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
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>
31 #include <linux/pgtable.h>
33 #include <asm/cacheflush.h>
34 #include <asm/cpu-type.h>
35 #include <asm/mmu_context.h>
38 #include <asm/setup.h>
39 #include <asm/tlbex.h>
41 static int mips_xpa_disabled;
43 static int __init xpa_disable(char *s)
45 mips_xpa_disabled = 1;
50 __setup("noxpa", xpa_disable);
53 * TLB load/store/modify handlers.
55 * Only the fastpath gets synthesized at runtime, the slowpath for
56 * do_page_fault remains normal asm.
58 extern void tlb_do_page_fault_0(void);
59 extern void tlb_do_page_fault_1(void);
61 struct work_registers {
70 } ____cacheline_aligned_in_smp;
72 static struct tlb_reg_save handler_reg_save[NR_CPUS];
74 static inline int r45k_bvahwbug(void)
76 /* XXX: We should probe for the presence of this bug, but we don't. */
80 static inline int r4k_250MHZhwbug(void)
82 /* XXX: We should probe for the presence of this bug, but we don't. */
86 extern int sb1250_m3_workaround_needed(void);
88 static inline int __maybe_unused bcm1250_m3_war(void)
90 if (IS_ENABLED(CONFIG_SB1_PASS_2_WORKAROUNDS))
91 return sb1250_m3_workaround_needed();
95 static inline int __maybe_unused r10000_llsc_war(void)
97 return IS_ENABLED(CONFIG_WAR_R10000_LLSC);
100 static int use_bbit_insns(void)
102 switch (current_cpu_type()) {
103 case CPU_CAVIUM_OCTEON:
104 case CPU_CAVIUM_OCTEON_PLUS:
105 case CPU_CAVIUM_OCTEON2:
106 case CPU_CAVIUM_OCTEON3:
113 static int use_lwx_insns(void)
115 switch (current_cpu_type()) {
116 case CPU_CAVIUM_OCTEON2:
117 case CPU_CAVIUM_OCTEON3:
123 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
124 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
125 static bool scratchpad_available(void)
129 static int scratchpad_offset(int i)
132 * CVMSEG starts at address -32768 and extends for
133 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
135 i += 1; /* Kernel use starts at the top and works down. */
136 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
139 static bool scratchpad_available(void)
143 static int scratchpad_offset(int i)
146 /* Really unreachable, but evidently some GCC want this. */
151 * Found by experiment: At least some revisions of the 4kc throw under
152 * some circumstances a machine check exception, triggered by invalid
153 * values in the index register. Delaying the tlbp instruction until
154 * after the next branch, plus adding an additional nop in front of
155 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
156 * why; it's not an issue caused by the core RTL.
159 static int m4kc_tlbp_war(void)
161 return current_cpu_type() == CPU_4KC;
164 /* Handle labels (which must be positive integers). */
166 label_second_part = 1,
171 label_split = label_tlbw_hazard_0 + 8,
172 label_tlbl_goaround1,
173 label_tlbl_goaround2,
177 label_smp_pgtable_change,
178 label_r3000_write_probe_fail,
179 label_large_segbits_fault,
180 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
181 label_tlb_huge_update,
185 UASM_L_LA(_second_part)
188 UASM_L_LA(_vmalloc_done)
189 /* _tlbw_hazard_x is handled differently. */
191 UASM_L_LA(_tlbl_goaround1)
192 UASM_L_LA(_tlbl_goaround2)
193 UASM_L_LA(_nopage_tlbl)
194 UASM_L_LA(_nopage_tlbs)
195 UASM_L_LA(_nopage_tlbm)
196 UASM_L_LA(_smp_pgtable_change)
197 UASM_L_LA(_r3000_write_probe_fail)
198 UASM_L_LA(_large_segbits_fault)
199 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
200 UASM_L_LA(_tlb_huge_update)
203 static int hazard_instance;
205 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
209 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
216 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
220 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
228 * pgtable bits are assigned dynamically depending on processor feature
229 * and statically based on kernel configuration. This spits out the actual
230 * values the kernel is using. Required to make sense from disassembled
231 * TLB exception handlers.
233 static void output_pgtable_bits_defines(void)
235 #define pr_define(fmt, ...) \
236 pr_debug("#define " fmt, ##__VA_ARGS__)
238 pr_debug("#include <asm/asm.h>\n");
239 pr_debug("#include <asm/regdef.h>\n");
242 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
243 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
244 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
245 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
246 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
247 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
248 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
250 #ifdef _PAGE_NO_EXEC_SHIFT
252 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
254 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
255 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
256 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
257 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
261 static inline void dump_handler(const char *symbol, const void *start, const void *end)
263 unsigned int count = (end - start) / sizeof(u32);
264 const u32 *handler = start;
267 pr_debug("LEAF(%s)\n", symbol);
269 pr_debug("\t.set push\n");
270 pr_debug("\t.set noreorder\n");
272 for (i = 0; i < count; i++)
273 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
275 pr_debug("\t.set\tpop\n");
277 pr_debug("\tEND(%s)\n", symbol);
280 /* The only general purpose registers allowed in TLB handlers. */
284 /* Some CP0 registers */
285 #define C0_INDEX 0, 0
286 #define C0_ENTRYLO0 2, 0
287 #define C0_TCBIND 2, 2
288 #define C0_ENTRYLO1 3, 0
289 #define C0_CONTEXT 4, 0
290 #define C0_PAGEMASK 5, 0
291 #define C0_PWBASE 5, 5
292 #define C0_PWFIELD 5, 6
293 #define C0_PWSIZE 5, 7
294 #define C0_PWCTL 6, 6
295 #define C0_BADVADDR 8, 0
297 #define C0_ENTRYHI 10, 0
299 #define C0_XCONTEXT 20, 0
302 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
304 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
307 /* The worst case length of the handler is around 18 instructions for
308 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
309 * Maximum space available is 32 instructions for R3000 and 64
310 * instructions for R4000.
312 * We deliberately chose a buffer size of 128, so we won't scribble
313 * over anything important on overflow before we panic.
315 static u32 tlb_handler[128];
317 /* simply assume worst case size for labels and relocs */
318 static struct uasm_label labels[128];
319 static struct uasm_reloc relocs[128];
321 static int check_for_high_segbits;
322 static bool fill_includes_sw_bits;
324 static unsigned int kscratch_used_mask;
326 static inline int __maybe_unused c0_kscratch(void)
328 switch (current_cpu_type()) {
337 static int allocate_kscratch(void)
340 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
347 r--; /* make it zero based */
349 kscratch_used_mask |= (1 << r);
354 static int scratch_reg;
356 EXPORT_SYMBOL_GPL(pgd_reg);
357 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
359 static struct work_registers build_get_work_registers(u32 **p)
361 struct work_registers r;
363 if (scratch_reg >= 0) {
364 /* Save in CPU local C0_KScratch? */
365 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
372 if (num_possible_cpus() > 1) {
373 /* Get smp_processor_id */
374 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
375 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
377 /* handler_reg_save index in K0 */
378 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
380 UASM_i_LA(p, K1, (long)&handler_reg_save);
381 UASM_i_ADDU(p, K0, K0, K1);
383 UASM_i_LA(p, K0, (long)&handler_reg_save);
385 /* K0 now points to save area, save $1 and $2 */
386 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
387 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
395 static void build_restore_work_registers(u32 **p)
397 if (scratch_reg >= 0) {
399 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
402 /* K0 already points to save area, restore $1 and $2 */
403 UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
404 UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
407 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
410 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
411 * we cannot do r3000 under these circumstances.
