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 completely 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/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.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>
39 static int mips_xpa_disabled;
41 static int __init xpa_disable(char *s)
43 mips_xpa_disabled = 1;
48 __setup("noxpa", xpa_disable);
51 * TLB load/store/modify handlers.
53 * Only the fastpath gets synthesized at runtime, the slowpath for
54 * do_page_fault remains normal asm.
56 extern void tlb_do_page_fault_0(void);
57 extern void tlb_do_page_fault_1(void);
59 struct work_registers {
68 } ____cacheline_aligned_in_smp;
70 static struct tlb_reg_save handler_reg_save[NR_CPUS];
72 static inline int r45k_bvahwbug(void)
74 /* XXX: We should probe for the presence of this bug, but we don't. */
78 static inline int r4k_250MHZhwbug(void)
80 /* XXX: We should probe for the presence of this bug, but we don't. */
84 static inline int __maybe_unused bcm1250_m3_war(void)
86 return BCM1250_M3_WAR;
89 static inline int __maybe_unused r10000_llsc_war(void)
91 return R10000_LLSC_WAR;
94 static int use_bbit_insns(void)
96 switch (current_cpu_type()) {
97 case CPU_CAVIUM_OCTEON:
98 case CPU_CAVIUM_OCTEON_PLUS:
99 case CPU_CAVIUM_OCTEON2:
100 case CPU_CAVIUM_OCTEON3:
107 static int use_lwx_insns(void)
109 switch (current_cpu_type()) {
110 case CPU_CAVIUM_OCTEON2:
111 case CPU_CAVIUM_OCTEON3:
117 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
118 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
119 static bool scratchpad_available(void)
123 static int scratchpad_offset(int i)
126 * CVMSEG starts at address -32768 and extends for
127 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
129 i += 1; /* Kernel use starts at the top and works down. */
130 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
133 static bool scratchpad_available(void)
137 static int scratchpad_offset(int i)
140 /* Really unreachable, but evidently some GCC want this. */
145 * Found by experiment: At least some revisions of the 4kc throw under
146 * some circumstances a machine check exception, triggered by invalid
147 * values in the index register. Delaying the tlbp instruction until
148 * after the next branch, plus adding an additional nop in front of
149 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
150 * why; it's not an issue caused by the core RTL.
153 static int m4kc_tlbp_war(void)
155 return (current_cpu_data.processor_id & 0xffff00) ==
156 (PRID_COMP_MIPS | PRID_IMP_4KC);
159 /* Handle labels (which must be positive integers). */
161 label_second_part = 1,
166 label_split = label_tlbw_hazard_0 + 8,
167 label_tlbl_goaround1,
168 label_tlbl_goaround2,
172 label_smp_pgtable_change,
173 label_r3000_write_probe_fail,
174 label_large_segbits_fault,
175 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
176 label_tlb_huge_update,
180 UASM_L_LA(_second_part)
183 UASM_L_LA(_vmalloc_done)
184 /* _tlbw_hazard_x is handled differently. */
186 UASM_L_LA(_tlbl_goaround1)
187 UASM_L_LA(_tlbl_goaround2)
188 UASM_L_LA(_nopage_tlbl)
189 UASM_L_LA(_nopage_tlbs)
190 UASM_L_LA(_nopage_tlbm)
191 UASM_L_LA(_smp_pgtable_change)
192 UASM_L_LA(_r3000_write_probe_fail)
193 UASM_L_LA(_large_segbits_fault)
194 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
195 UASM_L_LA(_tlb_huge_update)
198 static int hazard_instance;
200 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
204 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
211 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
215 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
223 * pgtable bits are assigned dynamically depending on processor feature
224 * and statically based on kernel configuration. This spits out the actual
225 * values the kernel is using. Required to make sense from disassembled
226 * TLB exception handlers.
228 static void output_pgtable_bits_defines(void)
230 #define pr_define(fmt, ...) \
231 pr_debug("#define " fmt, ##__VA_ARGS__)
233 pr_debug("#include <asm/asm.h>\n");
234 pr_debug("#include <asm/regdef.h>\n");
237 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
238 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
239 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
240 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
241 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
242 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
243 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
245 #ifdef _PAGE_NO_EXEC_SHIFT
247 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
249 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
250 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
251 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
252 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
256 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
260 pr_debug("LEAF(%s)\n", symbol);
262 pr_debug("\t.set push\n");
263 pr_debug("\t.set noreorder\n");
265 for (i = 0; i < count; i++)
266 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
268 pr_debug("\t.set\tpop\n");
270 pr_debug("\tEND(%s)\n", symbol);
273 /* The only general purpose registers allowed in TLB handlers. */
277 /* Some CP0 registers */
278 #define C0_INDEX 0, 0
279 #define C0_ENTRYLO0 2, 0
280 #define C0_TCBIND 2, 2
281 #define C0_ENTRYLO1 3, 0
282 #define C0_CONTEXT 4, 0
283 #define C0_PAGEMASK 5, 0
284 #define C0_PWBASE 5, 5
285 #define C0_PWFIELD 5, 6
286 #define C0_PWSIZE 5, 7
287 #define C0_PWCTL 6, 6
288 #define C0_BADVADDR 8, 0
290 #define C0_ENTRYHI 10, 0
292 #define C0_XCONTEXT 20, 0
295 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
297 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
300 /* The worst case length of the handler is around 18 instructions for
301 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
302 * Maximum space available is 32 instructions for R3000 and 64
303 * instructions for R4000.
305 * We deliberately chose a buffer size of 128, so we won't scribble
306 * over anything important on overflow before we panic.
308 static u32 tlb_handler[128];
310 /* simply assume worst case size for labels and relocs */
311 static struct uasm_label labels[128];
312 static struct uasm_reloc relocs[128];
314 static int check_for_high_segbits;
315 static bool fill_includes_sw_bits;
317 static unsigned int kscratch_used_mask;
319 static inline int __maybe_unused c0_kscratch(void)
321 switch (current_cpu_type()) {
330 static int allocate_kscratch(void)
333 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
340 r--; /* make it zero based */
342 kscratch_used_mask |= (1 << r);
347 static int scratch_reg;
349 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
351 static struct work_registers build_get_work_registers(u32 **p)
353 struct work_registers r;
355 if (scratch_reg >= 0) {
356 /* Save in CPU local C0_KScratch? */
357 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
364 if (num_possible_cpus() > 1) {
365 /* Get smp_processor_id */
366 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
367 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
369 /* handler_reg_save index in K0 */
370 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
372 UASM_i_LA(p, K1, (long)&handler_reg_save);
373 UASM_i_ADDU(p, K0, K0, K1);
375 UASM_i_LA(p, K0, (long)&handler_reg_save);
377 /* K0 now points to save area, save $1 and $2 */
378 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
379 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
387 static void build_restore_work_registers(u32 **p)
389 if (scratch_reg >= 0) {
390 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
393 /* K0 already points to save area, restore $1 and $2 */
394 UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
395 UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
398 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
401 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
402 * we cannot do r3000 under these circumstances.
