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/kernel.h>
26 #include <linux/types.h>
27 #include <linux/smp.h>
28 #include <linux/string.h>
29 #include <linux/cache.h>
31 #include <asm/cacheflush.h>
32 #include <asm/cpu-type.h>
33 #include <asm/pgtable.h>
36 #include <asm/setup.h>
38 static int mips_xpa_disabled;
40 static int __init xpa_disable(char *s)
42 mips_xpa_disabled = 1;
47 __setup("noxpa", xpa_disable);
50 * TLB load/store/modify handlers.
52 * Only the fastpath gets synthesized at runtime, the slowpath for
53 * do_page_fault remains normal asm.
55 extern void tlb_do_page_fault_0(void);
56 extern void tlb_do_page_fault_1(void);
58 struct work_registers {
67 } ____cacheline_aligned_in_smp;
69 static struct tlb_reg_save handler_reg_save[NR_CPUS];
71 static inline int r45k_bvahwbug(void)
73 /* XXX: We should probe for the presence of this bug, but we don't. */
77 static inline int r4k_250MHZhwbug(void)
79 /* XXX: We should probe for the presence of this bug, but we don't. */
83 static inline int __maybe_unused bcm1250_m3_war(void)
85 return BCM1250_M3_WAR;
88 static inline int __maybe_unused r10000_llsc_war(void)
90 return R10000_LLSC_WAR;
93 static int use_bbit_insns(void)
95 switch (current_cpu_type()) {
96 case CPU_CAVIUM_OCTEON:
97 case CPU_CAVIUM_OCTEON_PLUS:
98 case CPU_CAVIUM_OCTEON2:
99 case CPU_CAVIUM_OCTEON3:
106 static int use_lwx_insns(void)
108 switch (current_cpu_type()) {
109 case CPU_CAVIUM_OCTEON2:
110 case CPU_CAVIUM_OCTEON3:
116 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
117 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
118 static bool scratchpad_available(void)
122 static int scratchpad_offset(int i)
125 * CVMSEG starts at address -32768 and extends for
126 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
128 i += 1; /* Kernel use starts at the top and works down. */
129 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
132 static bool scratchpad_available(void)
136 static int scratchpad_offset(int i)
139 /* Really unreachable, but evidently some GCC want this. */
144 * Found by experiment: At least some revisions of the 4kc throw under
145 * some circumstances a machine check exception, triggered by invalid
146 * values in the index register. Delaying the tlbp instruction until
147 * after the next branch, plus adding an additional nop in front of
148 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
149 * why; it's not an issue caused by the core RTL.
152 static int m4kc_tlbp_war(void)
154 return (current_cpu_data.processor_id & 0xffff00) ==
155 (PRID_COMP_MIPS | PRID_IMP_4KC);
158 /* Handle labels (which must be positive integers). */
160 label_second_part = 1,
165 label_split = label_tlbw_hazard_0 + 8,
166 label_tlbl_goaround1,
167 label_tlbl_goaround2,
171 label_smp_pgtable_change,
172 label_r3000_write_probe_fail,
173 label_large_segbits_fault,
174 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
175 label_tlb_huge_update,
179 UASM_L_LA(_second_part)
182 UASM_L_LA(_vmalloc_done)
183 /* _tlbw_hazard_x is handled differently. */
185 UASM_L_LA(_tlbl_goaround1)
186 UASM_L_LA(_tlbl_goaround2)
187 UASM_L_LA(_nopage_tlbl)
188 UASM_L_LA(_nopage_tlbs)
189 UASM_L_LA(_nopage_tlbm)
190 UASM_L_LA(_smp_pgtable_change)
191 UASM_L_LA(_r3000_write_probe_fail)
192 UASM_L_LA(_large_segbits_fault)
193 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
194 UASM_L_LA(_tlb_huge_update)
197 static int hazard_instance;
199 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
203 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
210 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
214 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
222 * pgtable bits are assigned dynamically depending on processor feature
223 * and statically based on kernel configuration. This spits out the actual
224 * values the kernel is using. Required to make sense from disassembled
225 * TLB exception handlers.
227 static void output_pgtable_bits_defines(void)
229 #define pr_define(fmt, ...) \
230 pr_debug("#define " fmt, ##__VA_ARGS__)
232 pr_debug("#include <asm/asm.h>\n");
233 pr_debug("#include <asm/regdef.h>\n");
236 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
237 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
238 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
239 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
240 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
241 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
242 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
244 #ifdef _PAGE_NO_EXEC_SHIFT
246 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
248 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
249 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
250 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
251 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
255 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
259 pr_debug("LEAF(%s)\n", symbol);
261 pr_debug("\t.set push\n");
262 pr_debug("\t.set noreorder\n");
264 for (i = 0; i < count; i++)
265 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
267 pr_debug("\t.set\tpop\n");
269 pr_debug("\tEND(%s)\n", symbol);
272 /* The only general purpose registers allowed in TLB handlers. */
276 /* Some CP0 registers */
277 #define C0_INDEX 0, 0
278 #define C0_ENTRYLO0 2, 0
279 #define C0_TCBIND 2, 2
280 #define C0_ENTRYLO1 3, 0
281 #define C0_CONTEXT 4, 0
282 #define C0_PAGEMASK 5, 0
283 #define C0_PWBASE 5, 5
284 #define C0_PWFIELD 5, 6
285 #define C0_PWSIZE 5, 7
286 #define C0_PWCTL 6, 6
287 #define C0_BADVADDR 8, 0
289 #define C0_ENTRYHI 10, 0
291 #define C0_XCONTEXT 20, 0
294 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
296 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
299 /* The worst case length of the handler is around 18 instructions for
300 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
301 * Maximum space available is 32 instructions for R3000 and 64
302 * instructions for R4000.
304 * We deliberately chose a buffer size of 128, so we won't scribble
305 * over anything important on overflow before we panic.