413 * The R3000 TLB handler is simple.
415 static void build_r3000_tlb_refill_handler(void)
417 long pgdc = (long)pgd_current;
420 memset(tlb_handler, 0, sizeof(tlb_handler));
423 uasm_i_mfc0(&p, K0, C0_BADVADDR);
424 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
425 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
426 uasm_i_srl(&p, K0, K0, 22); /* load delay */
427 uasm_i_sll(&p, K0, K0, 2);
428 uasm_i_addu(&p, K1, K1, K0);
429 uasm_i_mfc0(&p, K0, C0_CONTEXT);
430 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
431 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
432 uasm_i_addu(&p, K1, K1, K0);
433 uasm_i_lw(&p, K0, 0, K1);
434 uasm_i_nop(&p); /* load delay */
435 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
436 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
437 uasm_i_tlbwr(&p); /* cp0 delay */
439 uasm_i_rfe(&p); /* branch delay */
441 if (p > tlb_handler + 32)
442 panic("TLB refill handler space exceeded");
444 pr_debug("Wrote TLB refill handler (%u instructions).\n",
445 (unsigned int)(p - tlb_handler));
447 memcpy((void *)ebase, tlb_handler, 0x80);
448 local_flush_icache_range(ebase, ebase + 0x80);
449 dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
451 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
454 * The R4000 TLB handler is much more complicated. We have two
455 * consecutive handler areas with 32 instructions space each.
456 * Since they aren't used at the same time, we can overflow in the
457 * other one.To keep things simple, we first assume linear space,
458 * then we relocate it to the final handler layout as needed.
460 static u32 final_handler[64];
465 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
466 * 2. A timing hazard exists for the TLBP instruction.
468 * stalling_instruction
471 * The JTLB is being read for the TLBP throughout the stall generated by the
472 * previous instruction. This is not really correct as the stalling instruction
473 * can modify the address used to access the JTLB. The failure symptom is that
474 * the TLBP instruction will use an address created for the stalling instruction
475 * and not the address held in C0_ENHI and thus report the wrong results.
477 * The software work-around is to not allow the instruction preceding the TLBP
478 * to stall - make it an NOP or some other instruction guaranteed not to stall.
480 * Errata 2 will not be fixed. This errata is also on the R5000.
482 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
484 static void __maybe_unused build_tlb_probe_entry(u32 **p)
486 switch (current_cpu_type()) {
487 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
502 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
503 struct uasm_reloc **r,
504 enum tlb_write_entry wmode)
506 void(*tlbw)(u32 **) = NULL;
509 case tlb_random: tlbw = uasm_i_tlbwr; break;
510 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
513 if (cpu_has_mips_r2_r6) {
514 if (cpu_has_mips_r2_exec_hazard)
520 switch (current_cpu_type()) {
528 * This branch uses up a mtc0 hazard nop slot and saves
529 * two nops after the tlbw instruction.
531 uasm_bgezl_hazard(p, r, hazard_instance);
533 uasm_bgezl_label(l, p, hazard_instance);
547 uasm_i_nop(p); /* QED specifies 2 nops hazard */
548 uasm_i_nop(p); /* QED specifies 2 nops hazard */
579 case CPU_LOONGSON2EF:
622 panic("No TLB refill handler yet (CPU type: %d)",
627 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
629 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
632 if (_PAGE_GLOBAL_SHIFT == 0) {
633 /* pte_t is already in EntryLo format */
637 if (cpu_has_rixi && !!_PAGE_NO_EXEC) {
638 if (fill_includes_sw_bits) {
639 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
641 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
642 UASM_i_ROTR(p, reg, reg,
643 ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
646 #ifdef CONFIG_PHYS_ADDR_T_64BIT
647 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
649 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
654 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
656 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
657 unsigned int tmp, enum label_id lid,
660 if (restore_scratch) {
662 * Ensure the MFC0 below observes the value written to the
663 * KScratch register by the prior MTC0.
665 if (scratch_reg >= 0)
668 /* Reset default page size */
669 if (PM_DEFAULT_MASK >> 16) {
670 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
671 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
672 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
673 uasm_il_b(p, r, lid);
674 } else if (PM_DEFAULT_MASK) {
675 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
676 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
677 uasm_il_b(p, r, lid);
679 uasm_i_mtc0(p, 0, C0_PAGEMASK);
680 uasm_il_b(p, r, lid);
682 if (scratch_reg >= 0)
683 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
685 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
687 /* Reset default page size */
688 if (PM_DEFAULT_MASK >> 16) {
689 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
690 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
691 uasm_il_b(p, r, lid);
692 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
693 } else if (PM_DEFAULT_MASK) {
694 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
695 uasm_il_b(p, r, lid);
696 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
698 uasm_il_b(p, r, lid);
699 uasm_i_mtc0(p, 0, C0_PAGEMASK);
704 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
705 struct uasm_reloc **r,
707 enum tlb_write_entry wmode,
710 /* Set huge page tlb entry size */
711 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
712 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
713 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
715 build_tlb_write_entry(p, l, r, wmode);
717 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
721 * Check if Huge PTE is present, if so then jump to LABEL.
724 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
725 unsigned int pmd, int lid)
727 UASM_i_LW(p, tmp, 0, pmd);
728 if (use_bbit_insns()) {
729 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
731 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
732 uasm_il_bnez(p, r, tmp, lid);
736 static void build_huge_update_entries(u32 **p, unsigned int pte,
742 * A huge PTE describes an area the size of the
743 * configured huge page size. This is twice the
744 * of the large TLB entry size we intend to use.
745 * A TLB entry half the size of the configured
746 * huge page size is configured into entrylo0
747 * and entrylo1 to cover the contiguous huge PTE
750 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
752 /* We can clobber tmp. It isn't used after this.*/
754 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
756 build_convert_pte_to_entrylo(p, pte);
757 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
758 /* convert to entrylo1 */
760 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
762 UASM_i_ADDU(p, pte, pte, tmp);
764 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
767 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
768 struct uasm_label **l,
774 UASM_i_SC(p, pte, 0, ptr);
775 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
776 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
778 UASM_i_SW(p, pte, 0, ptr);
780 if (cpu_has_ftlb && flush) {
781 BUG_ON(!cpu_has_tlbinv);
783 UASM_i_MFC0(p, ptr, C0_ENTRYHI);
784 uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
785 UASM_i_MTC0(p, ptr, C0_ENTRYHI);
786 build_tlb_write_entry(p, l, r, tlb_indexed);
788 uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
789 UASM_i_MTC0(p, ptr, C0_ENTRYHI);
790 build_huge_update_entries(p, pte, ptr);
791 build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
796 build_huge_update_entries(p, pte, ptr);
797 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
799 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
803 * TMP and PTR are scratch.
804 * TMP will be clobbered, PTR will hold the pmd entry.
806 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
807 unsigned int tmp, unsigned int ptr)
809 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
810 long pgdc = (long)pgd_current;
813 * The vmalloc handling is not in the hotpath.
