch >= 0 ? scratch : pte;
1664 uasm_i_andi(p, t, pte, _PAGE_WRITE);
1665 uasm_il_beqz(p, r, t, lid);
1667 /* You lose the SMP race :-(*/
1668 iPTE_LW(p, pte, ptr);
1672 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1676 * R3000 style TLB load/store/modify handlers.
1680 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1684 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1686 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1687 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1690 uasm_i_rfe(p); /* branch delay */
1694 * This places the pte into ENTRYLO0 and writes it with tlbwi
1695 * or tlbwr as appropriate. This is because the index register
1696 * may have the probe fail bit set as a result of a trap on a
1697 * kseg2 access, i.e. without refill. Then it returns.
1700 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1701 struct uasm_reloc **r, unsigned int pte,
1704 uasm_i_mfc0(p, tmp, C0_INDEX);
1705 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1706 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1707 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1708 uasm_i_tlbwi(p); /* cp0 delay */
1710 uasm_i_rfe(p); /* branch delay */
1711 uasm_l_r3000_write_probe_fail(l, *p);
1712 uasm_i_tlbwr(p); /* cp0 delay */
1714 uasm_i_rfe(p); /* branch delay */
1718 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1721 long pgdc = (long)pgd_current;
1723 uasm_i_mfc0(p, pte, C0_BADVADDR);
1724 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1725 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1726 uasm_i_srl(p, pte, pte, 22); /* load delay */
1727 uasm_i_sll(p, pte, pte, 2);
1728 uasm_i_addu(p, ptr, ptr, pte);
1729 uasm_i_mfc0(p, pte, C0_CONTEXT);
1730 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1731 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1732 uasm_i_addu(p, ptr, ptr, pte);
1733 uasm_i_lw(p, pte, 0, ptr);
1734 uasm_i_tlbp(p); /* load delay */
1737 static void build_r3000_tlb_load_handler(void)
1739 u32 *p = handle_tlbl;
1740 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1741 struct uasm_label *l = labels;
1742 struct uasm_reloc *r = relocs;
1744 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1745 memset(labels, 0, sizeof(labels));
1746 memset(relocs, 0, sizeof(relocs));
1748 build_r3000_tlbchange_handler_head(&p, K0, K1);
1749 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1750 uasm_i_nop(&p); /* load delay */
1751 build_make_valid(&p, &r, K0, K1);
1752 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1754 uasm_l_nopage_tlbl(&l, p);
1755 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1758 if (p >= handle_tlbl_end)
1759 panic("TLB load handler fastpath space exceeded");
1761 uasm_resolve_relocs(relocs, labels);
1762 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1763 (unsigned int)(p - handle_tlbl));
1765 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1768 static void build_r3000_tlb_store_handler(void)
1770 u32 *p = handle_tlbs;
1771 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1772 struct uasm_label *l = labels;
1773 struct uasm_reloc *r = relocs;
1775 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1776 memset(labels, 0, sizeof(labels));
1777 memset(relocs, 0, sizeof(relocs));
1779 build_r3000_tlbchange_handler_head(&p, K0, K1);
1780 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1781 uasm_i_nop(&p); /* load delay */
1782 build_make_write(&p, &r, K0, K1);
1783 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1785 uasm_l_nopage_tlbs(&l, p);
1786 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1789 if (p >= handle_tlbs_end)
1790 panic("TLB store handler fastpath space exceeded");
1792 uasm_resolve_relocs(relocs, labels);
1793 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1794 (unsigned int)(p - handle_tlbs));
1796 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1799 static void build_r3000_tlb_modify_handler(void)
1801 u32 *p = handle_tlbm;
1802 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1803 struct uasm_label *l = labels;
1804 struct uasm_reloc *r = relocs;
1806 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1807 memset(labels, 0, sizeof(labels));
1808 memset(relocs, 0, sizeof(relocs));
1810 build_r3000_tlbchange_handler_head(&p, K0, K1);
1811 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1812 uasm_i_nop(&p); /* load delay */
1813 build_make_write(&p, &r, K0, K1);
1814 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1816 uasm_l_nopage_tlbm(&l, p);
1817 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1820 if (p >= handle_tlbm_end)
1821 panic("TLB modify handler fastpath space exceeded");
1823 uasm_resolve_relocs(relocs, labels);
1824 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1825 (unsigned int)(p - handle_tlbm));
1827 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1829 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1832 * R4000 style TLB load/store/modify handlers.
1834 static struct work_registers
1835 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1836 struct uasm_reloc **r)
1838 struct work_registers wr = build_get_work_registers(p);
1841 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1843 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1846 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1848 * For huge tlb entries, pmd doesn't contain an address but
1849 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1850 * see if we need to jump to huge tlb processing.