307 static u32 tlb_handler[128];
309 /* simply assume worst case size for labels and relocs */
310 static struct uasm_label labels[128];
311 static struct uasm_reloc relocs[128];
313 static int check_for_high_segbits;
314 static bool fill_includes_sw_bits;
316 static unsigned int kscratch_used_mask;
318 static inline int __maybe_unused c0_kscratch(void)
320 switch (current_cpu_type()) {
329 static int allocate_kscratch(void)
332 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
339 r--; /* make it zero based */
341 kscratch_used_mask |= (1 << r);
346 static int scratch_reg;
348 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
350 static struct work_registers build_get_work_registers(u32 **p)
352 struct work_registers r;
354 if (scratch_reg >= 0) {
355 /* Save in CPU local C0_KScratch? */
356 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
363 if (num_possible_cpus() > 1) {
364 /* Get smp_processor_id */
365 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
366 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
368 /* handler_reg_save index in K0 */
369 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
371 UASM_i_LA(p, K1, (long)&handler_reg_save);
372 UASM_i_ADDU(p, K0, K0, K1);
374 UASM_i_LA(p, K0, (long)&handler_reg_save);
376 /* K0 now points to save area, save $1 and $2 */
377 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
378 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
386 static void build_restore_work_registers(u32 **p)
388 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) {
680 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
682 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
685 /* Reset default page size */
686 if (PM_DEFAULT_MASK >> 16) {
687 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
688 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
689 uasm_il_b(p, r, lid);
690 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
691 } else if (PM_DEFAULT_MASK) {
692 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
693 uasm_il_b(p, r, lid);
694 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
696 uasm_il_b(p, r, lid);
697 uasm_i_mtc0(p, 0, C0_PAGEMASK);
702 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
703 struct uasm_reloc **r,
705 enum tlb_write_entry wmode,
708 /* Set huge page tlb entry size */
709 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
710 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
711 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
713 build_tlb_write_entry(p, l, r, wmode);
715 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
719 * Check if Huge PTE is present, if so then jump to LABEL.
722 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
723 unsigned int pmd, int lid)
725 UASM_i_LW(p, tmp, 0, pmd);
726 if (use_bbit_insns()) {
727 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
729 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
730 uasm_il_bnez(p, r, tmp, lid);
734 static void build_huge_update_entries(u32 **p, unsigned int pte,
740 * A huge PTE describes an area the size of the
741 * configured huge page size. This is twice the
742 * of the large TLB entry size we intend to use.
743 * A TLB entry half the size of the configured
744 * huge page size is configured into entrylo0
745 * and entrylo1 to cover the contiguous huge PTE
748 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
750 /* We can clobber tmp. It isn't used after this.*/
752 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
754 build_convert_pte_to_entrylo(p, pte);
755 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
756 /* convert to entrylo1 */
758 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
760 UASM_i_ADDU(p, pte, pte, tmp);
762 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
765 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
766 struct uasm_label **l,
772 UASM_i_SC(p, pte, 0, ptr);
773 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
774 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
776 UASM_i_SW(p, pte, 0, ptr);
778 if (cpu_has_ftlb && flush) {
779 BUG_ON(!cpu_has_tlbinv);
781 UASM_i_MFC0(p, ptr, C0_ENTRYHI);
782 uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
783 UASM_i_MTC0(p, ptr, C0_ENTRYHI);
784 build_tlb_write_entry(p, l, r, tlb_indexed);
786 uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
787 UASM_i_MTC0(p, ptr, C0_ENTRYHI);
788 build_huge_update_entries(p, pte, ptr);
789 build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
794 build_huge_update_entries(p, pte, ptr);
795 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
797 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
801 * TMP and PTR are scratch.
802 * TMP will be clobbered, PTR will hold the pmd entry.
805 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
806 unsigned int tmp, unsigned int ptr)
808 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
809 long pgdc = (long)pgd_current;
812 * The vmalloc handling is not in the hotpath.
814 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
816 if (check_for_high_segbits) {
818 * The kernel currently implicitely assumes that the
819 * MIPS SEGBITS parameter for the processor is
820 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
821 * allocate virtual addresses outside the maximum
822 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
823 * that doesn't prevent user code from accessing the
824 * higher xuseg addresses. Here, we make sure that
825 * everything but the lower xuseg addresses goes down
826 * the module_alloc/vmalloc path.
828 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
829 uasm_il_bnez(p, r, ptr, label_vmalloc);
831 uasm_il_bltz(p, r, tmp, label_vmalloc);
833 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
836 /* pgd is in pgd_reg */
838 UASM_i_MFC0(p, ptr, C0_PWBASE);
840 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
842 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
844 * &pgd << 11 stored in CONTEXT [23..63].
846 UASM_i_MFC0(p, ptr, C0_CONTEXT);
848 /* Clear lower 23 bits of context. */
849 uasm_i_dins(p, ptr, 0, 0, 23);
851 /* 1 0 1 0 1 << 6 xkphys cached */
852 uasm_i_ori(p, ptr, ptr, 0x540);
853 uasm_i_drotr(p, ptr, ptr, 11);
854 #elif defined(CONFIG_SMP)
855 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
856 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
857 UASM_i_LA_mostly(p, tmp, pgdc);
858 uasm_i_daddu(p, ptr, ptr, tmp);
859 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
860 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
862 UASM_i_LA_mostly(p, ptr, pgdc);
863 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
867 uasm_l_vmalloc_done(l, *p);
869 /* get pgd offset in bytes */
870 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
872 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
873 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
874 #ifndef __PAGETABLE_PMD_FOLDED
875 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
876 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
877 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
878 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
879 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
884 * BVADDR is the faulting address, PTR is scratch.
885 * PTR will hold the pgd for vmalloc.
888 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
889 unsigned int bvaddr, unsigned int ptr,
890 enum vmalloc64_mode mode)
892 long swpd = (long)swapper_pg_dir;
893 int single_insn_swpd;
894 int did_vmalloc_branch = 0;
896 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
898 uasm_l_vmalloc(l, *p);
900 if (mode != not_refill && check_for_high_segbits) {
901 if (single_insn_swpd) {
902 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
903 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
904 did_vmalloc_branch = 1;
907 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
910 if (!did_vmalloc_branch) {
911 if (single_insn_swpd) {
912 uasm_il_b(p, r, label_vmalloc_done);
913 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
915 UASM_i_LA_mostly(p, ptr, swpd);
916 uasm_il_b(p, r, label_vmalloc_done);
917 if (uasm_in_compat_space_p(swpd))
918 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
920 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
923 if (mode != not_refill && check_for_high_segbits) {
924 uasm_l_large_segbits_fault(l, *p);
926 * We get here if we are an xsseg address, or if we are
927 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
929 * Ignoring xsseg (assume disabled so would generate
930 * (address errors?), the only remaining possibility
931 * is the upper xuseg addresses. On processors with
932 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
933 * addresses would have taken an address error. We try
934 * to mimic that here by taking a load/istream page
937 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
940 if (mode == refill_scratch) {
941 if (scratch_reg >= 0) {
943 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
945 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
953 #else /* !CONFIG_64BIT */
956 * TMP and PTR are scratch.