815 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
817 if (check_for_high_segbits) {
819 * The kernel currently implicitely assumes that the
820 * MIPS SEGBITS parameter for the processor is
821 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
822 * allocate virtual addresses outside the maximum
823 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
824 * that doesn't prevent user code from accessing the
825 * higher xuseg addresses. Here, we make sure that
826 * everything but the lower xuseg addresses goes down
827 * the module_alloc/vmalloc path.
829 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
830 uasm_il_bnez(p, r, ptr, label_vmalloc);
832 uasm_il_bltz(p, r, tmp, label_vmalloc);
834 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
837 /* pgd is in pgd_reg */
839 UASM_i_MFC0(p, ptr, C0_PWBASE);
841 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
843 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
845 * &pgd << 11 stored in CONTEXT [23..63].
847 UASM_i_MFC0(p, ptr, C0_CONTEXT);
849 /* Clear lower 23 bits of context. */
850 uasm_i_dins(p, ptr, 0, 0, 23);
852 /* insert bit[63:59] of CAC_BASE into bit[11:6] of ptr */
853 uasm_i_ori(p, ptr, ptr, ((u64)(CAC_BASE) >> 53));
854 uasm_i_drotr(p, ptr, ptr, 11);
855 #elif defined(CONFIG_SMP)
856 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
857 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
858 UASM_i_LA_mostly(p, tmp, pgdc);
859 uasm_i_daddu(p, ptr, ptr, tmp);
860 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
861 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
863 UASM_i_LA_mostly(p, ptr, pgdc);
864 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
868 uasm_l_vmalloc_done(l, *p);
870 /* get pgd offset in bytes */
871 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
873 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
874 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
875 #ifndef __PAGETABLE_PUD_FOLDED
876 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
877 uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
878 uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
879 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
880 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
882 #ifndef __PAGETABLE_PMD_FOLDED
883 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
884 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
885 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
886 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
887 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
890 EXPORT_SYMBOL_GPL(build_get_pmde64);
893 * BVADDR is the faulting address, PTR is scratch.
894 * PTR will hold the pgd for vmalloc.
897 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
898 unsigned int bvaddr, unsigned int ptr,
899 enum vmalloc64_mode mode)
901 long swpd = (long)swapper_pg_dir;
902 int single_insn_swpd;
903 int did_vmalloc_branch = 0;
905 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
907 uasm_l_vmalloc(l, *p);
909 if (mode != not_refill && check_for_high_segbits) {
910 if (single_insn_swpd) {
911 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
912 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
913 did_vmalloc_branch = 1;
916 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
919 if (!did_vmalloc_branch) {
920 if (single_insn_swpd) {
921 uasm_il_b(p, r, label_vmalloc_done);
922 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
924 UASM_i_LA_mostly(p, ptr, swpd);
925 uasm_il_b(p, r, label_vmalloc_done);
926 if (uasm_in_compat_space_p(swpd))
927 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
929 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
932 if (mode != not_refill && check_for_high_segbits) {
933 uasm_l_large_segbits_fault(l, *p);
935 if (mode == refill_scratch && scratch_reg >= 0)
939 * We get here if we are an xsseg address, or if we are
940 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
942 * Ignoring xsseg (assume disabled so would generate
943 * (address errors?), the only remaining possibility
944 * is the upper xuseg addresses. On processors with
945 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
946 * addresses would have taken an address error. We try
947 * to mimic that here by taking a load/istream page
950 if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
952 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
955 if (mode == refill_scratch) {
956 if (scratch_reg >= 0)
957 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
959 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
966 #else /* !CONFIG_64BIT */
969 * TMP and PTR are scratch.
970 * TMP will be clobbered, PTR will hold the pgd entry.
972 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
975 /* pgd is in pgd_reg */
976 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
977 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
979 long pgdc = (long)pgd_current;
981 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
983 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
984 UASM_i_LA_mostly(p, tmp, pgdc);
985 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
986 uasm_i_addu(p, ptr, tmp, ptr);
988 UASM_i_LA_mostly(p, ptr, pgdc);
990 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
991 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
993 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
994 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
995 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
997 EXPORT_SYMBOL_GPL(build_get_pgde32);
999 #endif /* !CONFIG_64BIT */
1001 static void build_adjust_context(u32 **p, unsigned int ctx)
1003 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
1004 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
1006 switch (current_cpu_type()) {
1023 UASM_i_SRL(p, ctx, ctx, shift);
1024 uasm_i_andi(p, ctx, ctx, mask);
1027 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1030 * Bug workaround for the Nevada. It seems as if under certain
1031 * circumstances the move from cp0_context might produce a
1032 * bogus result when the mfc0 instruction and its consumer are
1033 * in a different cacheline or a load instruction, probably any
1034 * memory reference, is between them.
1036 switch (current_cpu_type()) {
1038 UASM_i_LW(p, ptr, 0, ptr);
1039 GET_CONTEXT(p, tmp); /* get context reg */
1043 GET_CONTEXT(p, tmp); /* get context reg */
1044 UASM_i_LW(p, ptr, 0, ptr);
1048 build_adjust_context(p, tmp);
1049 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1051 EXPORT_SYMBOL_GPL(build_get_ptep);
1053 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1055 int pte_off_even = 0;
1056 int pte_off_odd = sizeof(pte_t);
1058 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1059 /* The low 32 bits of EntryLo is stored in pte_high */
1060 pte_off_even += offsetof(pte_t, pte_high);
1061 pte_off_odd += offsetof(pte_t, pte_high);
1064 if (IS_ENABLED(CONFIG_XPA)) {
1065 uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1066 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1067 UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1069 if (cpu_has_xpa && !mips_xpa_disabled) {
1070 uasm_i_lw(p, tmp, 0, ptep);
1071 uasm_i_ext(p, tmp, tmp, 0, 24);
1072 uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1075 uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1076 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1077 UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1079 if (cpu_has_xpa && !mips_xpa_disabled) {
1080 uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1081 uasm_i_ext(p, tmp, tmp, 0, 24);
1082 uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1087 UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1088 UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1089 if (r45k_bvahwbug())
1090 build_tlb_probe_entry(p);
1091 build_convert_pte_to_entrylo(p, tmp);
1092 if (r4k_250MHZhwbug())
1093 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1094 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1095 build_convert_pte_to_entrylo(p, ptep);
1096 if (r45k_bvahwbug())
1097 uasm_i_mfc0(p, tmp, C0_INDEX);
1098 if (r4k_250MHZhwbug())
1099 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1100 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1102 EXPORT_SYMBOL_GPL(build_update_entries);
1104 struct mips_huge_tlb_info {
1106 int restore_scratch;
1107 bool need_reload_pte;
1110 static struct mips_huge_tlb_info
1111 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1112 struct uasm_reloc **r, unsigned int tmp,
1113 unsigned int ptr, int c0_scratch_reg)
1115 struct mips_huge_tlb_info rv;
1116 unsigned int even, odd;
1117 int vmalloc_branch_delay_filled = 0;
1118 const int scratch = 1; /* Our extra working register */
1120 rv.huge_pte = scratch;