1852 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1855 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1856 UASM_i_LW(p, wr.r2, 0, wr.r2);
1857 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1858 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1859 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1862 uasm_l_smp_pgtable_change(l, *p);
1864 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1865 if (!m4kc_tlbp_war())
1866 build_tlb_probe_entry(p);
1871 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1872 struct uasm_reloc **r, unsigned int tmp,
1875 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1876 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1877 build_update_entries(p, tmp, ptr);
1878 build_tlb_write_entry(p, l, r, tlb_indexed);
1879 uasm_l_leave(l, *p);
1880 build_restore_work_registers(p);
1881 uasm_i_eret(p); /* return from trap */
1884 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1888 static void build_r4000_tlb_load_handler(void)
1890 u32 *p = handle_tlbl;
1891 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1892 struct uasm_label *l = labels;
1893 struct uasm_reloc *r = relocs;
1894 struct work_registers wr;
1896 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1897 memset(labels, 0, sizeof(labels));
1898 memset(relocs, 0, sizeof(relocs));
1900 if (bcm1250_m3_war()) {
1901 unsigned int segbits = 44;
1903 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1904 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1905 uasm_i_xor(&p, K0, K0, K1);
1906 uasm_i_dsrl_safe(&p, K1, K0, 62);
1907 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1908 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1909 uasm_i_or(&p, K0, K0, K1);
1910 uasm_il_bnez(&p, &r, K0, label_leave);
1911 /* No need for uasm_i_nop */
1914 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1915 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1916 if (m4kc_tlbp_war())
1917 build_tlb_probe_entry(&p);
1921 * If the page is not _PAGE_VALID, RI or XI could not
1922 * have triggered it. Skip the expensive test..
1924 if (use_bbit_insns()) {
1925 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1926 label_tlbl_goaround1);
1928 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1929 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1935 switch (current_cpu_type()) {
1937 if (cpu_has_mips_r2) {
1940 case CPU_CAVIUM_OCTEON:
1941 case CPU_CAVIUM_OCTEON_PLUS:
1942 case CPU_CAVIUM_OCTEON2:
1947 /* Examine entrylo 0 or 1 based on ptr. */
1948 if (use_bbit_insns()) {
1949 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1951 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1952 uasm_i_beqz(&p, wr.r3, 8);
1954 /* load it in the delay slot*/
1955 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1956 /* load it if ptr is odd */
1957 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1959 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1960 * XI must have triggered it.
1962 if (use_bbit_insns()) {
1963 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1965 uasm_l_tlbl_goaround1(&l, p);
1967 uasm_i_andi(&p, wr.r3, wr.r3, 2);
1968 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1971 uasm_l_tlbl_goaround1(&l, p);
1973 build_make_valid(&p, &r, wr.r1, wr.r2);
1974 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1976 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1978 * This is the entry point when build_r4000_tlbchange_handler_head
1979 * spots a huge page.
1981 uasm_l_tlb_huge_update(&l, p);
1982 iPTE_LW(&p, wr.r1, wr.r2);
1983 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1984 build_tlb_probe_entry(&p);
1988 * If the page is not _PAGE_VALID, RI or XI could not
1989 * have triggered it. Skip the expensive test..
1991 if (use_bbit_insns()) {
1992 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1993 label_tlbl_goaround2);
1995 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1996 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2002 switch (current_cpu_type()) {
2004 if (cpu_has_mips_r2) {
2007 case CPU_CAVIUM_OCTEON:
2008 case CPU_CAVIUM_OCTEON_PLUS:
2009 case CPU_CAVIUM_OCTEON2:
2014 /* Examine entrylo 0 or 1 based on ptr. */
2015 if (use_bbit_insns()) {
2016 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2018 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2019 uasm_i_beqz(&p, wr.r3, 8);
2021 /* load it in the delay slot*/
2022 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2023 /* load it if ptr is odd */
2024 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2026 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2027 * XI must have triggered it.
2029 if (use_bbit_insns()) {
2030 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2032 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2033 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2035 if (PM_DEFAULT_MASK == 0)
2038 * We clobbered C0_PAGEMASK, restore it. On the other branch
2039 * it is restored in build_huge_tlb_write_entry.