957 * TMP will be clobbered, PTR will hold the pgd entry.
959 static void __maybe_unused
960 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
963 /* pgd is in pgd_reg */
964 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
965 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
967 long pgdc = (long)pgd_current;
969 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
971 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
972 UASM_i_LA_mostly(p, tmp, pgdc);
973 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
974 uasm_i_addu(p, ptr, tmp, ptr);
976 UASM_i_LA_mostly(p, ptr, pgdc);
978 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
979 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
981 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
982 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
983 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
986 #endif /* !CONFIG_64BIT */
988 static void build_adjust_context(u32 **p, unsigned int ctx)
990 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
991 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
993 switch (current_cpu_type()) {
1010 UASM_i_SRL(p, ctx, ctx, shift);
1011 uasm_i_andi(p, ctx, ctx, mask);
1014 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1017 * Bug workaround for the Nevada. It seems as if under certain
1018 * circumstances the move from cp0_context might produce a
1019 * bogus result when the mfc0 instruction and its consumer are
1020 * in a different cacheline or a load instruction, probably any
1021 * memory reference, is between them.
1023 switch (current_cpu_type()) {
1025 UASM_i_LW(p, ptr, 0, ptr);
1026 GET_CONTEXT(p, tmp); /* get context reg */
1030 GET_CONTEXT(p, tmp); /* get context reg */
1031 UASM_i_LW(p, ptr, 0, ptr);
1035 build_adjust_context(p, tmp);
1036 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1039 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1041 int pte_off_even = 0;
1042 int pte_off_odd = sizeof(pte_t);
1044 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1045 /* The low 32 bits of EntryLo is stored in pte_high */
1046 pte_off_even += offsetof(pte_t, pte_high);
1047 pte_off_odd += offsetof(pte_t, pte_high);
1050 if (IS_ENABLED(CONFIG_XPA)) {
1051 uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1052 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1053 UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1055 if (cpu_has_xpa && !mips_xpa_disabled) {
1056 uasm_i_lw(p, tmp, 0, ptep);
1057 uasm_i_ext(p, tmp, tmp, 0, 24);
1058 uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1061 uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1062 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1063 UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1065 if (cpu_has_xpa && !mips_xpa_disabled) {
1066 uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1067 uasm_i_ext(p, tmp, tmp, 0, 24);
1068 uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1073 UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1074 UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1075 if (r45k_bvahwbug())
1076 build_tlb_probe_entry(p);
1077 build_convert_pte_to_entrylo(p, tmp);
1078 if (r4k_250MHZhwbug())
1079 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1080 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1081 build_convert_pte_to_entrylo(p, ptep);
1082 if (r45k_bvahwbug())
1083 uasm_i_mfc0(p, tmp, C0_INDEX);
1084 if (r4k_250MHZhwbug())
1085 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1086 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1089 struct mips_huge_tlb_info {
1091 int restore_scratch;
1092 bool need_reload_pte;
1095 static struct mips_huge_tlb_info
1096 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1097 struct uasm_reloc **r, unsigned int tmp,
1098 unsigned int ptr, int c0_scratch_reg)
1100 struct mips_huge_tlb_info rv;
1101 unsigned int even, odd;
1102 int vmalloc_branch_delay_filled = 0;
1103 const int scratch = 1; /* Our extra working register */
1105 rv.huge_pte = scratch;
1106 rv.restore_scratch = 0;
1107 rv.need_reload_pte = false;
1109 if (check_for_high_segbits) {
1110 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1113 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1115 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1117 if (c0_scratch_reg >= 0)
1118 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1120 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1122 uasm_i_dsrl_safe(p, scratch, tmp,
1123 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1124 uasm_il_bnez(p, r, scratch, label_vmalloc);
1126 if (pgd_reg == -1) {
1127 vmalloc_branch_delay_filled = 1;
1128 /* Clear lower 23 bits of context. */
1129 uasm_i_dins(p, ptr, 0, 0, 23);
1133 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1135 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1137 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1139 if (c0_scratch_reg >= 0)
1140 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1142 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1145 /* Clear lower 23 bits of context. */
1146 uasm_i_dins(p, ptr, 0, 0, 23);
1148 uasm_il_bltz(p, r, tmp, label_vmalloc);
1151 if (pgd_reg == -1) {
1152 vmalloc_branch_delay_filled = 1;
1153 /* 1 0 1 0 1 << 6 xkphys cached */
1154 uasm_i_ori(p, ptr, ptr, 0x540);
1155 uasm_i_drotr(p, ptr, ptr, 11);
1158 #ifdef __PAGETABLE_PMD_FOLDED
1159 #define LOC_PTEP scratch
1161 #define LOC_PTEP ptr
1164 if (!vmalloc_branch_delay_filled)
1165 /* get pgd offset in bytes */
1166 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1168 uasm_l_vmalloc_done(l, *p);
1172 * fall-through case = badvaddr *pgd_current
1173 * vmalloc case = badvaddr swapper_pg_dir
1176 if (vmalloc_branch_delay_filled)
1177 /* get pgd offset in bytes */
1178 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1180 #ifdef __PAGETABLE_PMD_FOLDED
1181 GET_CONTEXT(p, tmp); /* get context reg */
1183 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1185 if (use_lwx_insns()) {
1186 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1188 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1189 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1192 #ifndef __PAGETABLE_PMD_FOLDED
1193 /* get pmd offset in bytes */
1194 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1195 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1196 GET_CONTEXT(p, tmp); /* get context reg */
1198 if (use_lwx_insns()) {
1199 UASM_i_LWX(p, scratch, scratch, ptr);
1201 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1202 UASM_i_LW(p, scratch, 0, ptr);
1205 /* Adjust the context during the load latency. */
1206 build_adjust_context(p, tmp);
1208 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1209 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1211 * The in the LWX case we don't want to do the load in the
1212 * delay slot. It cannot issue in the same cycle and may be
1213 * speculative and unneeded.
1215 if (use_lwx_insns())
1217 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1220 /* build_update_entries */
1221 if (use_lwx_insns()) {
1224 UASM_i_LWX(p, even, scratch, tmp);
1225 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1226 UASM_i_LWX(p, odd, scratch, tmp);
1228 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1231 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1232 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1235 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1236 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1237 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1239 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1240 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1241 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1243 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1245 if (c0_scratch_reg >= 0) {
1247 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1248 build_tlb_write_entry(p, l, r, tlb_random);
1249 uasm_l_leave(l, *p);
1250 rv.restore_scratch = 1;
1251 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1252 build_tlb_write_entry(p, l, r, tlb_random);
1253 uasm_l_leave(l, *p);
1254 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1256 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1257 build_tlb_write_entry(p, l, r, tlb_random);
1258 uasm_l_leave(l, *p);
1259 rv.restore_scratch = 1;
1262 uasm_i_eret(p); /* return from trap */
1268 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1269 * because EXL == 0. If we wrap, we can also use the 32 instruction
1270 * slots before the XTLB refill exception handler which belong to the
1271 * unused TLB refill exception.