1121 rv.restore_scratch = 0;
1122 rv.need_reload_pte = false;
1124 if (check_for_high_segbits) {
1125 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1128 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1130 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1132 if (c0_scratch_reg >= 0)
1133 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1135 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1137 uasm_i_dsrl_safe(p, scratch, tmp,
1138 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1139 uasm_il_bnez(p, r, scratch, label_vmalloc);
1141 if (pgd_reg == -1) {
1142 vmalloc_branch_delay_filled = 1;
1143 /* Clear lower 23 bits of context. */
1144 uasm_i_dins(p, ptr, 0, 0, 23);
1148 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1150 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1152 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1154 if (c0_scratch_reg >= 0)
1155 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1157 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1160 /* Clear lower 23 bits of context. */
1161 uasm_i_dins(p, ptr, 0, 0, 23);
1163 uasm_il_bltz(p, r, tmp, label_vmalloc);
1166 if (pgd_reg == -1) {
1167 vmalloc_branch_delay_filled = 1;
1168 /* insert bit[63:59] of CAC_BASE into bit[11:6] of ptr */
1169 uasm_i_ori(p, ptr, ptr, ((u64)(CAC_BASE) >> 53));
1171 uasm_i_drotr(p, ptr, ptr, 11);
1174 #ifdef __PAGETABLE_PMD_FOLDED
1175 #define LOC_PTEP scratch
1177 #define LOC_PTEP ptr
1180 if (!vmalloc_branch_delay_filled)
1181 /* get pgd offset in bytes */
1182 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1184 uasm_l_vmalloc_done(l, *p);
1188 * fall-through case = badvaddr *pgd_current
1189 * vmalloc case = badvaddr swapper_pg_dir
1192 if (vmalloc_branch_delay_filled)
1193 /* get pgd offset in bytes */
1194 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1196 #ifdef __PAGETABLE_PMD_FOLDED
1197 GET_CONTEXT(p, tmp); /* get context reg */
1199 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1201 if (use_lwx_insns()) {
1202 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1204 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1205 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1208 #ifndef __PAGETABLE_PUD_FOLDED
1209 /* get pud offset in bytes */
1210 uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1211 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1213 if (use_lwx_insns()) {
1214 UASM_i_LWX(p, ptr, scratch, ptr);
1216 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1217 UASM_i_LW(p, ptr, 0, ptr);
1219 /* ptr contains a pointer to PMD entry */
1220 /* tmp contains the address */
1223 #ifndef __PAGETABLE_PMD_FOLDED
1224 /* get pmd offset in bytes */
1225 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1226 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1227 GET_CONTEXT(p, tmp); /* get context reg */
1229 if (use_lwx_insns()) {
1230 UASM_i_LWX(p, scratch, scratch, ptr);
1232 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1233 UASM_i_LW(p, scratch, 0, ptr);
1236 /* Adjust the context during the load latency. */
1237 build_adjust_context(p, tmp);
1239 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1240 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1242 * The in the LWX case we don't want to do the load in the
1243 * delay slot. It cannot issue in the same cycle and may be
1244 * speculative and unneeded.
1246 if (use_lwx_insns())
1248 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1251 /* build_update_entries */
1252 if (use_lwx_insns()) {
1255 UASM_i_LWX(p, even, scratch, tmp);
1256 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1257 UASM_i_LWX(p, odd, scratch, tmp);
1259 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1262 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1263 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1266 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1267 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1268 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1270 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1271 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1272 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1274 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1276 if (c0_scratch_reg >= 0) {
1278 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1279 build_tlb_write_entry(p, l, r, tlb_random);
1280 uasm_l_leave(l, *p);
1281 rv.restore_scratch = 1;
1282 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1283 build_tlb_write_entry(p, l, r, tlb_random);
1284 uasm_l_leave(l, *p);
1285 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1287 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1288 build_tlb_write_entry(p, l, r, tlb_random);
1289 uasm_l_leave(l, *p);
1290 rv.restore_scratch = 1;
1293 uasm_i_eret(p); /* return from trap */
1299 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1300 * because EXL == 0. If we wrap, we can also use the 32 instruction
1301 * slots before the XTLB refill exception handler which belong to the
1302 * unused TLB refill exception.
1304 #define MIPS64_REFILL_INSNS 32
1306 static void build_r4000_tlb_refill_handler(void)
1308 u32 *p = tlb_handler;
1309 struct uasm_label *l = labels;
1310 struct uasm_reloc *r = relocs;
1312 unsigned int final_len;
1313 struct mips_huge_tlb_info htlb_info __maybe_unused;
1314 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1316 memset(tlb_handler, 0, sizeof(tlb_handler));
1317 memset(labels, 0, sizeof(labels));
1318 memset(relocs, 0, sizeof(relocs));
1319 memset(final_handler, 0, sizeof(final_handler));
1321 if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1322 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1324 vmalloc_mode = refill_scratch;
1326 htlb_info.huge_pte = K0;
1327 htlb_info.restore_scratch = 0;
1328 htlb_info.need_reload_pte = true;
1329 vmalloc_mode = refill_noscratch;
1331 * create the plain linear handler
1333 if (bcm1250_m3_war()) {
1334 unsigned int segbits = 44;
1336 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1337 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1338 uasm_i_xor(&p, K0, K0, K1);
1339 uasm_i_dsrl_safe(&p, K1, K0, 62);
1340 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1341 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1342 uasm_i_or(&p, K0, K0, K1);
1343 uasm_il_bnez(&p, &r, K0, label_leave);
1344 /* No need for uasm_i_nop */
1348 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1350 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1353 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1354 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1357 build_get_ptep(&p, K0, K1);
1358 build_update_entries(&p, K0, K1);
1359 build_tlb_write_entry(&p, &l, &r, tlb_random);
1360 uasm_l_leave(&l, p);
1361 uasm_i_eret(&p); /* return from trap */
1363 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1364 uasm_l_tlb_huge_update(&l, p);
1365 if (htlb_info.need_reload_pte)
1366 UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1367 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1368 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1369 htlb_info.restore_scratch);
1373 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1377 * Overflow check: For the 64bit handler, we need at least one
1378 * free instruction slot for the wrap-around branch. In worst
1379 * case, if the intended insertion point is a delay slot, we
1380 * need three, with the second nop'ed and the third being
1383 switch (boot_cpu_type()) {
1385 if (sizeof(long) == 4) {
1387 case CPU_LOONGSON2EF:
1388 /* Loongson2 ebase is different than r4k, we have more space */
1389 if ((p - tlb_handler) > 64)
1390 panic("TLB refill handler space exceeded");
1392 * Now fold the handler in the TLB refill handler space.
1395 /* Simplest case, just copy the handler. */
1396 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1397 final_len = p - tlb_handler;
1400 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1401 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1402 && uasm_insn_has_bdelay(relocs,
1403 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1404 panic("TLB refill handler space exceeded");
1406 * Now fold the handler in the TLB refill handler space.
1408 f = final_handler + MIPS64_REFILL_INSNS;
1409 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1410 /* Just copy the handler. */
1411 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1412 final_len = p - tlb_handler;
1414 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1415 const enum label_id ls = label_tlb_huge_update;
1417 const enum label_id ls = label_vmalloc;
1423 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1425 BUG_ON(i == ARRAY_SIZE(labels));
1426 split = labels[i].addr;
1429 * See if we have overflown one way or the other.