404 * Declare pgd_current here instead of including mmu_context.h to avoid type
405 * conflicts for tlbmiss_handler_setup_pgd
407 extern unsigned long pgd_current[];
410 * The R3000 TLB handler is simple.
412 static void build_r3000_tlb_refill_handler(void)
414 long pgdc = (long)pgd_current;
417 memset(tlb_handler, 0, sizeof(tlb_handler));
420 uasm_i_mfc0(&p, K0, C0_BADVADDR);
421 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
422 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
423 uasm_i_srl(&p, K0, K0, 22); /* load delay */
424 uasm_i_sll(&p, K0, K0, 2);
425 uasm_i_addu(&p, K1, K1, K0);
426 uasm_i_mfc0(&p, K0, C0_CONTEXT);
427 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
428 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
429 uasm_i_addu(&p, K1, K1, K0);
430 uasm_i_lw(&p, K0, 0, K1);
431 uasm_i_nop(&p); /* load delay */
432 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
433 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
434 uasm_i_tlbwr(&p); /* cp0 delay */
436 uasm_i_rfe(&p); /* branch delay */
438 if (p > tlb_handler + 32)
439 panic("TLB refill handler space exceeded");
441 pr_debug("Wrote TLB refill handler (%u instructions).\n",
442 (unsigned int)(p - tlb_handler));
444 memcpy((void *)ebase, tlb_handler, 0x80);
445 local_flush_icache_range(ebase, ebase + 0x80);
447 dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
449 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
452 * The R4000 TLB handler is much more complicated. We have two
453 * consecutive handler areas with 32 instructions space each.
454 * Since they aren't used at the same time, we can overflow in the
455 * other one.To keep things simple, we first assume linear space,
456 * then we relocate it to the final handler layout as needed.
458 static u32 final_handler[64];
463 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
464 * 2. A timing hazard exists for the TLBP instruction.
466 * stalling_instruction
469 * The JTLB is being read for the TLBP throughout the stall generated by the
470 * previous instruction. This is not really correct as the stalling instruction
471 * can modify the address used to access the JTLB. The failure symptom is that
472 * the TLBP instruction will use an address created for the stalling instruction
473 * and not the address held in C0_ENHI and thus report the wrong results.
475 * The software work-around is to not allow the instruction preceding the TLBP
476 * to stall - make it an NOP or some other instruction guaranteed not to stall.
478 * Errata 2 will not be fixed. This errata is also on the R5000.
480 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
482 static void __maybe_unused build_tlb_probe_entry(u32 **p)
484 switch (current_cpu_type()) {
485 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
501 * Write random or indexed TLB entry, and care about the hazards from
502 * the preceding mtc0 and for the following eret.
504 enum tlb_write_entry { tlb_random, tlb_indexed };
506 static void build_tlb_write_entry(u32 **p, struct uasm_label **l,
507 struct uasm_reloc **r,
508 enum tlb_write_entry wmode)
510 void(*tlbw)(u32 **) = NULL;
513 case tlb_random: tlbw = uasm_i_tlbwr; break;
514 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
517 if (cpu_has_mips_r2_r6) {
518 if (cpu_has_mips_r2_exec_hazard)
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 */
626 panic("No TLB refill handler yet (CPU type: %d)",
632 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
635 if (_PAGE_GLOBAL_SHIFT == 0) {
636 /* pte_t is already in EntryLo format */
640 if (cpu_has_rixi && _PAGE_NO_EXEC) {
641 if (fill_includes_sw_bits) {
642 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
644 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
645 UASM_i_ROTR(p, reg, reg,
646 ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
649 #ifdef CONFIG_PHYS_ADDR_T_64BIT
650 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
652 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
657 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
659 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
660 unsigned int tmp, enum label_id lid,
663 if (restore_scratch) {
664 /* Reset default page size */
665 if (PM_DEFAULT_MASK >> 16) {
666 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
667 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
668 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
669 uasm_il_b(p, r, lid);
670 } else if (PM_DEFAULT_MASK) {
671 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
672 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
673 uasm_il_b(p, r, lid);
675 uasm_i_mtc0(p, 0, C0_PAGEMASK);
676 uasm_il_b(p, r, lid);
678 if (scratch_reg >= 0)
679 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
681 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
683 /* Reset default page size */
684 if (PM_DEFAULT_MASK >> 16) {
685 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
686 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
687 uasm_il_b(p, r, lid);
688 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
689 } else if (PM_DEFAULT_MASK) {
690 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
691 uasm_il_b(p, r, lid);
692 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
694 uasm_il_b(p, r, lid);
695 uasm_i_mtc0(p, 0, C0_PAGEMASK);
700 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
701 struct uasm_reloc **r,
703 enum tlb_write_entry wmode,
706 /* Set huge page tlb entry size */
707 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
708 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
709 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
711 build_tlb_write_entry(p, l, r, wmode);
713 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
717 * Check if Huge PTE is present, if so then jump to LABEL.
720 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
721 unsigned int pmd, int lid)
723 UASM_i_LW(p, tmp, 0, pmd);
724 if (use_bbit_insns()) {
725 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
727 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
728 uasm_il_bnez(p, r, tmp, lid);
732 static void build_huge_update_entries(u32 **p, unsigned int pte,
738 * A huge PTE describes an area the size of the
739 * configured huge page size. This is twice the
740 * of the large TLB entry size we intend to use.
741 * A TLB entry half the size of the configured
742 * huge page size is configured into entrylo0
743 * and entrylo1 to cover the contiguous huge PTE
746 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
748 /* We can clobber tmp. It isn't used after this.*/
750 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
752 build_convert_pte_to_entrylo(p, pte);
753 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
754 /* convert to entrylo1 */
756 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
758 UASM_i_ADDU(p, pte, pte, tmp);
760 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
763 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
764 struct uasm_label **l,
769 UASM_i_SC(p, pte, 0, ptr);
770 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
771 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
773 UASM_i_SW(p, pte, 0, ptr);
775 build_huge_update_entries(p, pte, ptr);
776 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
778 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
782 * TMP and PTR are scratch.
783 * TMP will be clobbered, PTR will hold the pmd entry.
786 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
787 unsigned int tmp, unsigned int ptr)
789 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
790 long pgdc = (long)pgd_current;
793 * The vmalloc handling is not in the hotpath.
795 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
797 if (check_for_high_segbits) {
799 * The kernel currently implicitely assumes that the
800 * MIPS SEGBITS parameter for the processor is
801 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
802 * allocate virtual addresses outside the maximum
803 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
804 * that doesn't prevent user code from accessing the
805 * higher xuseg addresses. Here, we make sure that
806 * everything but the lower xuseg addresses goes down
807 * the module_alloc/vmalloc path.
809 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
810 uasm_il_bnez(p, r, ptr, label_vmalloc);
812 uasm_il_bltz(p, r, tmp, label_vmalloc);
814 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
817 /* pgd is in pgd_reg */
819 UASM_i_MFC0(p, ptr, C0_PWBASE);
821 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
823 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
825 * &pgd << 11 stored in CONTEXT [23..63].