1273 #define MIPS64_REFILL_INSNS 32
1275 static void build_r4000_tlb_refill_handler(void)
1277 u32 *p = tlb_handler;
1278 struct uasm_label *l = labels;
1279 struct uasm_reloc *r = relocs;
1281 unsigned int final_len;
1282 struct mips_huge_tlb_info htlb_info __maybe_unused;
1283 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1285 memset(tlb_handler, 0, sizeof(tlb_handler));
1286 memset(labels, 0, sizeof(labels));
1287 memset(relocs, 0, sizeof(relocs));
1288 memset(final_handler, 0, sizeof(final_handler));
1290 if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1291 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1293 vmalloc_mode = refill_scratch;
1295 htlb_info.huge_pte = K0;
1296 htlb_info.restore_scratch = 0;
1297 htlb_info.need_reload_pte = true;
1298 vmalloc_mode = refill_noscratch;
1300 * create the plain linear handler
1302 if (bcm1250_m3_war()) {
1303 unsigned int segbits = 44;
1305 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1306 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1307 uasm_i_xor(&p, K0, K0, K1);
1308 uasm_i_dsrl_safe(&p, K1, K0, 62);
1309 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1310 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1311 uasm_i_or(&p, K0, K0, K1);
1312 uasm_il_bnez(&p, &r, K0, label_leave);
1313 /* No need for uasm_i_nop */
1317 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1319 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1322 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1323 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1326 build_get_ptep(&p, K0, K1);
1327 build_update_entries(&p, K0, K1);
1328 build_tlb_write_entry(&p, &l, &r, tlb_random);
1329 uasm_l_leave(&l, p);
1330 uasm_i_eret(&p); /* return from trap */
1332 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1333 uasm_l_tlb_huge_update(&l, p);
1334 if (htlb_info.need_reload_pte)
1335 UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1336 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1337 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1338 htlb_info.restore_scratch);
1342 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1346 * Overflow check: For the 64bit handler, we need at least one
1347 * free instruction slot for the wrap-around branch. In worst
1348 * case, if the intended insertion point is a delay slot, we
1349 * need three, with the second nop'ed and the third being
1352 switch (boot_cpu_type()) {
1354 if (sizeof(long) == 4) {
1356 /* Loongson2 ebase is different than r4k, we have more space */
1357 if ((p - tlb_handler) > 64)
1358 panic("TLB refill handler space exceeded");
1360 * Now fold the handler in the TLB refill handler space.
1363 /* Simplest case, just copy the handler. */
1364 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1365 final_len = p - tlb_handler;
1368 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1369 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1370 && uasm_insn_has_bdelay(relocs,
1371 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1372 panic("TLB refill handler space exceeded");
1374 * Now fold the handler in the TLB refill handler space.
1376 f = final_handler + MIPS64_REFILL_INSNS;
1377 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1378 /* Just copy the handler. */
1379 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1380 final_len = p - tlb_handler;
1382 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1383 const enum label_id ls = label_tlb_huge_update;
1385 const enum label_id ls = label_vmalloc;
1391 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1393 BUG_ON(i == ARRAY_SIZE(labels));
1394 split = labels[i].addr;
1397 * See if we have overflown one way or the other.
1399 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1400 split < p - MIPS64_REFILL_INSNS)
1405 * Split two instructions before the end. One
1406 * for the branch and one for the instruction
1407 * in the delay slot.
1409 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1412 * If the branch would fall in a delay slot,
1413 * we must back up an additional instruction
1414 * so that it is no longer in a delay slot.
1416 if (uasm_insn_has_bdelay(relocs, split - 1))
1419 /* Copy first part of the handler. */
1420 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1421 f += split - tlb_handler;
1424 /* Insert branch. */
1425 uasm_l_split(&l, final_handler);
1426 uasm_il_b(&f, &r, label_split);
1427 if (uasm_insn_has_bdelay(relocs, split))
1430 uasm_copy_handler(relocs, labels,
1431 split, split + 1, f);
1432 uasm_move_labels(labels, f, f + 1, -1);
1438 /* Copy the rest of the handler. */
1439 uasm_copy_handler(relocs, labels, split, p, final_handler);
1440 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1447 uasm_resolve_relocs(relocs, labels);
1448 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1451 memcpy((void *)ebase, final_handler, 0x100);
1452 local_flush_icache_range(ebase, ebase + 0x100);
1454 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1457 static void setup_pw(void)
1459 unsigned long pgd_i, pgd_w;
1460 #ifndef __PAGETABLE_PMD_FOLDED
1461 unsigned long pmd_i, pmd_w;
1463 unsigned long pt_i, pt_w;
1464 unsigned long pte_i, pte_w;
1465 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1468 psn = ilog2(_PAGE_HUGE); /* bit used to indicate huge page */
1470 pgd_i = PGDIR_SHIFT; /* 1st level PGD */
1471 #ifndef __PAGETABLE_PMD_FOLDED
1472 pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1474 pmd_i = PMD_SHIFT; /* 2nd level PMD */
1475 pmd_w = PMD_SHIFT - PAGE_SHIFT;
1477 pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1480 pt_i = PAGE_SHIFT; /* 3rd level PTE */
1481 pt_w = PAGE_SHIFT - 3;
1483 pte_i = ilog2(_PAGE_GLOBAL);
1486 #ifndef __PAGETABLE_PMD_FOLDED
1487 write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1488 write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1490 write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1491 write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1494 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1495 write_c0_pwctl(1 << 6 | psn);
1497 write_c0_kpgd(swapper_pg_dir);
1498 kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1501 static void build_loongson3_tlb_refill_handler(void)
1503 u32 *p = tlb_handler;
1504 struct uasm_label *l = labels;
1505 struct uasm_reloc *r = relocs;
1507 memset(labels, 0, sizeof(labels));
1508 memset(relocs, 0, sizeof(relocs));
1509 memset(tlb_handler, 0, sizeof(tlb_handler));
1511 if (check_for_high_segbits) {
1512 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1513 uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1514 uasm_il_beqz(&p, &r, K1, label_vmalloc);
1517 uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1519 uasm_l_vmalloc(&l, p);
1522 uasm_i_dmfc0(&p, K1, C0_PGD);