1431 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1432 split < p - MIPS64_REFILL_INSNS)
1437 * Split two instructions before the end. One
1438 * for the branch and one for the instruction
1439 * in the delay slot.
1441 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1444 * If the branch would fall in a delay slot,
1445 * we must back up an additional instruction
1446 * so that it is no longer in a delay slot.
1448 if (uasm_insn_has_bdelay(relocs, split - 1))
1451 /* Copy first part of the handler. */
1452 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1453 f += split - tlb_handler;
1456 /* Insert branch. */
1457 uasm_l_split(&l, final_handler);
1458 uasm_il_b(&f, &r, label_split);
1459 if (uasm_insn_has_bdelay(relocs, split))
1462 uasm_copy_handler(relocs, labels,
1463 split, split + 1, f);
1464 uasm_move_labels(labels, f, f + 1, -1);
1470 /* Copy the rest of the handler. */
1471 uasm_copy_handler(relocs, labels, split, p, final_handler);
1472 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1479 uasm_resolve_relocs(relocs, labels);
1480 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1483 memcpy((void *)ebase, final_handler, 0x100);
1484 local_flush_icache_range(ebase, ebase + 0x100);
1485 dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1488 static void setup_pw(void)
1491 unsigned long pgd_i, pgd_w;
1492 #ifndef __PAGETABLE_PMD_FOLDED
1493 unsigned long pmd_i, pmd_w;
1495 unsigned long pt_i, pt_w;
1496 unsigned long pte_i, pte_w;
1497 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1500 psn = ilog2(_PAGE_HUGE); /* bit used to indicate huge page */
1502 pgd_i = PGDIR_SHIFT; /* 1st level PGD */
1503 #ifndef __PAGETABLE_PMD_FOLDED
1504 pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1506 pmd_i = PMD_SHIFT; /* 2nd level PMD */
1507 pmd_w = PMD_SHIFT - PAGE_SHIFT;
1509 pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1512 pt_i = PAGE_SHIFT; /* 3rd level PTE */
1513 pt_w = PAGE_SHIFT - 3;
1515 pte_i = ilog2(_PAGE_GLOBAL);
1517 pwctl = 1 << 30; /* Set PWDirExt */
1519 #ifndef __PAGETABLE_PMD_FOLDED
1520 write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1521 write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1523 write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1524 write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1527 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1528 pwctl |= (1 << 6 | psn);
1530 write_c0_pwctl(pwctl);
1531 write_c0_kpgd((long)swapper_pg_dir);
1532 kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1535 static void build_loongson3_tlb_refill_handler(void)
1537 u32 *p = tlb_handler;
1538 struct uasm_label *l = labels;
1539 struct uasm_reloc *r = relocs;
1541 memset(labels, 0, sizeof(labels));
1542 memset(relocs, 0, sizeof(relocs));
1543 memset(tlb_handler, 0, sizeof(tlb_handler));
1545 if (check_for_high_segbits) {
1546 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1547 uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1548 uasm_il_beqz(&p, &r, K1, label_vmalloc);
1551 uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1553 uasm_l_vmalloc(&l, p);
1556 uasm_i_dmfc0(&p, K1, C0_PGD);
1558 uasm_i_lddir(&p, K0, K1, 3); /* global page dir */
1559 #ifndef __PAGETABLE_PMD_FOLDED
1560 uasm_i_lddir(&p, K1, K0, 1); /* middle page dir */
1562 uasm_i_ldpte(&p, K1, 0); /* even */
1563 uasm_i_ldpte(&p, K1, 1); /* odd */
1566 /* restore page mask */
1567 if (PM_DEFAULT_MASK >> 16) {
1568 uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1569 uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1570 uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1571 } else if (PM_DEFAULT_MASK) {
1572 uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1573 uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1575 uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1580 if (check_for_high_segbits) {
1581 uasm_l_large_segbits_fault(&l, p);
1582 UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1587 uasm_resolve_relocs(relocs, labels);
1588 memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1589 local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1590 dump_handler("loongson3_tlb_refill",
1591 (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1594 static void build_setup_pgd(void)
1597 const int __maybe_unused a1 = 5;
1598 const int __maybe_unused a2 = 6;
1599 u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1600 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1601 long pgdc = (long)pgd_current;
1604 memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1605 memset(labels, 0, sizeof(labels));
1606 memset(relocs, 0, sizeof(relocs));
1607 pgd_reg = allocate_kscratch();
1608 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1609 if (pgd_reg == -1) {
1610 struct uasm_label *l = labels;
1611 struct uasm_reloc *r = relocs;
1613 /* PGD << 11 in c0_Context */
1615 * If it is a ckseg0 address, convert to a physical
1616 * address. Shifting right by 29 and adding 4 will
1617 * result in zero for these addresses.
1620 UASM_i_SRA(&p, a1, a0, 29);
1621 UASM_i_ADDIU(&p, a1, a1, 4);
1622 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1624 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1625 uasm_l_tlbl_goaround1(&l, p);
1626 UASM_i_SLL(&p, a0, a0, 11);
1627 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1631 /* PGD in c0_KScratch */
1633 UASM_i_MTC0(&p, a0, C0_PWBASE);
1635 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1641 /* Save PGD to pgd_current[smp_processor_id()] */
1642 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1643 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1644 UASM_i_LA_mostly(&p, a2, pgdc);
1645 UASM_i_ADDU(&p, a2, a2, a1);
1646 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1648 UASM_i_LA_mostly(&p, a2, pgdc);
1649 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1652 /* if pgd_reg is allocated, save PGD also to scratch register */
1653 if (pgd_reg != -1) {
1654 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1662 if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1663 panic("tlbmiss_handler_setup_pgd space exceeded");
1665 uasm_resolve_relocs(relocs, labels);
1666 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1667 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1669 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1670 tlbmiss_handler_setup_pgd_end);
1674 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1677 if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
1679 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1681 uasm_i_lld(p, pte, 0, ptr);
1684 UASM_i_LL(p, pte, 0, ptr);
1686 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1688 uasm_i_ld(p, pte, 0, ptr);
1691 UASM_i_LW(p, pte, 0, ptr);
1696 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1697 unsigned int mode, unsigned int scratch)
1699 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1700 unsigned int swmode = mode & ~hwmode;
1702 if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1703 uasm_i_lui(p, scratch, swmode >> 16);
1704 uasm_i_or(p, pte, pte, scratch);
1705 BUG_ON(swmode & 0xffff);
1707 uasm_i_ori(p, pte, pte, mode);
1711 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1713 uasm_i_scd(p, pte, 0, ptr);
1716 UASM_i_SC(p, pte, 0, ptr);
1718 if (r10000_llsc_war())
1719 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1721 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1723 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1724 if (!cpu_has_64bits) {
1725 /* no uasm_i_nop needed */
1726 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1727 uasm_i_ori(p, pte, pte, hwmode);
1728 BUG_ON(hwmode & ~0xffff);
1729 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1730 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1731 /* no uasm_i_nop needed */
1732 uasm_i_lw(p, pte, 0, ptr);
1739 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1741 uasm_i_sd(p, pte, 0, ptr);
1744 UASM_i_SW(p, pte, 0, ptr);
1746 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1747 if (!cpu_has_64bits) {
1748 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1749 uasm_i_ori(p, pte, pte, hwmode);
1750 BUG_ON(hwmode & ~0xffff);
1751 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1752 uasm_i_lw(p, pte, 0, ptr);
1759 * Check if PTE is present, if not then jump to LABEL. PTR points to
1760 * the page table where this PTE is located, PTE will be re-loaded
1761 * with it's original value.