827 UASM_i_MFC0(p, ptr, C0_CONTEXT);
829 /* Clear lower 23 bits of context. */
830 uasm_i_dins(p, ptr, 0, 0, 23);
832 /* 1 0 1 0 1 << 6 xkphys cached */
833 uasm_i_ori(p, ptr, ptr, 0x540);
834 uasm_i_drotr(p, ptr, ptr, 11);
835 #elif defined(CONFIG_SMP)
836 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
837 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
838 UASM_i_LA_mostly(p, tmp, pgdc);
839 uasm_i_daddu(p, ptr, ptr, tmp);
840 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
841 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
843 UASM_i_LA_mostly(p, ptr, pgdc);
844 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
848 uasm_l_vmalloc_done(l, *p);
850 /* get pgd offset in bytes */
851 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
853 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
854 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
855 #ifndef __PAGETABLE_PMD_FOLDED
856 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
857 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
858 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
859 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
860 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
865 * BVADDR is the faulting address, PTR is scratch.
866 * PTR will hold the pgd for vmalloc.
869 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
870 unsigned int bvaddr, unsigned int ptr,
871 enum vmalloc64_mode mode)
873 long swpd = (long)swapper_pg_dir;
874 int single_insn_swpd;
875 int did_vmalloc_branch = 0;
877 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
879 uasm_l_vmalloc(l, *p);
881 if (mode != not_refill && check_for_high_segbits) {
882 if (single_insn_swpd) {
883 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
884 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
885 did_vmalloc_branch = 1;
888 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
891 if (!did_vmalloc_branch) {
892 if (single_insn_swpd) {
893 uasm_il_b(p, r, label_vmalloc_done);
894 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
896 UASM_i_LA_mostly(p, ptr, swpd);
897 uasm_il_b(p, r, label_vmalloc_done);
898 if (uasm_in_compat_space_p(swpd))
899 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
901 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
904 if (mode != not_refill && check_for_high_segbits) {
905 uasm_l_large_segbits_fault(l, *p);
907 * We get here if we are an xsseg address, or if we are
908 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
910 * Ignoring xsseg (assume disabled so would generate
911 * (address errors?), the only remaining possibility
912 * is the upper xuseg addresses. On processors with
913 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
914 * addresses would have taken an address error. We try
915 * to mimic that here by taking a load/istream page
918 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
921 if (mode == refill_scratch) {
922 if (scratch_reg >= 0)
923 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
925 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
932 #else /* !CONFIG_64BIT */
935 * TMP and PTR are scratch.
936 * TMP will be clobbered, PTR will hold the pgd entry.
938 static void __maybe_unused
939 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
942 /* pgd is in pgd_reg */
943 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
944 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
946 long pgdc = (long)pgd_current;
948 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
950 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
951 UASM_i_LA_mostly(p, tmp, pgdc);
952 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
953 uasm_i_addu(p, ptr, tmp, ptr);
955 UASM_i_LA_mostly(p, ptr, pgdc);
957 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
958 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
960 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
961 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
962 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
965 #endif /* !CONFIG_64BIT */
967 static void build_adjust_context(u32 **p, unsigned int ctx)
969 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
970 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
972 switch (current_cpu_type()) {
989 UASM_i_SRL(p, ctx, ctx, shift);
990 uasm_i_andi(p, ctx, ctx, mask);
993 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
996 * Bug workaround for the Nevada. It seems as if under certain
997 * circumstances the move from cp0_context might produce a
998 * bogus result when the mfc0 instruction and its consumer are
999 * in a different cacheline or a load instruction, probably any
1000 * memory reference, is between them.
1002 switch (current_cpu_type()) {
1004 UASM_i_LW(p, ptr, 0, ptr);
1005 GET_CONTEXT(p, tmp); /* get context reg */
1009 GET_CONTEXT(p, tmp); /* get context reg */
1010 UASM_i_LW(p, ptr, 0, ptr);
1014 build_adjust_context(p, tmp);
1015 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1018 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1020 int pte_off_even = 0;
1021 int pte_off_odd = sizeof(pte_t);
1023 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1024 /* The low 32 bits of EntryLo is stored in pte_high */
1025 pte_off_even += offsetof(pte_t, pte_high);
1026 pte_off_odd += offsetof(pte_t, pte_high);
1029 if (IS_ENABLED(CONFIG_XPA)) {
1030 uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1031 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1032 UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1034 if (cpu_has_xpa && !mips_xpa_disabled) {
1035 uasm_i_lw(p, tmp, 0, ptep);
1036 uasm_i_ext(p, tmp, tmp, 0, 24);
1037 uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1040 uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1041 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1042 UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1044 if (cpu_has_xpa && !mips_xpa_disabled) {
1045 uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1046 uasm_i_ext(p, tmp, tmp, 0, 24);
1047 uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1052 UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1053 UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1054 if (r45k_bvahwbug())
1055 build_tlb_probe_entry(p);
1056 build_convert_pte_to_entrylo(p, tmp);
1057 if (r4k_250MHZhwbug())
1058 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1059 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1060 build_convert_pte_to_entrylo(p, ptep);
1061 if (r45k_bvahwbug())
1062 uasm_i_mfc0(p, tmp, C0_INDEX);
1063 if (r4k_250MHZhwbug())
1064 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1065 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1068 struct mips_huge_tlb_info {
1070 int restore_scratch;
1071 bool need_reload_pte;
1074 static struct mips_huge_tlb_info
1075 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1076 struct uasm_reloc **r, unsigned int tmp,
1077 unsigned int ptr, int c0_scratch_reg)
1079 struct mips_huge_tlb_info rv;
1080 unsigned int even, odd;
1081 int vmalloc_branch_delay_filled = 0;
1082 const int scratch = 1; /* Our extra working register */
1084 rv.huge_pte = scratch;
1085 rv.restore_scratch = 0;
1086 rv.need_reload_pte = false;
1088 if (check_for_high_segbits) {
1089 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1092 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1094 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1096 if (c0_scratch_reg >= 0)
1097 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1099 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1101 uasm_i_dsrl_safe(p, scratch, tmp,
1102 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1103 uasm_il_bnez(p, r, scratch, label_vmalloc);
1105 if (pgd_reg == -1) {
1106 vmalloc_branch_delay_filled = 1;
1107 /* Clear lower 23 bits of context. */
1108 uasm_i_dins(p, ptr, 0, 0, 23);
1112 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1114 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1116 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1118 if (c0_scratch_reg >= 0)
1119 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1121 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1124 /* Clear lower 23 bits of context. */
1125 uasm_i_dins(p, ptr, 0, 0, 23);
1127 uasm_il_bltz(p, r, tmp, label_vmalloc);
1130 if (pgd_reg == -1) {
1131 vmalloc_branch_delay_filled = 1;
1132 /* 1 0 1 0 1 << 6 xkphys cached */
1133 uasm_i_ori(p, ptr, ptr, 0x540);
1134 uasm_i_drotr(p, ptr, ptr, 11);
1137 #ifdef __PAGETABLE_PMD_FOLDED
1138 #define LOC_PTEP scratch
1140 #define LOC_PTEP ptr
1143 if (!vmalloc_branch_delay_filled)
1144 /* get pgd offset in bytes */
1145 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1147 uasm_l_vmalloc_done(l, *p);
1151 * fall-through case = badvaddr *pgd_current
1152 * vmalloc case = badvaddr swapper_pg_dir
1155 if (vmalloc_branch_delay_filled)
1156 /* get pgd offset in bytes */
1157 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1159 #ifdef __PAGETABLE_PMD_FOLDED
1160 GET_CONTEXT(p, tmp); /* get context reg */
1162 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1164 if (use_lwx_insns()) {
1165 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1167 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1168 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1171 #ifndef __PAGETABLE_PMD_FOLDED
1172 /* get pmd offset in bytes */
1173 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1174 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1175 GET_CONTEXT(p, tmp); /* get context reg */
1177 if (use_lwx_insns()) {
1178 UASM_i_LWX(p, scratch, scratch, ptr);
1180 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1181 UASM_i_LW(p, scratch, 0, ptr);
1184 /* Adjust the context during the load latency. */
1185 build_adjust_context(p, tmp);
1187 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1188 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1190 * The in the LWX case we don't want to do the load in the
1191 * delay slot. It cannot issue in the same cycle and may be
1192 * speculative and unneeded.