1524 uasm_i_lddir(&p, K0, K1, 3); /* global page dir */
1525 #ifndef __PAGETABLE_PMD_FOLDED
1526 uasm_i_lddir(&p, K1, K0, 1); /* middle page dir */
1528 uasm_i_ldpte(&p, K1, 0); /* even */
1529 uasm_i_ldpte(&p, K1, 1); /* odd */
1532 /* restore page mask */
1533 if (PM_DEFAULT_MASK >> 16) {
1534 uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1535 uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1536 uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1537 } else if (PM_DEFAULT_MASK) {
1538 uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1539 uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1541 uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1546 if (check_for_high_segbits) {
1547 uasm_l_large_segbits_fault(&l, p);
1548 UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1553 uasm_resolve_relocs(relocs, labels);
1554 memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1555 local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1556 dump_handler("loongson3_tlb_refill", (u32 *)(ebase + 0x80), 32);
1559 extern u32 handle_tlbl[], handle_tlbl_end[];
1560 extern u32 handle_tlbs[], handle_tlbs_end[];
1561 extern u32 handle_tlbm[], handle_tlbm_end[];
1562 extern u32 tlbmiss_handler_setup_pgd_start[], tlbmiss_handler_setup_pgd[];
1563 extern u32 tlbmiss_handler_setup_pgd_end[];
1565 static void build_setup_pgd(void)
1568 const int __maybe_unused a1 = 5;
1569 const int __maybe_unused a2 = 6;
1570 u32 *p = tlbmiss_handler_setup_pgd_start;
1571 const int tlbmiss_handler_setup_pgd_size =
1572 tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1573 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1574 long pgdc = (long)pgd_current;
1577 memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1578 sizeof(tlbmiss_handler_setup_pgd[0]));
1579 memset(labels, 0, sizeof(labels));
1580 memset(relocs, 0, sizeof(relocs));
1581 pgd_reg = allocate_kscratch();
1582 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1583 if (pgd_reg == -1) {
1584 struct uasm_label *l = labels;
1585 struct uasm_reloc *r = relocs;
1587 /* PGD << 11 in c0_Context */
1589 * If it is a ckseg0 address, convert to a physical
1590 * address. Shifting right by 29 and adding 4 will
1591 * result in zero for these addresses.
1594 UASM_i_SRA(&p, a1, a0, 29);
1595 UASM_i_ADDIU(&p, a1, a1, 4);
1596 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1598 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1599 uasm_l_tlbl_goaround1(&l, p);
1600 UASM_i_SLL(&p, a0, a0, 11);
1601 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1605 /* PGD in c0_KScratch */
1607 UASM_i_MTC0(&p, a0, C0_PWBASE);
1609 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1615 /* Save PGD to pgd_current[smp_processor_id()] */
1616 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1617 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1618 UASM_i_LA_mostly(&p, a2, pgdc);
1619 UASM_i_ADDU(&p, a2, a2, a1);
1620 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1622 UASM_i_LA_mostly(&p, a2, pgdc);
1623 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1626 /* if pgd_reg is allocated, save PGD also to scratch register */
1627 if (pgd_reg != -1) {
1628 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1636 if (p >= tlbmiss_handler_setup_pgd_end)
1637 panic("tlbmiss_handler_setup_pgd space exceeded");
1639 uasm_resolve_relocs(relocs, labels);
1640 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1641 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1643 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1644 tlbmiss_handler_setup_pgd_size);
1648 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1651 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1653 uasm_i_lld(p, pte, 0, ptr);
1656 UASM_i_LL(p, pte, 0, ptr);
1658 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1660 uasm_i_ld(p, pte, 0, ptr);
1663 UASM_i_LW(p, pte, 0, ptr);
1668 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1669 unsigned int mode, unsigned int scratch)
1671 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1672 unsigned int swmode = mode & ~hwmode;
1674 if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1675 uasm_i_lui(p, scratch, swmode >> 16);
1676 uasm_i_or(p, pte, pte, scratch);
1677 BUG_ON(swmode & 0xffff);
1679 uasm_i_ori(p, pte, pte, mode);
1683 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1685 uasm_i_scd(p, pte, 0, ptr);
1688 UASM_i_SC(p, pte, 0, ptr);
1690 if (r10000_llsc_war())
1691 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1693 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1695 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1696 if (!cpu_has_64bits) {
1697 /* no uasm_i_nop needed */
1698 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1699 uasm_i_ori(p, pte, pte, hwmode);
1700 BUG_ON(hwmode & ~0xffff);
1701 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1702 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1703 /* no uasm_i_nop needed */
1704 uasm_i_lw(p, pte, 0, ptr);
1711 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1713 uasm_i_sd(p, pte, 0, ptr);
1716 UASM_i_SW(p, pte, 0, ptr);
1718 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1719 if (!cpu_has_64bits) {
1720 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1721 uasm_i_ori(p, pte, pte, hwmode);
1722 BUG_ON(hwmode & ~0xffff);
1723 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1724 uasm_i_lw(p, pte, 0, ptr);
1731 * Check if PTE is present, if not then jump to LABEL. PTR points to
1732 * the page table where this PTE is located, PTE will be re-loaded
1733 * with it's original value.
1736 build_pte_present(u32 **p, struct uasm_reloc **r,
1737 int pte, int ptr, int scratch, enum label_id lid)
1739 int t = scratch >= 0 ? scratch : pte;
1743 if (use_bbit_insns()) {
1744 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1747 if (_PAGE_PRESENT_SHIFT) {
1748 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1751 uasm_i_andi(p, t, cur, 1);
1752 uasm_il_beqz(p, r, t, lid);
1754 /* You lose the SMP race :-(*/
1755 iPTE_LW(p, pte, ptr);
1758 if (_PAGE_PRESENT_SHIFT) {
1759 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1762 uasm_i_andi(p, t, cur,
1763 (_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1764 uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1765 uasm_il_bnez(p, r, t, lid);
1767 /* You lose the SMP race :-(*/
1768 iPTE_LW(p, pte, ptr);
1772 /* Make PTE valid, store result in PTR. */
1774 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1775 unsigned int ptr, unsigned int scratch)
1777 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1779 iPTE_SW(p, r, pte, ptr, mode, scratch);
1783 * Check if PTE can be written to, if not branch to LABEL. Regardless
1784 * restore PTE with value from PTR when done.