1764 build_pte_present(u32 **p, struct uasm_reloc **r,
1765 int pte, int ptr, int scratch, enum label_id lid)
1767 int t = scratch >= 0 ? scratch : pte;
1771 if (use_bbit_insns()) {
1772 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1775 if (_PAGE_PRESENT_SHIFT) {
1776 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1779 uasm_i_andi(p, t, cur, 1);
1780 uasm_il_beqz(p, r, t, lid);
1782 /* You lose the SMP race :-(*/
1783 iPTE_LW(p, pte, ptr);
1786 if (_PAGE_PRESENT_SHIFT) {
1787 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1790 uasm_i_andi(p, t, cur,
1791 (_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1792 uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1793 uasm_il_bnez(p, r, t, lid);
1795 /* You lose the SMP race :-(*/
1796 iPTE_LW(p, pte, ptr);
1800 /* Make PTE valid, store result in PTR. */
1802 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1803 unsigned int ptr, unsigned int scratch)
1805 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1807 iPTE_SW(p, r, pte, ptr, mode, scratch);
1811 * Check if PTE can be written to, if not branch to LABEL. Regardless
1812 * restore PTE with value from PTR when done.
1815 build_pte_writable(u32 **p, struct uasm_reloc **r,
1816 unsigned int pte, unsigned int ptr, int scratch,
1819 int t = scratch >= 0 ? scratch : pte;
1822 if (_PAGE_PRESENT_SHIFT) {
1823 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1826 uasm_i_andi(p, t, cur,
1827 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1828 uasm_i_xori(p, t, t,
1829 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1830 uasm_il_bnez(p, r, t, lid);
1832 /* You lose the SMP race :-(*/
1833 iPTE_LW(p, pte, ptr);
1838 /* Make PTE writable, update software status bits as well, then store
1842 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1843 unsigned int ptr, unsigned int scratch)
1845 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1848 iPTE_SW(p, r, pte, ptr, mode, scratch);
1852 * Check if PTE can be modified, if not branch to LABEL. Regardless
1853 * restore PTE with value from PTR when done.
1856 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1857 unsigned int pte, unsigned int ptr, int scratch,
1860 if (use_bbit_insns()) {
1861 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1864 int t = scratch >= 0 ? scratch : pte;
1865 uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1866 uasm_i_andi(p, t, t, 1);
1867 uasm_il_beqz(p, r, t, lid);
1869 /* You lose the SMP race :-(*/
1870 iPTE_LW(p, pte, ptr);
1874 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1878 * R3000 style TLB load/store/modify handlers.
1882 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1886 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1888 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1889 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1892 uasm_i_rfe(p); /* branch delay */
1896 * This places the pte into ENTRYLO0 and writes it with tlbwi
1897 * or tlbwr as appropriate. This is because the index register
1898 * may have the probe fail bit set as a result of a trap on a
1899 * kseg2 access, i.e. without refill. Then it returns.
1902 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1903 struct uasm_reloc **r, unsigned int pte,
1906 uasm_i_mfc0(p, tmp, C0_INDEX);
1907 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1908 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1909 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1910 uasm_i_tlbwi(p); /* cp0 delay */
1912 uasm_i_rfe(p); /* branch delay */
1913 uasm_l_r3000_write_probe_fail(l, *p);
1914 uasm_i_tlbwr(p); /* cp0 delay */
1916 uasm_i_rfe(p); /* branch delay */
1920 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1923 long pgdc = (long)pgd_current;
1925 uasm_i_mfc0(p, pte, C0_BADVADDR);
1926 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1927 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1928 uasm_i_srl(p, pte, pte, 22); /* load delay */
1929 uasm_i_sll(p, pte, pte, 2);
1930 uasm_i_addu(p, ptr, ptr, pte);
1931 uasm_i_mfc0(p, pte, C0_CONTEXT);
1932 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1933 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1934 uasm_i_addu(p, ptr, ptr, pte);
1935 uasm_i_lw(p, pte, 0, ptr);
1936 uasm_i_tlbp(p); /* load delay */
1939 static void build_r3000_tlb_load_handler(void)
1941 u32 *p = (u32 *)handle_tlbl;
1942 struct uasm_label *l = labels;
1943 struct uasm_reloc *r = relocs;
1945 memset(p, 0, handle_tlbl_end - (char *)p);
1946 memset(labels, 0, sizeof(labels));
1947 memset(relocs, 0, sizeof(relocs));
1949 build_r3000_tlbchange_handler_head(&p, K0, K1);
1950 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1951 uasm_i_nop(&p); /* load delay */
1952 build_make_valid(&p, &r, K0, K1, -1);
1953 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1955 uasm_l_nopage_tlbl(&l, p);
1956 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1959 if (p >= (u32 *)handle_tlbl_end)
1960 panic("TLB load handler fastpath space exceeded");
1962 uasm_resolve_relocs(relocs, labels);
1963 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1964 (unsigned int)(p - (u32 *)handle_tlbl));
1966 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1969 static void build_r3000_tlb_store_handler(void)
1971 u32 *p = (u32 *)handle_tlbs;
1972 struct uasm_label *l = labels;
1973 struct uasm_reloc *r = relocs;
1975 memset(p, 0, handle_tlbs_end - (char *)p);
1976 memset(labels, 0, sizeof(labels));
1977 memset(relocs, 0, sizeof(relocs));
1979 build_r3000_tlbchange_handler_head(&p, K0, K1);
1980 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1981 uasm_i_nop(&p); /* load delay */
1982 build_make_write(&p, &r, K0, K1, -1);
1983 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1985 uasm_l_nopage_tlbs(&l, p);
1986 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1989 if (p >= (u32 *)handle_tlbs_end)
1990 panic("TLB store handler fastpath space exceeded");
1992 uasm_resolve_relocs(relocs, labels);
1993 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1994 (unsigned int)(p - (u32 *)handle_tlbs));
1996 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1999 static void build_r3000_tlb_modify_handler(void)
2001 u32 *p = (u32 *)handle_tlbm;
2002 struct uasm_label *l = labels;
2003 struct uasm_reloc *r = relocs;
2005 memset(p, 0, handle_tlbm_end - (char *)p);
2006 memset(labels, 0, sizeof(labels));
2007 memset(relocs, 0, sizeof(relocs));
2009 build_r3000_tlbchange_handler_head(&p, K0, K1);
2010 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
2011 uasm_i_nop(&p); /* load delay */
2012 build_make_write(&p, &r, K0, K1, -1);
2013 build_r3000_pte_reload_tlbwi(&p, K0, K1);
2015 uasm_l_nopage_tlbm(&l, p);
2016 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2019 if (p >= (u32 *)handle_tlbm_end)
2020 panic("TLB modify handler fastpath space exceeded");
2022 uasm_resolve_relocs(relocs, labels);
2023 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2024 (unsigned int)(p - (u32 *)handle_tlbm));
2026 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
2028 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2030 static bool cpu_has_tlbex_tlbp_race(void)
2033 * When a Hardware Table Walker is running it can replace TLB entries
2034 * at any time, leading to a race between it & the CPU.