1194 if (use_lwx_insns())
1196 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1199 /* build_update_entries */
1200 if (use_lwx_insns()) {
1203 UASM_i_LWX(p, even, scratch, tmp);
1204 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1205 UASM_i_LWX(p, odd, scratch, tmp);
1207 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1210 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1211 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1214 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1215 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1216 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1218 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1219 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1220 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1222 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1224 if (c0_scratch_reg >= 0) {
1225 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1226 build_tlb_write_entry(p, l, r, tlb_random);
1227 uasm_l_leave(l, *p);
1228 rv.restore_scratch = 1;
1229 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1230 build_tlb_write_entry(p, l, r, tlb_random);
1231 uasm_l_leave(l, *p);
1232 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1234 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1235 build_tlb_write_entry(p, l, r, tlb_random);
1236 uasm_l_leave(l, *p);
1237 rv.restore_scratch = 1;
1240 uasm_i_eret(p); /* return from trap */
1246 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1247 * because EXL == 0. If we wrap, we can also use the 32 instruction
1248 * slots before the XTLB refill exception handler which belong to the
1249 * unused TLB refill exception.
1251 #define MIPS64_REFILL_INSNS 32
1253 static void build_r4000_tlb_refill_handler(void)
1255 u32 *p = tlb_handler;
1256 struct uasm_label *l = labels;
1257 struct uasm_reloc *r = relocs;
1259 unsigned int final_len;
1260 struct mips_huge_tlb_info htlb_info __maybe_unused;
1261 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1263 memset(tlb_handler, 0, sizeof(tlb_handler));
1264 memset(labels, 0, sizeof(labels));
1265 memset(relocs, 0, sizeof(relocs));
1266 memset(final_handler, 0, sizeof(final_handler));
1268 if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1269 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1271 vmalloc_mode = refill_scratch;
1273 htlb_info.huge_pte = K0;
1274 htlb_info.restore_scratch = 0;
1275 htlb_info.need_reload_pte = true;
1276 vmalloc_mode = refill_noscratch;
1278 * create the plain linear handler
1280 if (bcm1250_m3_war()) {
1281 unsigned int segbits = 44;
1283 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1284 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1285 uasm_i_xor(&p, K0, K0, K1);
1286 uasm_i_dsrl_safe(&p, K1, K0, 62);
1287 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1288 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1289 uasm_i_or(&p, K0, K0, K1);
1290 uasm_il_bnez(&p, &r, K0, label_leave);
1291 /* No need for uasm_i_nop */
1295 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1297 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1300 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1301 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1304 build_get_ptep(&p, K0, K1);
1305 build_update_entries(&p, K0, K1);
1306 build_tlb_write_entry(&p, &l, &r, tlb_random);
1307 uasm_l_leave(&l, p);
1308 uasm_i_eret(&p); /* return from trap */
1310 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1311 uasm_l_tlb_huge_update(&l, p);
1312 if (htlb_info.need_reload_pte)
1313 UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1314 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1315 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1316 htlb_info.restore_scratch);
1320 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1324 * Overflow check: For the 64bit handler, we need at least one
1325 * free instruction slot for the wrap-around branch. In worst
1326 * case, if the intended insertion point is a delay slot, we
1327 * need three, with the second nop'ed and the third being
1330 switch (boot_cpu_type()) {
1332 if (sizeof(long) == 4) {
1334 /* Loongson2 ebase is different than r4k, we have more space */
1335 if ((p - tlb_handler) > 64)
1336 panic("TLB refill handler space exceeded");
1338 * Now fold the handler in the TLB refill handler space.
1341 /* Simplest case, just copy the handler. */
1342 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1343 final_len = p - tlb_handler;
1346 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1347 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1348 && uasm_insn_has_bdelay(relocs,
1349 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1350 panic("TLB refill handler space exceeded");
1352 * Now fold the handler in the TLB refill handler space.
1354 f = final_handler + MIPS64_REFILL_INSNS;
1355 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1356 /* Just copy the handler. */
1357 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1358 final_len = p - tlb_handler;
1360 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1361 const enum label_id ls = label_tlb_huge_update;
1363 const enum label_id ls = label_vmalloc;
1369 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1371 BUG_ON(i == ARRAY_SIZE(labels));
1372 split = labels[i].addr;
1375 * See if we have overflown one way or the other.
1377 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1378 split < p - MIPS64_REFILL_INSNS)
1383 * Split two instructions before the end. One
1384 * for the branch and one for the instruction
1385 * in the delay slot.
1387 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1390 * If the branch would fall in a delay slot,
1391 * we must back up an additional instruction
1392 * so that it is no longer in a delay slot.