1787 build_pte_writable(u32 **p, struct uasm_reloc **r,
1788 unsigned int pte, unsigned int ptr, int scratch,
1791 int t = scratch >= 0 ? scratch : pte;
1794 if (_PAGE_PRESENT_SHIFT) {
1795 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1798 uasm_i_andi(p, t, cur,
1799 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1800 uasm_i_xori(p, t, t,
1801 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1802 uasm_il_bnez(p, r, t, lid);
1804 /* You lose the SMP race :-(*/
1805 iPTE_LW(p, pte, ptr);
1810 /* Make PTE writable, update software status bits as well, then store
1814 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1815 unsigned int ptr, unsigned int scratch)
1817 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1820 iPTE_SW(p, r, pte, ptr, mode, scratch);
1824 * Check if PTE can be modified, if not branch to LABEL. Regardless
1825 * restore PTE with value from PTR when done.
1828 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1829 unsigned int pte, unsigned int ptr, int scratch,
1832 if (use_bbit_insns()) {
1833 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1836 int t = scratch >= 0 ? scratch : pte;
1837 uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1838 uasm_i_andi(p, t, t, 1);
1839 uasm_il_beqz(p, r, t, lid);
1841 /* You lose the SMP race :-(*/
1842 iPTE_LW(p, pte, ptr);
1846 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1850 * R3000 style TLB load/store/modify handlers.
1854 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1858 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1860 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1861 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1864 uasm_i_rfe(p); /* branch delay */
1868 * This places the pte into ENTRYLO0 and writes it with tlbwi
1869 * or tlbwr as appropriate. This is because the index register
1870 * may have the probe fail bit set as a result of a trap on a
1871 * kseg2 access, i.e. without refill. Then it returns.
1874 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1875 struct uasm_reloc **r, unsigned int pte,
1878 uasm_i_mfc0(p, tmp, C0_INDEX);
1879 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1880 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1881 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1882 uasm_i_tlbwi(p); /* cp0 delay */
1884 uasm_i_rfe(p); /* branch delay */
1885 uasm_l_r3000_write_probe_fail(l, *p);
1886 uasm_i_tlbwr(p); /* cp0 delay */
1888 uasm_i_rfe(p); /* branch delay */
1892 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1895 long pgdc = (long)pgd_current;
1897 uasm_i_mfc0(p, pte, C0_BADVADDR);
1898 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1899 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1900 uasm_i_srl(p, pte, pte, 22); /* load delay */
1901 uasm_i_sll(p, pte, pte, 2);
1902 uasm_i_addu(p, ptr, ptr, pte);
1903 uasm_i_mfc0(p, pte, C0_CONTEXT);
1904 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1905 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1906 uasm_i_addu(p, ptr, ptr, pte);
1907 uasm_i_lw(p, pte, 0, ptr);
1908 uasm_i_tlbp(p); /* load delay */
1911 static void build_r3000_tlb_load_handler(void)
1913 u32 *p = handle_tlbl;
1914 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1915 struct uasm_label *l = labels;
1916 struct uasm_reloc *r = relocs;
1918 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1919 memset(labels, 0, sizeof(labels));
1920 memset(relocs, 0, sizeof(relocs));
1922 build_r3000_tlbchange_handler_head(&p, K0, K1);
1923 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1924 uasm_i_nop(&p); /* load delay */
1925 build_make_valid(&p, &r, K0, K1, -1);
1926 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1928 uasm_l_nopage_tlbl(&l, p);
1929 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1932 if (p >= handle_tlbl_end)
1933 panic("TLB load handler fastpath space exceeded");
1935 uasm_resolve_relocs(relocs, labels);
1936 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1937 (unsigned int)(p - handle_tlbl));
1939 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1942 static void build_r3000_tlb_store_handler(void)
1944 u32 *p = handle_tlbs;
1945 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1946 struct uasm_label *l = labels;
1947 struct uasm_reloc *r = relocs;
1949 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1950 memset(labels, 0, sizeof(labels));
1951 memset(relocs, 0, sizeof(relocs));
1953 build_r3000_tlbchange_handler_head(&p, K0, K1);
1954 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1955 uasm_i_nop(&p); /* load delay */
1956 build_make_write(&p, &r, K0, K1, -1);
1957 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1959 uasm_l_nopage_tlbs(&l, p);
1960 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1963 if (p >= handle_tlbs_end)
1964 panic("TLB store handler fastpath space exceeded");
1966 uasm_resolve_relocs(relocs, labels);
1967 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1968 (unsigned int)(p - handle_tlbs));
1970 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1973 static void build_r3000_tlb_modify_handler(void)
1975 u32 *p = handle_tlbm;
1976 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1977 struct uasm_label *l = labels;
1978 struct uasm_reloc *r = relocs;
1980 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1981 memset(labels, 0, sizeof(labels));
1982 memset(relocs, 0, sizeof(relocs));
1984 build_r3000_tlbchange_handler_head(&p, K0, K1);
1985 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1986 uasm_i_nop(&p); /* load delay */
1987 build_make_write(&p, &r, K0, K1, -1);
1988 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1990 uasm_l_nopage_tlbm(&l, p);
1991 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1994 if (p >= handle_tlbm_end)
1995 panic("TLB modify handler fastpath space exceeded");
1997 uasm_resolve_relocs(relocs, labels);
1998 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1999 (unsigned int)(p - handle_tlbm));
2001 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
2003 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2006 * R4000 style TLB load/store/modify handlers.
2008 static struct work_registers
2009 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2010 struct uasm_reloc **r)
2012 struct work_registers wr = build_get_work_registers(p);
2015 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2017 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2020 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2022 * For huge tlb entries, pmd doesn't contain an address but
2023 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2024 * see if we need to jump to huge tlb processing.