2040 * If the CPU shares FTLB RAM with its siblings then our entry may be
2041 * replaced at any time by a sibling performing a write to the FTLB.
2043 if (cpu_has_shared_ftlb_ram)
2046 /* In all other cases there ought to be no race condition to handle */
2051 * R4000 style TLB load/store/modify handlers.
2053 static struct work_registers
2054 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2055 struct uasm_reloc **r)
2057 struct work_registers wr = build_get_work_registers(p);
2060 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2062 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2065 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2067 * For huge tlb entries, pmd doesn't contain an address but
2068 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2069 * see if we need to jump to huge tlb processing.
2071 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2074 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2075 UASM_i_LW(p, wr.r2, 0, wr.r2);
2076 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2077 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2078 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2081 uasm_l_smp_pgtable_change(l, *p);
2083 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2084 if (!m4kc_tlbp_war()) {
2085 build_tlb_probe_entry(p);
2086 if (cpu_has_tlbex_tlbp_race()) {
2087 /* race condition happens, leaving */
2089 uasm_i_mfc0(p, wr.r3, C0_INDEX);
2090 uasm_il_bltz(p, r, wr.r3, label_leave);
2098 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2099 struct uasm_reloc **r, unsigned int tmp,
2102 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2103 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2104 build_update_entries(p, tmp, ptr);
2105 build_tlb_write_entry(p, l, r, tlb_indexed);
2106 uasm_l_leave(l, *p);
2107 build_restore_work_registers(p);
2108 uasm_i_eret(p); /* return from trap */
2111 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2115 static void build_r4000_tlb_load_handler(void)
2117 u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2118 struct uasm_label *l = labels;
2119 struct uasm_reloc *r = relocs;
2120 struct work_registers wr;
2122 memset(p, 0, handle_tlbl_end - (char *)p);
2123 memset(labels, 0, sizeof(labels));
2124 memset(relocs, 0, sizeof(relocs));
2126 if (bcm1250_m3_war()) {
2127 unsigned int segbits = 44;
2129 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2130 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2131 uasm_i_xor(&p, K0, K0, K1);
2132 uasm_i_dsrl_safe(&p, K1, K0, 62);
2133 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2134 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2135 uasm_i_or(&p, K0, K0, K1);
2136 uasm_il_bnez(&p, &r, K0, label_leave);
2137 /* No need for uasm_i_nop */
2140 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2141 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2142 if (m4kc_tlbp_war())
2143 build_tlb_probe_entry(&p);
2145 if (cpu_has_rixi && !cpu_has_rixiex) {
2147 * If the page is not _PAGE_VALID, RI or XI could not
2148 * have triggered it. Skip the expensive test..
2150 if (use_bbit_insns()) {
2151 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2152 label_tlbl_goaround1);
2154 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2155 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2160 * Warn if something may race with us & replace the TLB entry
2161 * before we read it here. Everything with such races should
2162 * also have dedicated RiXi exception handlers, so this
2165 WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2169 switch (current_cpu_type()) {
2171 if (cpu_has_mips_r2_exec_hazard) {
2175 case CPU_CAVIUM_OCTEON:
2176 case CPU_CAVIUM_OCTEON_PLUS:
2177 case CPU_CAVIUM_OCTEON2:
2182 /* Examine entrylo 0 or 1 based on ptr. */
2183 if (use_bbit_insns()) {
2184 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2186 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2187 uasm_i_beqz(&p, wr.r3, 8);
2189 /* load it in the delay slot*/
2190 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2191 /* load it if ptr is odd */
2192 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2194 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2195 * XI must have triggered it.
2197 if (use_bbit_insns()) {
2198 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2200 uasm_l_tlbl_goaround1(&l, p);
2202 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2203 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2206 uasm_l_tlbl_goaround1(&l, p);
2208 build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2209 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2211 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2213 * This is the entry point when build_r4000_tlbchange_handler_head
2214 * spots a huge page.
2216 uasm_l_tlb_huge_update(&l, p);
2217 iPTE_LW(&p, wr.r1, wr.r2);
2218 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2219 build_tlb_probe_entry(&p);
2221 if (cpu_has_rixi && !cpu_has_rixiex) {
2223 * If the page is not _PAGE_VALID, RI or XI could not
2224 * have triggered it. Skip the expensive test..
2226 if (use_bbit_insns()) {
2227 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2228 label_tlbl_goaround2);
2230 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2231 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2236 * Warn if something may race with us & replace the TLB entry
2237 * before we read it here. Everything with such races should
2238 * also have dedicated RiXi exception handlers, so this
2241 WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2245 switch (current_cpu_type()) {
2247 if (cpu_has_mips_r2_exec_hazard) {
2250 case CPU_CAVIUM_OCTEON:
2251 case CPU_CAVIUM_OCTEON_PLUS:
2252 case CPU_CAVIUM_OCTEON2:
2257 /* Examine entrylo 0 or 1 based on ptr. */
2258 if (use_bbit_insns()) {
2259 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2261 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2262 uasm_i_beqz(&p, wr.r3, 8);
2264 /* load it in the delay slot*/
2265 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2266 /* load it if ptr is odd */
2267 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2269 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2270 * XI must have triggered it.
2272 if (use_bbit_insns()) {
2273 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2275 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2276 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2278 if (PM_DEFAULT_MASK == 0)
2281 * We clobbered C0_PAGEMASK, restore it. On the other branch
2282 * it is restored in build_huge_tlb_write_entry.
2284 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2286 uasm_l_tlbl_goaround2(&l, p);
2288 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2289 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2292 uasm_l_nopage_tlbl(&l, p);
2293 if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2295 build_restore_work_registers(&p);
2296 #ifdef CONFIG_CPU_MICROMIPS
2297 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2298 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2299 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2303 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2306 if (p >= (u32 *)handle_tlbl_end)
2307 panic("TLB load handler fastpath space exceeded");
2309 uasm_resolve_relocs(relocs, labels);
2310 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2311 (unsigned int)(p - (u32 *)handle_tlbl));
2313 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2316 static void build_r4000_tlb_store_handler(void)
2318 u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2319 struct uasm_label *l = labels;
2320 struct uasm_reloc *r = relocs;
2321 struct work_registers wr;
2323 memset(p, 0, handle_tlbs_end - (char *)p);
2324 memset(labels, 0, sizeof(labels));
2325 memset(relocs, 0, sizeof(relocs));
2327 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2328 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2329 if (m4kc_tlbp_war())
2330 build_tlb_probe_entry(&p);
2331 build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2332 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2334 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2336 * This is the entry point when
2337 * build_r4000_tlbchange_handler_head spots a huge page.