1394 if (uasm_insn_has_bdelay(relocs, split - 1))
1397 /* Copy first part of the handler. */
1398 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1399 f += split - tlb_handler;
1402 /* Insert branch. */
1403 uasm_l_split(&l, final_handler);
1404 uasm_il_b(&f, &r, label_split);
1405 if (uasm_insn_has_bdelay(relocs, split))
1408 uasm_copy_handler(relocs, labels,
1409 split, split + 1, f);
1410 uasm_move_labels(labels, f, f + 1, -1);
1416 /* Copy the rest of the handler. */
1417 uasm_copy_handler(relocs, labels, split, p, final_handler);
1418 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1425 uasm_resolve_relocs(relocs, labels);
1426 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1429 memcpy((void *)ebase, final_handler, 0x100);
1430 local_flush_icache_range(ebase, ebase + 0x100);
1432 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1435 static void setup_pw(void)
1437 unsigned long pgd_i, pgd_w;
1438 #ifndef __PAGETABLE_PMD_FOLDED
1439 unsigned long pmd_i, pmd_w;
1441 unsigned long pt_i, pt_w;
1442 unsigned long pte_i, pte_w;
1443 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1446 psn = ilog2(_PAGE_HUGE); /* bit used to indicate huge page */
1448 pgd_i = PGDIR_SHIFT; /* 1st level PGD */
1449 #ifndef __PAGETABLE_PMD_FOLDED
1450 pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1452 pmd_i = PMD_SHIFT; /* 2nd level PMD */
1453 pmd_w = PMD_SHIFT - PAGE_SHIFT;
1455 pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1458 pt_i = PAGE_SHIFT; /* 3rd level PTE */
1459 pt_w = PAGE_SHIFT - 3;
1461 pte_i = ilog2(_PAGE_GLOBAL);
1464 #ifndef __PAGETABLE_PMD_FOLDED
1465 write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1466 write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1468 write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1469 write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1472 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1473 write_c0_pwctl(1 << 6 | psn);
1475 write_c0_kpgd(swapper_pg_dir);
1476 kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1479 static void build_loongson3_tlb_refill_handler(void)
1481 u32 *p = tlb_handler;
1482 struct uasm_label *l = labels;
1483 struct uasm_reloc *r = relocs;
1485 memset(labels, 0, sizeof(labels));
1486 memset(relocs, 0, sizeof(relocs));
1487 memset(tlb_handler, 0, sizeof(tlb_handler));
1489 if (check_for_high_segbits) {
1490 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1491 uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1492 uasm_il_beqz(&p, &r, K1, label_vmalloc);
1495 uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1497 uasm_l_vmalloc(&l, p);
1500 uasm_i_dmfc0(&p, K1, C0_PGD);
1502 uasm_i_lddir(&p, K0, K1, 3); /* global page dir */
1503 #ifndef __PAGETABLE_PMD_FOLDED
1504 uasm_i_lddir(&p, K1, K0, 1); /* middle page dir */
1506 uasm_i_ldpte(&p, K1, 0); /* even */
1507 uasm_i_ldpte(&p, K1, 1); /* odd */
1510 /* restore page mask */
1511 if (PM_DEFAULT_MASK >> 16) {
1512 uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1513 uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1514 uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1515 } else if (PM_DEFAULT_MASK) {
1516 uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1517 uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1519 uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1524 if (check_for_high_segbits) {
1525 uasm_l_large_segbits_fault(&l, p);
1526 UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1531 uasm_resolve_relocs(relocs, labels);
1532 memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1533 local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1534 dump_handler("loongson3_tlb_refill", (u32 *)(ebase + 0x80), 32);
1537 extern u32 handle_tlbl[], handle_tlbl_end[];
1538 extern u32 handle_tlbs[], handle_tlbs_end[];
1539 extern u32 handle_tlbm[], handle_tlbm_end[];
1540 extern u32 tlbmiss_handler_setup_pgd_start[];
1541 extern u32 tlbmiss_handler_setup_pgd[];
1542 EXPORT_SYMBOL_GPL(tlbmiss_handler_setup_pgd);
1543 extern u32 tlbmiss_handler_setup_pgd_end[];
1545 static void build_setup_pgd(void)
1548 const int __maybe_unused a1 = 5;
1549 const int __maybe_unused a2 = 6;
1550 u32 *p = tlbmiss_handler_setup_pgd_start;
1551 const int tlbmiss_handler_setup_pgd_size =
1552 tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1553 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1554 long pgdc = (long)pgd_current;
1557 memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1558 sizeof(tlbmiss_handler_setup_pgd[0]));
1559 memset(labels, 0, sizeof(labels));
1560 memset(relocs, 0, sizeof(relocs));
1561 pgd_reg = allocate_kscratch();
1562 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1563 if (pgd_reg == -1) {
1564 struct uasm_label *l = labels;
1565 struct uasm_reloc *r = relocs;
1567 /* PGD << 11 in c0_Context */
1569 * If it is a ckseg0 address, convert to a physical
1570 * address. Shifting right by 29 and adding 4 will
1571 * result in zero for these addresses.
1574 UASM_i_SRA(&p, a1, a0, 29);
1575 UASM_i_ADDIU(&p, a1, a1, 4);
1576 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1578 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1579 uasm_l_tlbl_goaround1(&l, p);
1580 UASM_i_SLL(&p, a0, a0, 11);
1582 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1584 /* PGD in c0_KScratch */
1587 UASM_i_MTC0(&p, a0, C0_PWBASE);
1589 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1593 /* Save PGD to pgd_current[smp_processor_id()] */
1594 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1595 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1596 UASM_i_LA_mostly(&p, a2, pgdc);
1597 UASM_i_ADDU(&p, a2, a2, a1);
1598 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1600 UASM_i_LA_mostly(&p, a2, pgdc);
1601 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1605 /* if pgd_reg is allocated, save PGD also to scratch register */
1607 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1611 if (p >= tlbmiss_handler_setup_pgd_end)
1612 panic("tlbmiss_handler_setup_pgd space exceeded");
1614 uasm_resolve_relocs(relocs, labels);
1615 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1616 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1618 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1619 tlbmiss_handler_setup_pgd_size);
1623 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1626 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1628 uasm_i_lld(p, pte, 0, ptr);
1631 UASM_i_LL(p, pte, 0, ptr);
1633 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1635 uasm_i_ld(p, pte, 0, ptr);
1638 UASM_i_LW(p, pte, 0, ptr);
1643 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1644 unsigned int mode, unsigned int scratch)
1646 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1647 unsigned int swmode = mode & ~hwmode;
1649 if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1650 uasm_i_lui(p, scratch, swmode >> 16);
1651 uasm_i_or(p, pte, pte, scratch);
1652 BUG_ON(swmode & 0xffff);
1654 uasm_i_ori(p, pte, pte, mode);
1658 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1660 uasm_i_scd(p, pte, 0, ptr);
1663 UASM_i_SC(p, pte, 0, ptr);
1665 if (r10000_llsc_war())
1666 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1668 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1670 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1671 if (!cpu_has_64bits) {
1672 /* no uasm_i_nop needed */
1673 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1674 uasm_i_ori(p, pte, pte, hwmode);
1675 BUG_ON(hwmode & ~0xffff);
1676 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1677 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1678 /* no uasm_i_nop needed */
1679 uasm_i_lw(p, pte, 0, ptr);
1686 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1688 uasm_i_sd(p, pte, 0, ptr);
1691 UASM_i_SW(p, pte, 0, ptr);
1693 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1694 if (!cpu_has_64bits) {
1695 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1696 uasm_i_ori(p, pte, pte, hwmode);
1697 BUG_ON(hwmode & ~0xffff);
1698 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1699 uasm_i_lw(p, pte, 0, ptr);
1706 * Check if PTE is present, if not then jump to LABEL. PTR points to
1707 * the page table where this PTE is located, PTE will be re-loaded
1708 * with it's original value.