2026 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2029 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2030 UASM_i_LW(p, wr.r2, 0, wr.r2);
2031 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2032 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2033 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2036 uasm_l_smp_pgtable_change(l, *p);
2038 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2039 if (!m4kc_tlbp_war()) {
2040 build_tlb_probe_entry(p);
2042 /* race condition happens, leaving */
2044 uasm_i_mfc0(p, wr.r3, C0_INDEX);
2045 uasm_il_bltz(p, r, wr.r3, label_leave);
2053 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2054 struct uasm_reloc **r, unsigned int tmp,
2057 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2058 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2059 build_update_entries(p, tmp, ptr);
2060 build_tlb_write_entry(p, l, r, tlb_indexed);
2061 uasm_l_leave(l, *p);
2062 build_restore_work_registers(p);
2063 uasm_i_eret(p); /* return from trap */
2066 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2070 static void build_r4000_tlb_load_handler(void)
2072 u32 *p = handle_tlbl;
2073 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
2074 struct uasm_label *l = labels;
2075 struct uasm_reloc *r = relocs;
2076 struct work_registers wr;
2078 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
2079 memset(labels, 0, sizeof(labels));
2080 memset(relocs, 0, sizeof(relocs));
2082 if (bcm1250_m3_war()) {
2083 unsigned int segbits = 44;
2085 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2086 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2087 uasm_i_xor(&p, K0, K0, K1);
2088 uasm_i_dsrl_safe(&p, K1, K0, 62);
2089 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2090 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2091 uasm_i_or(&p, K0, K0, K1);
2092 uasm_il_bnez(&p, &r, K0, label_leave);
2093 /* No need for uasm_i_nop */
2096 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2097 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2098 if (m4kc_tlbp_war())
2099 build_tlb_probe_entry(&p);
2101 if (cpu_has_rixi && !cpu_has_rixiex) {
2103 * If the page is not _PAGE_VALID, RI or XI could not
2104 * have triggered it. Skip the expensive test..
2106 if (use_bbit_insns()) {
2107 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2108 label_tlbl_goaround1);
2110 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2111 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2117 switch (current_cpu_type()) {
2119 if (cpu_has_mips_r2_exec_hazard) {
2122 case CPU_CAVIUM_OCTEON:
2123 case CPU_CAVIUM_OCTEON_PLUS:
2124 case CPU_CAVIUM_OCTEON2:
2129 /* Examine entrylo 0 or 1 based on ptr. */
2130 if (use_bbit_insns()) {
2131 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2133 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2134 uasm_i_beqz(&p, wr.r3, 8);
2136 /* load it in the delay slot*/
2137 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2138 /* load it if ptr is odd */
2139 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2141 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2142 * XI must have triggered it.
2144 if (use_bbit_insns()) {
2145 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2147 uasm_l_tlbl_goaround1(&l, p);
2149 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2150 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2153 uasm_l_tlbl_goaround1(&l, p);
2155 build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2156 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2158 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2160 * This is the entry point when build_r4000_tlbchange_handler_head
2161 * spots a huge page.
2163 uasm_l_tlb_huge_update(&l, p);
2164 iPTE_LW(&p, wr.r1, wr.r2);
2165 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2166 build_tlb_probe_entry(&p);
2168 if (cpu_has_rixi && !cpu_has_rixiex) {
2170 * If the page is not _PAGE_VALID, RI or XI could not
2171 * have triggered it. Skip the expensive test..
2173 if (use_bbit_insns()) {
2174 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2175 label_tlbl_goaround2);
2177 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2178 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2184 switch (current_cpu_type()) {
2186 if (cpu_has_mips_r2_exec_hazard) {
2189 case CPU_CAVIUM_OCTEON:
2190 case CPU_CAVIUM_OCTEON_PLUS:
2191 case CPU_CAVIUM_OCTEON2:
2196 /* Examine entrylo 0 or 1 based on ptr. */
2197 if (use_bbit_insns()) {
2198 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2200 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2201 uasm_i_beqz(&p, wr.r3, 8);
2203 /* load it in the delay slot*/
2204 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2205 /* load it if ptr is odd */
2206 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2208 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2209 * XI must have triggered it.
2211 if (use_bbit_insns()) {
2212 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2214 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2215 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2217 if (PM_DEFAULT_MASK == 0)
2220 * We clobbered C0_PAGEMASK, restore it. On the other branch
2221 * it is restored in build_huge_tlb_write_entry.
2223 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2225 uasm_l_tlbl_goaround2(&l, p);
2227 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2228 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2231 uasm_l_nopage_tlbl(&l, p);
2232 build_restore_work_registers(&p);
2233 #ifdef CONFIG_CPU_MICROMIPS
2234 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2235 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2236 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2240 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2243 if (p >= handle_tlbl_end)
2244 panic("TLB load handler fastpath space exceeded");
2246 uasm_resolve_relocs(relocs, labels);
2247 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2248 (unsigned int)(p - handle_tlbl));
2250 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2253 static void build_r4000_tlb_store_handler(void)
2255 u32 *p = handle_tlbs;
2256 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2257 struct uasm_label *l = labels;
2258 struct uasm_reloc *r = relocs;
2259 struct work_registers wr;
2261 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2262 memset(labels, 0, sizeof(labels));
2263 memset(relocs, 0, sizeof(relocs));
2265 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2266 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2267 if (m4kc_tlbp_war())
2268 build_tlb_probe_entry(&p);
2269 build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2270 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2272 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2274 * This is the entry point when
2275 * build_r4000_tlbchange_handler_head spots a huge page.
2277 uasm_l_tlb_huge_update(&l, p);
2278 iPTE_LW(&p, wr.r1, wr.r2);
2279 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2280 build_tlb_probe_entry(&p);
2281 uasm_i_ori(&p, wr.r1, wr.r1,
2282 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2283 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2286 uasm_l_nopage_tlbs(&l, p);
2287 build_restore_work_registers(&p);
2288 #ifdef CONFIG_CPU_MICROMIPS
2289 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2290 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2291 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2295 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2298 if (p >= handle_tlbs_end)
2299 panic("TLB store handler fastpath space exceeded");
2301 uasm_resolve_relocs(relocs, labels);
2302 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2303 (unsigned int)(p - handle_tlbs));
2305 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2308 static void build_r4000_tlb_modify_handler(void)
2310 u32 *p = handle_tlbm;
2311 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2312 struct uasm_label *l = labels;
2313 struct uasm_reloc *r = relocs;
2314 struct work_registers wr;
2316 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2317 memset(labels, 0, sizeof(labels));
2318 memset(relocs, 0, sizeof(relocs));
2320 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2321 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2322 if (m4kc_tlbp_war())
2323 build_tlb_probe_entry(&p);
2324 /* Present and writable bits set, set accessed and dirty bits. */
2325 build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2326 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2328 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2330 * This is the entry point when
2331 * build_r4000_tlbchange_handler_head spots a huge page.