2041 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2043 uasm_l_tlbl_goaround2(&l, p);
2045 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2046 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2049 uasm_l_nopage_tlbl(&l, p);
2050 build_restore_work_registers(&p);
2051 #ifdef CONFIG_CPU_MICROMIPS
2052 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2053 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2054 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2058 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2061 if (p >= handle_tlbl_end)
2062 panic("TLB load handler fastpath space exceeded");
2064 uasm_resolve_relocs(relocs, labels);
2065 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2066 (unsigned int)(p - handle_tlbl));
2068 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2071 static void build_r4000_tlb_store_handler(void)
2073 u32 *p = handle_tlbs;
2074 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2075 struct uasm_label *l = labels;
2076 struct uasm_reloc *r = relocs;
2077 struct work_registers wr;
2079 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2080 memset(labels, 0, sizeof(labels));
2081 memset(relocs, 0, sizeof(relocs));
2083 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2084 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2085 if (m4kc_tlbp_war())
2086 build_tlb_probe_entry(&p);
2087 build_make_write(&p, &r, wr.r1, wr.r2);
2088 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2090 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2092 * This is the entry point when
2093 * build_r4000_tlbchange_handler_head spots a huge page.
2095 uasm_l_tlb_huge_update(&l, p);
2096 iPTE_LW(&p, wr.r1, wr.r2);
2097 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2098 build_tlb_probe_entry(&p);
2099 uasm_i_ori(&p, wr.r1, wr.r1,
2100 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2101 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2104 uasm_l_nopage_tlbs(&l, p);
2105 build_restore_work_registers(&p);
2106 #ifdef CONFIG_CPU_MICROMIPS
2107 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2108 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2109 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2113 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2116 if (p >= handle_tlbs_end)
2117 panic("TLB store handler fastpath space exceeded");
2119 uasm_resolve_relocs(relocs, labels);
2120 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2121 (unsigned int)(p - handle_tlbs));
2123 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2126 static void build_r4000_tlb_modify_handler(void)
2128 u32 *p = handle_tlbm;
2129 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2130 struct uasm_label *l = labels;
2131 struct uasm_reloc *r = relocs;
2132 struct work_registers wr;
2134 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2135 memset(labels, 0, sizeof(labels));
2136 memset(relocs, 0, sizeof(relocs));
2138 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2139 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2140 if (m4kc_tlbp_war())
2141 build_tlb_probe_entry(&p);
2142 /* Present and writable bits set, set accessed and dirty bits. */
2143 build_make_write(&p, &r, wr.r1, wr.r2);
2144 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2146 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2148 * This is the entry point when
2149 * build_r4000_tlbchange_handler_head spots a huge page.
2151 uasm_l_tlb_huge_update(&l, p);
2152 iPTE_LW(&p, wr.r1, wr.r2);
2153 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2154 build_tlb_probe_entry(&p);
2155 uasm_i_ori(&p, wr.r1, wr.r1,
2156 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2157 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2160 uasm_l_nopage_tlbm(&l, p);
2161 build_restore_work_registers(&p);
2162 #ifdef CONFIG_CPU_MICROMIPS
2163 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2164 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2165 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2169 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2172 if (p >= handle_tlbm_end)
2173 panic("TLB modify handler fastpath space exceeded");
2175 uasm_resolve_relocs(relocs, labels);
2176 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2177 (unsigned int)(p - handle_tlbm));
2179 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2182 static void flush_tlb_handlers(void)
2184 local_flush_icache_range((unsigned long)handle_tlbl,
2185 (unsigned long)handle_tlbl_end);
2186 local_flush_icache_range((unsigned long)handle_tlbs,
2187 (unsigned long)handle_tlbs_end);
2188 local_flush_icache_range((unsigned long)handle_tlbm,
2189 (unsigned long)handle_tlbm_end);
2190 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2191 (unsigned long)tlbmiss_handler_setup_pgd_end);
2194 void build_tlb_refill_handler(void)
2197 * The refill handler is generated per-CPU, multi-node systems
2198 * may have local storage for it. The other handlers are only
2201 static int run_once = 0;
2203 output_pgtable_bits_defines();
2206 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2209 switch (current_cpu_type()) {
2217 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2218 if (cpu_has_local_ebase)
2219 build_r3000_tlb_refill_handler();
2221 if (!cpu_has_local_ebase)
2222 build_r3000_tlb_refill_handler();
2224 build_r3000_tlb_load_handler();
2225 build_r3000_tlb_store_handler();
2226 build_r3000_tlb_modify_handler();
2227 flush_tlb_handlers();
2231 panic("No R3000 TLB refill handler");
2237 panic("No R6000 TLB refill handler yet");
2241 panic("No R8000 TLB refill handler yet");
2246 scratch_reg = allocate_kscratch();
2248 build_r4000_tlb_load_handler();
2249 build_r4000_tlb_store_handler();
2250 build_r4000_tlb_modify_handler();
2251 if (!cpu_has_local_ebase)
2252 build_r4000_tlb_refill_handler();
2253 flush_tlb_handlers();
2256 if (cpu_has_local_ebase)
2257 build_r4000_tlb_refill_handler();