1711 build_pte_present(u32 **p, struct uasm_reloc **r,
1712 int pte, int ptr, int scratch, enum label_id lid)
1714 int t = scratch >= 0 ? scratch : pte;
1718 if (use_bbit_insns()) {
1719 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1722 if (_PAGE_PRESENT_SHIFT) {
1723 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1726 uasm_i_andi(p, t, cur, 1);
1727 uasm_il_beqz(p, r, t, lid);
1729 /* You lose the SMP race :-(*/
1730 iPTE_LW(p, pte, ptr);
1733 if (_PAGE_PRESENT_SHIFT) {
1734 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1737 uasm_i_andi(p, t, cur,
1738 (_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1739 uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1740 uasm_il_bnez(p, r, t, lid);
1742 /* You lose the SMP race :-(*/
1743 iPTE_LW(p, pte, ptr);
1747 /* Make PTE valid, store result in PTR. */
1749 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1750 unsigned int ptr, unsigned int scratch)
1752 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1754 iPTE_SW(p, r, pte, ptr, mode, scratch);
1758 * Check if PTE can be written to, if not branch to LABEL. Regardless
1759 * restore PTE with value from PTR when done.
1762 build_pte_writable(u32 **p, struct uasm_reloc **r,
1763 unsigned int pte, unsigned int ptr, int scratch,
1766 int t = scratch >= 0 ? scratch : pte;
1769 if (_PAGE_PRESENT_SHIFT) {
1770 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1773 uasm_i_andi(p, t, cur,
1774 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1775 uasm_i_xori(p, t, t,
1776 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1777 uasm_il_bnez(p, r, t, lid);
1779 /* You lose the SMP race :-(*/
1780 iPTE_LW(p, pte, ptr);
1785 /* Make PTE writable, update software status bits as well, then store
1789 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1790 unsigned int ptr, unsigned int scratch)
1792 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1795 iPTE_SW(p, r, pte, ptr, mode, scratch);
1799 * Check if PTE can be modified, if not branch to LABEL. Regardless
1800 * restore PTE with value from PTR when done.
1803 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1804 unsigned int pte, unsigned int ptr, int scratch,
1807 if (use_bbit_insns()) {
1808 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1811 int t = scratch >= 0 ? scratch : pte;
1812 uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1813 uasm_i_andi(p, t, t, 1);
1814 uasm_il_beqz(p, r, t, lid);
1816 /* You lose the SMP race :-(*/
1817 iPTE_LW(p, pte, ptr);
1821 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1825 * R3000 style TLB load/store/modify handlers.
1829 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1833 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1835 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1836 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1839 uasm_i_rfe(p); /* branch delay */
1843 * This places the pte into ENTRYLO0 and writes it with tlbwi
1844 * or tlbwr as appropriate. This is because the index register
1845 * may have the probe fail bit set as a result of a trap on a
1846 * kseg2 access, i.e. without refill. Then it returns.
1849 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1850 struct uasm_reloc **r, unsigned int pte,
1853 uasm_i_mfc0(p, tmp, C0_INDEX);
1854 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1855 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1856 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1857 uasm_i_tlbwi(p); /* cp0 delay */
1859 uasm_i_rfe(p); /* branch delay */
1860 uasm_l_r3000_write_probe_fail(l, *p);
1861 uasm_i_tlbwr(p); /* cp0 delay */
1863 uasm_i_rfe(p); /* branch delay */
1867 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1870 long pgdc = (long)pgd_current;
1872 uasm_i_mfc0(p, pte, C0_BADVADDR);
1873 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1874 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1875 uasm_i_srl(p, pte, pte, 22); /* load delay */
1876 uasm_i_sll(p, pte, pte, 2);
1877 uasm_i_addu(p, ptr, ptr, pte);
1878 uasm_i_mfc0(p, pte, C0_CONTEXT);
1879 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1880 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1881 uasm_i_addu(p, ptr, ptr, pte);
1882 uasm_i_lw(p, pte, 0, ptr);
1883 uasm_i_tlbp(p); /* load delay */
1886 static void build_r3000_tlb_load_handler(void)
1888 u32 *p = handle_tlbl;
1889 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1890 struct uasm_label *l = labels;
1891 struct uasm_reloc *r = relocs;
1893 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1894 memset(labels, 0, sizeof(labels));
1895 memset(relocs, 0, sizeof(relocs));
1897 build_r3000_tlbchange_handler_head(&p, K0, K1);
1898 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1899 uasm_i_nop(&p); /* load delay */
1900 build_make_valid(&p, &r, K0, K1, -1);
1901 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1903 uasm_l_nopage_tlbl(&l, p);
1904 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1907 if (p >= handle_tlbl_end)
1908 panic("TLB load handler fastpath space exceeded");
1910 uasm_resolve_relocs(relocs, labels);
1911 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1912 (unsigned int)(p - handle_tlbl));
1914 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1917 static void build_r3000_tlb_store_handler(void)
1919 u32 *p = handle_tlbs;
1920 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1921 struct uasm_label *l = labels;
1922 struct uasm_reloc *r = relocs;
1924 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1925 memset(labels, 0, sizeof(labels));
1926 memset(relocs, 0, sizeof(relocs));
1928 build_r3000_tlbchange_handler_head(&p, K0, K1);
1929 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1930 uasm_i_nop(&p); /* load delay */
1931 build_make_write(&p, &r, K0, K1, -1);
1932 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1934 uasm_l_nopage_tlbs(&l, p);
1935 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1938 if (p >= handle_tlbs_end)
1939 panic("TLB store handler fastpath space exceeded");
1941 uasm_resolve_relocs(relocs, labels);
1942 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1943 (unsigned int)(p - handle_tlbs));
1945 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1948 static void build_r3000_tlb_modify_handler(void)
1950 u32 *p = handle_tlbm;
1951 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1952 struct uasm_label *l = labels;
1953 struct uasm_reloc *r = relocs;
1955 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1956 memset(labels, 0, sizeof(labels));
1957 memset(relocs, 0, sizeof(relocs));
1959 build_r3000_tlbchange_handler_head(&p, K0, K1);
1960 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1961 uasm_i_nop(&p); /* load delay */
1962 build_make_write(&p, &r, K0, K1, -1);
1963 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1965 uasm_l_nopage_tlbm(&l, p);
1966 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1969 if (p >= handle_tlbm_end)
1970 panic("TLB modify handler fastpath space exceeded");
1972 uasm_resolve_relocs(relocs, labels);
1973 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1974 (unsigned int)(p - handle_tlbm));
1976 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1978 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1981 * R4000 style TLB load/store/modify handlers.
1983 static struct work_registers
1984 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1985 struct uasm_reloc **r)
1987 struct work_registers wr = build_get_work_registers(p);
1990 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1992 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1995 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1997 * For huge tlb entries, pmd doesn't contain an address but
1998 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1999 * see if we need to jump to huge tlb processing.