2339 uasm_l_tlb_huge_update(&l, p);
2340 iPTE_LW(&p, wr.r1, wr.r2);
2341 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2342 build_tlb_probe_entry(&p);
2343 uasm_i_ori(&p, wr.r1, wr.r1,
2344 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2345 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2348 uasm_l_nopage_tlbs(&l, p);
2349 if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2351 build_restore_work_registers(&p);
2352 #ifdef CONFIG_CPU_MICROMIPS
2353 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2354 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2355 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2359 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2362 if (p >= (u32 *)handle_tlbs_end)
2363 panic("TLB store handler fastpath space exceeded");
2365 uasm_resolve_relocs(relocs, labels);
2366 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2367 (unsigned int)(p - (u32 *)handle_tlbs));
2369 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2372 static void build_r4000_tlb_modify_handler(void)
2374 u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2375 struct uasm_label *l = labels;
2376 struct uasm_reloc *r = relocs;
2377 struct work_registers wr;
2379 memset(p, 0, handle_tlbm_end - (char *)p);
2380 memset(labels, 0, sizeof(labels));
2381 memset(relocs, 0, sizeof(relocs));
2383 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2384 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2385 if (m4kc_tlbp_war())
2386 build_tlb_probe_entry(&p);
2387 /* Present and writable bits set, set accessed and dirty bits. */
2388 build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2389 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2391 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2393 * This is the entry point when
2394 * build_r4000_tlbchange_handler_head spots a huge page.
2396 uasm_l_tlb_huge_update(&l, p);
2397 iPTE_LW(&p, wr.r1, wr.r2);
2398 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2399 build_tlb_probe_entry(&p);
2400 uasm_i_ori(&p, wr.r1, wr.r1,
2401 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2402 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2405 uasm_l_nopage_tlbm(&l, p);
2406 if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2408 build_restore_work_registers(&p);
2409 #ifdef CONFIG_CPU_MICROMIPS
2410 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2411 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2412 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2416 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2419 if (p >= (u32 *)handle_tlbm_end)
2420 panic("TLB modify handler fastpath space exceeded");
2422 uasm_resolve_relocs(relocs, labels);
2423 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2424 (unsigned int)(p - (u32 *)handle_tlbm));
2426 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2429 static void flush_tlb_handlers(void)
2431 local_flush_icache_range((unsigned long)handle_tlbl,
2432 (unsigned long)handle_tlbl_end);
2433 local_flush_icache_range((unsigned long)handle_tlbs,
2434 (unsigned long)handle_tlbs_end);
2435 local_flush_icache_range((unsigned long)handle_tlbm,
2436 (unsigned long)handle_tlbm_end);
2437 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2438 (unsigned long)tlbmiss_handler_setup_pgd_end);
2441 static void print_htw_config(void)
2443 unsigned long config;
2445 const int field = 2 * sizeof(unsigned long);
2447 config = read_c0_pwfield();
2448 pr_debug("PWField (0x%0*lx): GDI: 0x%02lx UDI: 0x%02lx MDI: 0x%02lx PTI: 0x%02lx PTEI: 0x%02lx\n",
2450 (config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2451 (config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2452 (config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2453 (config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2454 (config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2456 config = read_c0_pwsize();
2457 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",
2459 (config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2460 (config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2461 (config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2462 (config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2463 (config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2464 (config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2466 pwctl = read_c0_pwctl();
2467 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",
2469 (pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2470 (pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2471 (pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2472 (pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2473 (pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2474 (pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2475 (pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2478 static void config_htw_params(void)
2480 unsigned long pwfield, pwsize, ptei;
2481 unsigned int config;
2484 * We are using 2-level page tables, so we only need to
2485 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2486 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2487 * write values less than 0xc in these fields because the entire
2488 * write will be dropped. As a result of which, we must preserve
2489 * the original reset values and overwrite only what we really want.
2492 pwfield = read_c0_pwfield();
2493 /* re-initialize the GDI field */
2494 pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2495 pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2496 /* re-initialize the PTI field including the even/odd bit */
2497 pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2498 pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2499 if (CONFIG_PGTABLE_LEVELS >= 3) {
2500 pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2501 pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2503 /* Set the PTEI right shift */
2504 ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2506 write_c0_pwfield(pwfield);
2507 /* Check whether the PTEI value is supported */
2508 back_to_back_c0_hazard();
2509 pwfield = read_c0_pwfield();
2510 if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2512 pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2515 * Drop option to avoid HTW being enabled via another path
2518 current_cpu_data.options &= ~MIPS_CPU_HTW;
2522 pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2523 pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2524 if (CONFIG_PGTABLE_LEVELS >= 3)
2525 pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2527 /* Set pointer size to size of directory pointers */
2528 if (IS_ENABLED(CONFIG_64BIT))
2529 pwsize |= MIPS_PWSIZE_PS_MASK;
2530 /* PTEs may be multiple pointers long (e.g. with XPA) */
2531 pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2532 & MIPS_PWSIZE_PTEW_MASK;
2534 write_c0_pwsize(pwsize);
2536 /* Make sure everything is set before we enable the HTW */
2537 back_to_back_c0_hazard();
2540 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2543 config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2544 if (IS_ENABLED(CONFIG_64BIT))
2545 config |= MIPS_PWCTL_XU_MASK;
2546 write_c0_pwctl(config);
2547 pr_info("Hardware Page Table Walker enabled\n");
2552 static void config_xpa_params(void)
2555 unsigned int pagegrain;
2557 if (mips_xpa_disabled) {
2558 pr_info("Extended Physical Addressing (XPA) disabled\n");
2562 pagegrain = read_c0_pagegrain();
2563 write_c0_pagegrain(pagegrain | PG_ELPA);
2564 back_to_back_c0_hazard();
2565 pagegrain = read_c0_pagegrain();
2567 if (pagegrain & PG_ELPA)
2568 pr_info("Extended Physical Addressing (XPA) enabled\n");
2570 panic("Extended Physical Addressing (XPA) disabled");
2574 static void check_pabits(void)
2576 unsigned long entry;
2577 unsigned pabits, fillbits;
2579 if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2581 * We'll only be making use of the fact that we can rotate bits
2582 * into the fill if the CPU supports RIXI, so don't bother
2583 * probing this for CPUs which don't.
2588 write_c0_entrylo0(~0ul);
2589 back_to_back_c0_hazard();
2590 entry = read_c0_entrylo0();
2592 /* clear all non-PFN bits */
2593 entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2594 entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2596 /* find a lower bound on PABITS, and upper bound on fill bits */
2597 pabits = fls_long(entry) + 6;
2598 fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2600 /* minus the RI & XI bits */
2601 fillbits -= min_t(unsigned, fillbits, 2);
2603 if (fillbits >= ilog2(_PAGE_NO_EXEC))
2604 fill_includes_sw_bits = true;
2606 pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2609 void build_tlb_refill_handler(void)
2612 * The refill handler is generated per-CPU, multi-node systems
2613 * may have local storage for it. The other handlers are only
2616 static int run_once = 0;
2618 if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2619 panic("Kernels supporting XPA currently require CPUs with RIXI");
2621 output_pgtable_bits_defines();
2625 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2629 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2632 build_r3000_tlb_refill_handler();
2633 build_r3000_tlb_load_handler();
2634 build_r3000_tlb_store_handler();
2635 build_r3000_tlb_modify_handler();
2636 flush_tlb_handlers();
2640 panic("No R3000 TLB refill handler");
2649 scratch_reg = allocate_kscratch();
2651 build_r4000_tlb_load_handler();
2652 build_r4000_tlb_store_handler();
2653 build_r4000_tlb_modify_handler();
2655 build_loongson3_tlb_refill_handler();
2657 build_r4000_tlb_refill_handler();
2658 flush_tlb_handlers();
2662 config_xpa_params();
2664 config_htw_params();