2333 uasm_l_tlb_huge_update(&l, p);
2334 iPTE_LW(&p, wr.r1, wr.r2);
2335 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2336 build_tlb_probe_entry(&p);
2337 uasm_i_ori(&p, wr.r1, wr.r1,
2338 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2339 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2342 uasm_l_nopage_tlbm(&l, p);
2343 build_restore_work_registers(&p);
2344 #ifdef CONFIG_CPU_MICROMIPS
2345 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2346 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2347 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2351 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2354 if (p >= handle_tlbm_end)
2355 panic("TLB modify handler fastpath space exceeded");
2357 uasm_resolve_relocs(relocs, labels);
2358 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2359 (unsigned int)(p - handle_tlbm));
2361 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2364 static void flush_tlb_handlers(void)
2366 local_flush_icache_range((unsigned long)handle_tlbl,
2367 (unsigned long)handle_tlbl_end);
2368 local_flush_icache_range((unsigned long)handle_tlbs,
2369 (unsigned long)handle_tlbs_end);
2370 local_flush_icache_range((unsigned long)handle_tlbm,
2371 (unsigned long)handle_tlbm_end);
2372 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2373 (unsigned long)tlbmiss_handler_setup_pgd_end);
2376 static void print_htw_config(void)
2378 unsigned long config;
2380 const int field = 2 * sizeof(unsigned long);
2382 config = read_c0_pwfield();
2383 pr_debug("PWField (0x%0*lx): GDI: 0x%02lx UDI: 0x%02lx MDI: 0x%02lx PTI: 0x%02lx PTEI: 0x%02lx\n",
2385 (config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2386 (config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2387 (config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2388 (config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2389 (config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2391 config = read_c0_pwsize();
2392 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",
2394 (config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2395 (config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2396 (config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2397 (config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2398 (config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2399 (config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2401 pwctl = read_c0_pwctl();
2402 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",
2404 (pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2405 (pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2406 (pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2407 (pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2408 (pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2409 (pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2410 (pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2413 static void config_htw_params(void)
2415 unsigned long pwfield, pwsize, ptei;
2416 unsigned int config;
2419 * We are using 2-level page tables, so we only need to
2420 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2421 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2422 * write values less than 0xc in these fields because the entire
2423 * write will be dropped. As a result of which, we must preserve
2424 * the original reset values and overwrite only what we really want.
2427 pwfield = read_c0_pwfield();
2428 /* re-initialize the GDI field */
2429 pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2430 pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2431 /* re-initialize the PTI field including the even/odd bit */
2432 pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2433 pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2434 if (CONFIG_PGTABLE_LEVELS >= 3) {
2435 pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2436 pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2438 /* Set the PTEI right shift */
2439 ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2441 write_c0_pwfield(pwfield);
2442 /* Check whether the PTEI value is supported */
2443 back_to_back_c0_hazard();
2444 pwfield = read_c0_pwfield();
2445 if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2447 pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2450 * Drop option to avoid HTW being enabled via another path
2453 current_cpu_data.options &= ~MIPS_CPU_HTW;
2457 pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2458 pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2459 if (CONFIG_PGTABLE_LEVELS >= 3)
2460 pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2462 /* Set pointer size to size of directory pointers */
2463 if (IS_ENABLED(CONFIG_64BIT))
2464 pwsize |= MIPS_PWSIZE_PS_MASK;
2465 /* PTEs may be multiple pointers long (e.g. with XPA) */
2466 pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2467 & MIPS_PWSIZE_PTEW_MASK;
2469 write_c0_pwsize(pwsize);
2471 /* Make sure everything is set before we enable the HTW */
2472 back_to_back_c0_hazard();
2475 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2478 config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2479 if (IS_ENABLED(CONFIG_64BIT))
2480 config |= MIPS_PWCTL_XU_MASK;
2481 write_c0_pwctl(config);
2482 pr_info("Hardware Page Table Walker enabled\n");
2487 static void config_xpa_params(void)
2490 unsigned int pagegrain;
2492 if (mips_xpa_disabled) {
2493 pr_info("Extended Physical Addressing (XPA) disabled\n");
2497 pagegrain = read_c0_pagegrain();
2498 write_c0_pagegrain(pagegrain | PG_ELPA);
2499 back_to_back_c0_hazard();
2500 pagegrain = read_c0_pagegrain();
2502 if (pagegrain & PG_ELPA)
2503 pr_info("Extended Physical Addressing (XPA) enabled\n");
2505 panic("Extended Physical Addressing (XPA) disabled");
2509 static void check_pabits(void)
2511 unsigned long entry;
2512 unsigned pabits, fillbits;
2514 if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2516 * We'll only be making use of the fact that we can rotate bits
2517 * into the fill if the CPU supports RIXI, so don't bother
2518 * probing this for CPUs which don't.
2523 write_c0_entrylo0(~0ul);
2524 back_to_back_c0_hazard();
2525 entry = read_c0_entrylo0();
2527 /* clear all non-PFN bits */
2528 entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2529 entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2531 /* find a lower bound on PABITS, and upper bound on fill bits */
2532 pabits = fls_long(entry) + 6;
2533 fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2535 /* minus the RI & XI bits */
2536 fillbits -= min_t(unsigned, fillbits, 2);
2538 if (fillbits >= ilog2(_PAGE_NO_EXEC))
2539 fill_includes_sw_bits = true;
2541 pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2544 void build_tlb_refill_handler(void)
2547 * The refill handler is generated per-CPU, multi-node systems
2548 * may have local storage for it. The other handlers are only
2551 static int run_once = 0;
2553 if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2554 panic("Kernels supporting XPA currently require CPUs with RIXI");
2556 output_pgtable_bits_defines();
2560 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2563 switch (current_cpu_type()) {
2571 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2572 if (cpu_has_local_ebase)
2573 build_r3000_tlb_refill_handler();
2575 if (!cpu_has_local_ebase)
2576 build_r3000_tlb_refill_handler();
2578 build_r3000_tlb_load_handler();
2579 build_r3000_tlb_store_handler();
2580 build_r3000_tlb_modify_handler();
2581 flush_tlb_handlers();
2585 panic("No R3000 TLB refill handler");
2591 panic("No R6000 TLB refill handler yet");
2595 panic("No R8000 TLB refill handler yet");
2603 scratch_reg = allocate_kscratch();
2605 build_r4000_tlb_load_handler();
2606 build_r4000_tlb_store_handler();
2607 build_r4000_tlb_modify_handler();
2609 build_loongson3_tlb_refill_handler();
2610 else if (!cpu_has_local_ebase)
2611 build_r4000_tlb_refill_handler();
2612 flush_tlb_handlers();
2615 if (cpu_has_local_ebase)
2616 build_r4000_tlb_refill_handler();
2618 config_xpa_params();
2620 config_htw_params();