2001 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2004 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2005 UASM_i_LW(p, wr.r2, 0, wr.r2);
2006 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2007 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2008 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2011 uasm_l_smp_pgtable_change(l, *p);
2013 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2014 if (!m4kc_tlbp_war()) {
2015 build_tlb_probe_entry(p);
2017 /* race condition happens, leaving */
2019 uasm_i_mfc0(p, wr.r3, C0_INDEX);
2020 uasm_il_bltz(p, r, wr.r3, label_leave);
2028 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2029 struct uasm_reloc **r, unsigned int tmp,
2032 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2033 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2034 build_update_entries(p, tmp, ptr);
2035 build_tlb_write_entry(p, l, r, tlb_indexed);
2036 uasm_l_leave(l, *p);
2037 build_restore_work_registers(p);
2038 uasm_i_eret(p); /* return from trap */
2041 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2045 static void build_r4000_tlb_load_handler(void)
2047 u32 *p = handle_tlbl;
2048 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
2049 struct uasm_label *l = labels;
2050 struct uasm_reloc *r = relocs;
2051 struct work_registers wr;
2053 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
2054 memset(labels, 0, sizeof(labels));
2055 memset(relocs, 0, sizeof(relocs));
2057 if (bcm1250_m3_war()) {
2058 unsigned int segbits = 44;
2060 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2061 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2062 uasm_i_xor(&p, K0, K0, K1);
2063 uasm_i_dsrl_safe(&p, K1, K0, 62);
2064 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2065 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2066 uasm_i_or(&p, K0, K0, K1);
2067 uasm_il_bnez(&p, &r, K0, label_leave);
2068 /* No need for uasm_i_nop */
2071 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2072 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2073 if (m4kc_tlbp_war())
2074 build_tlb_probe_entry(&p);
2076 if (cpu_has_rixi && !cpu_has_rixiex) {
2078 * If the page is not _PAGE_VALID, RI or XI could not
2079 * have triggered it. Skip the expensive test..
2081 if (use_bbit_insns()) {
2082 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2083 label_tlbl_goaround1);
2085 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2086 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2092 switch (current_cpu_type()) {
2094 if (cpu_has_mips_r2_exec_hazard) {
2097 case CPU_CAVIUM_OCTEON:
2098 case CPU_CAVIUM_OCTEON_PLUS:
2099 case CPU_CAVIUM_OCTEON2:
2104 /* Examine entrylo 0 or 1 based on ptr. */
2105 if (use_bbit_insns()) {
2106 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2108 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2109 uasm_i_beqz(&p, wr.r3, 8);
2111 /* load it in the delay slot*/
2112 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2113 /* load it if ptr is odd */
2114 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2116 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2117 * XI must have triggered it.
2119 if (use_bbit_insns()) {
2120 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2122 uasm_l_tlbl_goaround1(&l, p);
2124 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2125 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2128 uasm_l_tlbl_goaround1(&l, p);
2130 build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2131 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2133 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2135 * This is the entry point when build_r4000_tlbchange_handler_head
2136 * spots a huge page.
2138 uasm_l_tlb_huge_update(&l, p);
2139 iPTE_LW(&p, wr.r1, wr.r2);
2140 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2141 build_tlb_probe_entry(&p);
2143 if (cpu_has_rixi && !cpu_has_rixiex) {
2145 * If the page is not _PAGE_VALID, RI or XI could not
2146 * have triggered it. Skip the expensive test..
2148 if (use_bbit_insns()) {
2149 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2150 label_tlbl_goaround2);
2152 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2153 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2159 switch (current_cpu_type()) {
2161 if (cpu_has_mips_r2_exec_hazard) {
2164 case CPU_CAVIUM_OCTEON:
2165 case CPU_CAVIUM_OCTEON_PLUS:
2166 case CPU_CAVIUM_OCTEON2:
2171 /* Examine entrylo 0 or 1 based on ptr. */
2172 if (use_bbit_insns()) {
2173 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2175 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2176 uasm_i_beqz(&p, wr.r3, 8);
2178 /* load it in the delay slot*/
2179 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2180 /* load it if ptr is odd */
2181 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2183 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2184 * XI must have triggered it.
2186 if (use_bbit_insns()) {
2187 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2189 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2190 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2192 if (PM_DEFAULT_MASK == 0)
2195 * We clobbered C0_PAGEMASK, restore it. On the other branch
2196 * it is restored in build_huge_tlb_write_entry.
2198 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2200 uasm_l_tlbl_goaround2(&l, p);
2202 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2203 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2206 uasm_l_nopage_tlbl(&l, p);
2207 build_restore_work_registers(&p);
2208 #ifdef CONFIG_CPU_MICROMIPS
2209 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2210 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2211 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2215 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2218 if (p >= handle_tlbl_end)
2219 panic("TLB load handler fastpath space exceeded");
2221 uasm_resolve_relocs(relocs, labels);
2222 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2223 (unsigned int)(p - handle_tlbl));
2225 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2228 static void build_r4000_tlb_store_handler(void)
2230 u32 *p = handle_tlbs;
2231 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2232 struct uasm_label *l = labels;
2233 struct uasm_reloc *r = relocs;
2234 struct work_registers wr;
2236 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2237 memset(labels, 0, sizeof(labels));
2238 memset(relocs, 0, sizeof(relocs));
2240 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2241 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2242 if (m4kc_tlbp_war())
2243 build_tlb_probe_entry(&p);
2244 build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2245 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2247 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2249 * This is the entry point when
2250 * build_r4000_tlbchange_handler_head spots a huge page.
2252 uasm_l_tlb_huge_update(&l, p);
2253 iPTE_LW(&p, wr.r1, wr.r2);
2254 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2255 build_tlb_probe_entry(&p);
2256 uasm_i_ori(&p, wr.r1, wr.r1,
2257 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2258 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2261 uasm_l_nopage_tlbs(&l, p);
2262 build_restore_work_registers(&p);
2263 #ifdef CONFIG_CPU_MICROMIPS
2264 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2265 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2266 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2270 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2273 if (p >= handle_tlbs_end)
2274 panic("TLB store handler fastpath space exceeded");
2276 uasm_resolve_relocs(relocs, labels);
2277 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2278 (unsigned int)(p - handle_tlbs));
2280 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2283 static void build_r4000_tlb_modify_handler(void)
2285 u32 *p = handle_tlbm;
2286 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2287 struct uasm_label *l = labels;
2288 struct uasm_reloc *r = relocs;
2289 struct work_registers wr;
2291 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2292 memset(labels, 0, sizeof(labels));
2293 memset(relocs, 0, sizeof(relocs));
2295 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2296 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2297 if (m4kc_tlbp_war())
2298 build_tlb_probe_entry(&p);
2299 /* Present and writable bits set, set accessed and dirty bits. */
2300 build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2301 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2303 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2305 * This is the entry point when
2306 * build_r4000_tlbchange_handler_head spots a huge page.
2308 uasm_l_tlb_huge_update(&l, p);
2309 iPTE_LW(&p, wr.r1, wr.r2);
2310 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2311 build_tlb_probe_entry(&p);
2312 uasm_i_ori(&p, wr.r1, wr.r1,
2313 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2314 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2317 uasm_l_nopage_tlbm(&l, p);
2318 build_restore_work_registers(&p);
2319 #ifdef CONFIG_CPU_MICROMIPS
2320 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2321 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2322 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2326 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2329 if (p >= handle_tlbm_end)
2330 panic("TLB modify handler fastpath space exceeded");
2332 uasm_resolve_relocs(relocs, labels);
2333 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2334 (unsigned int)(p - handle_tlbm));
2336 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2339 static void flush_tlb_handlers(void)
2341 local_flush_icache_range((unsigned long)handle_tlbl,
2342 (unsigned long)handle_tlbl_end);
2343 local_flush_icache_range((unsigned long)handle_tlbs,
2344 (unsigned long)handle_tlbs_end);
2345 local_flush_icache_range((unsigned long)handle_tlbm,
2346 (unsigned long)handle_tlbm_end);
2347 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2348 (unsigned long)tlbmiss_handler_setup_pgd_end);
2351 static void print_htw_config(void)
2353 unsigned long config;
2355 const int field = 2 * sizeof(unsigned long);
2357 config = read_c0_pwfield();
2358 pr_debug("PWField (0x%0*lx): GDI: 0x%02lx UDI: 0x%02lx MDI: 0x%02lx PTI: 0x%02lx PTEI: 0x%02lx\n",
2360 (config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2361 (config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2362 (config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2363 (config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2364 (config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2366 config = read_c0_pwsize();
2367 pr_debug("PWSize (0x%0*lx): PS: 0x%lx GDW: 0x%02lx UDW: 0x%02lx MDW: 0x%02lx PTW: 0x%02lx PTEW: 0x%02lx\n",
2369 (config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2370 (config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2371 (config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2372 (config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2373 (config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2374 (config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2376 pwctl = read_c0_pwctl();
2377 pr_debug("PWCtl (0x%x): PWEn: 0x%x XK: 0x%x XS: 0x%x XU: 0x%x DPH: 0x%x HugePg: 0x%x Psn: 0x%x\n",
2379 (pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2380 (pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2381 (pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2382 (pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2383 (pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2384 (pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2385 (pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2388 static void config_htw_params(void)
2390 unsigned long pwfield, pwsize, ptei;
2391 unsigned int config;
2394 * We are using 2-level page tables, so we only need to
2395 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2396 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2397 * write values less than 0xc in these fields because the entire
2398 * write will be dropped. As a result of which, we must preserve
2399 * the original reset values and overwrite only what we really want.
2402 pwfield = read_c0_pwfield();
2403 /* re-initialize the GDI field */
2404 pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2405 pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2406 /* re-initialize the PTI field including the even/odd bit */
2407 pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2408 pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2409 if (CONFIG_PGTABLE_LEVELS >= 3) {
2410 pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2411 pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2413 /* Set the PTEI right shift */
2414 ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2416 write_c0_pwfield(pwfield);
2417 /* Check whether the PTEI value is supported */
2418 back_to_back_c0_hazard();
2419 pwfield = read_c0_pwfield();
2420 if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2422 pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2425 * Drop option to avoid HTW being enabled via another path
2428 current_cpu_data.options &= ~MIPS_CPU_HTW;
2432 pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2433 pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2434 if (CONFIG_PGTABLE_LEVELS >= 3)
2435 pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2437 /* Set pointer size to size of directory pointers */
2438 if (IS_ENABLED(CONFIG_64BIT))
2439 pwsize |= MIPS_PWSIZE_PS_MASK;
2440 /* PTEs may be multiple pointers long (e.g. with XPA) */
2441 pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2442 & MIPS_PWSIZE_PTEW_MASK;
2444 write_c0_pwsize(pwsize);
2446 /* Make sure everything is set before we enable the HTW */
2447 back_to_back_c0_hazard();
2450 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2453 config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2454 if (IS_ENABLED(CONFIG_64BIT))
2455 config |= MIPS_PWCTL_XU_MASK;
2456 write_c0_pwctl(config);
2457 pr_info("Hardware Page Table Walker enabled\n");
2462 static void config_xpa_params(void)
2465 unsigned int pagegrain;
2467 if (mips_xpa_disabled) {
2468 pr_info("Extended Physical Addressing (XPA) disabled\n");
2472 pagegrain = read_c0_pagegrain();
2473 write_c0_pagegrain(pagegrain | PG_ELPA);
2474 back_to_back_c0_hazard();
2475 pagegrain = read_c0_pagegrain();
2477 if (pagegrain & PG_ELPA)
2478 pr_info("Extended Physical Addressing (XPA) enabled\n");
2480 panic("Extended Physical Addressing (XPA) disabled");
2484 static void check_pabits(void)
2486 unsigned long entry;
2487 unsigned pabits, fillbits;
2489 if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2491 * We'll only be making use of the fact that we can rotate bits
2492 * into the fill if the CPU supports RIXI, so don't bother
2493 * probing this for CPUs which don't.
2498 write_c0_entrylo0(~0ul);
2499 back_to_back_c0_hazard();
2500 entry = read_c0_entrylo0();
2502 /* clear all non-PFN bits */
2503 entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2504 entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2506 /* find a lower bound on PABITS, and upper bound on fill bits */
2507 pabits = fls_long(entry) + 6;
2508 fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2510 /* minus the RI & XI bits */
2511 fillbits -= min_t(unsigned, fillbits, 2);
2513 if (fillbits >= ilog2(_PAGE_NO_EXEC))
2514 fill_includes_sw_bits = true;
2516 pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2519 void build_tlb_refill_handler(void)
2522 * The refill handler is generated per-CPU, multi-node systems
2523 * may have local storage for it. The other handlers are only
2526 static int run_once = 0;
2528 if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2529 panic("Kernels supporting XPA currently require CPUs with RIXI");
2531 output_pgtable_bits_defines();
2535 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2538 switch (current_cpu_type()) {
2546 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2547 if (cpu_has_local_ebase)
2548 build_r3000_tlb_refill_handler();
2550 if (!cpu_has_local_ebase)
2551 build_r3000_tlb_refill_handler();
2553 build_r3000_tlb_load_handler();
2554 build_r3000_tlb_store_handler();
2555 build_r3000_tlb_modify_handler();
2556 flush_tlb_handlers();
2560 panic("No R3000 TLB refill handler");
2566 panic("No R6000 TLB refill handler yet");
2570 panic("No R8000 TLB refill handler yet");
2578 scratch_reg = allocate_kscratch();
2580 build_r4000_tlb_load_handler();
2581 build_r4000_tlb_store_handler();
2582 build_r4000_tlb_modify_handler();
2584 build_loongson3_tlb_refill_handler();
2585 else if (!cpu_has_local_ebase)
2586 build_r4000_tlb_refill_handler();
2587 flush_tlb_handlers();
2590 if (cpu_has_local_ebase)
2591 build_r4000_tlb_refill_handler();
2593 config_xpa_params();
2595 config_htw_params();