2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Synthesize TLB refill handlers at runtime.
8 * Copyright (C) 2004, 2005, 2006, 2008 Thiemo Seufer
9 * Copyright (C) 2005, 2007, 2008, 2009 Maciej W. Rozycki
10 * Copyright (C) 2006 Ralf Baechle (ralf@linux-mips.org)
11 * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12 * Copyright (C) 2011 MIPS Technologies, Inc.
14 * ... and the days got worse and worse and now you see
15 * I've gone completely out of my mind.
17 * They're coming to take me a away haha
18 * they're coming to take me a away hoho hihi haha
19 * to the funny farm where code is beautiful all the time ...
21 * (Condolences to Napoleon XIV)
24 #include <linux/bug.h>
25 #include <linux/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.h>
30 #include <linux/cache.h>
32 #include <asm/cacheflush.h>
33 #include <asm/cpu-type.h>
34 #include <asm/mmu_context.h>
35 #include <asm/pgtable.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 static inline int __maybe_unused bcm1250_m3_war(void)
88 return BCM1250_M3_WAR;
91 static inline int __maybe_unused r10000_llsc_war(void)
93 return R10000_LLSC_WAR;
96 static int use_bbit_insns(void)
98 switch (current_cpu_type()) {
99 case CPU_CAVIUM_OCTEON:
100 case CPU_CAVIUM_OCTEON_PLUS:
101 case CPU_CAVIUM_OCTEON2:
102 case CPU_CAVIUM_OCTEON3:
109 static int use_lwx_insns(void)
111 switch (current_cpu_type()) {
112 case CPU_CAVIUM_OCTEON2:
113 case CPU_CAVIUM_OCTEON3:
119 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
120 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
121 static bool scratchpad_available(void)
125 static int scratchpad_offset(int i)
128 * CVMSEG starts at address -32768 and extends for
129 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
131 i += 1; /* Kernel use starts at the top and works down. */
132 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
135 static bool scratchpad_available(void)
139 static int scratchpad_offset(int i)
142 /* Really unreachable, but evidently some GCC want this. */
147 * Found by experiment: At least some revisions of the 4kc throw under
148 * some circumstances a machine check exception, triggered by invalid
149 * values in the index register. Delaying the tlbp instruction until
150 * after the next branch, plus adding an additional nop in front of
151 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
152 * why; it's not an issue caused by the core RTL.
155 static int m4kc_tlbp_war(void)
157 return current_cpu_type() == CPU_4KC;
160 /* Handle labels (which must be positive integers). */
162 label_second_part = 1,
167 label_split = label_tlbw_hazard_0 + 8,
168 label_tlbl_goaround1,
169 label_tlbl_goaround2,
173 label_smp_pgtable_change,
174 label_r3000_write_probe_fail,
175 label_large_segbits_fault,
176 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
177 label_tlb_huge_update,
181 UASM_L_LA(_second_part)
184 UASM_L_LA(_vmalloc_done)
185 /* _tlbw_hazard_x is handled differently. */
187 UASM_L_LA(_tlbl_goaround1)
188 UASM_L_LA(_tlbl_goaround2)
189 UASM_L_LA(_nopage_tlbl)
190 UASM_L_LA(_nopage_tlbs)
191 UASM_L_LA(_nopage_tlbm)
192 UASM_L_LA(_smp_pgtable_change)
193 UASM_L_LA(_r3000_write_probe_fail)
194 UASM_L_LA(_large_segbits_fault)
195 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
196 UASM_L_LA(_tlb_huge_update)
199 static int hazard_instance;
201 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
205 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
212 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
216 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
224 * pgtable bits are assigned dynamically depending on processor feature
225 * and statically based on kernel configuration. This spits out the actual
226 * values the kernel is using. Required to make sense from disassembled
227 * TLB exception handlers.
229 static void output_pgtable_bits_defines(void)
231 #define pr_define(fmt, ...) \
232 pr_debug("#define " fmt, ##__VA_ARGS__)
234 pr_debug("#include <asm/asm.h>\n");
235 pr_debug("#include <asm/regdef.h>\n");
238 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
239 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
240 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
241 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
242 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
243 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
244 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
246 #ifdef _PAGE_NO_EXEC_SHIFT
248 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
250 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
251 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
252 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
253 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
257 static inline void dump_handler(const char *symbol, const void *start, const void *end)
259 unsigned int count = (end - start) / sizeof(u32);
260 const u32 *handler = start;
263 pr_debug("LEAF(%s)\n", symbol);
265 pr_debug("\t.set push\n");
266 pr_debug("\t.set noreorder\n");
268 for (i = 0; i < count; i++)
269 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
271 pr_debug("\t.set\tpop\n");
273 pr_debug("\tEND(%s)\n", symbol);
276 /* The only general purpose registers allowed in TLB handlers. */
280 /* Some CP0 registers */
281 #define C0_INDEX 0, 0
282 #define C0_ENTRYLO0 2, 0
283 #define C0_TCBIND 2, 2
284 #define C0_ENTRYLO1 3, 0
285 #define C0_CONTEXT 4, 0
286 #define C0_PAGEMASK 5, 0
287 #define C0_PWBASE 5, 5
288 #define C0_PWFIELD 5, 6
289 #define C0_PWSIZE 5, 7
290 #define C0_PWCTL 6, 6
291 #define C0_BADVADDR 8, 0
293 #define C0_ENTRYHI 10, 0
295 #define C0_XCONTEXT 20, 0
298 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
300 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
303 /* The worst case length of the handler is around 18 instructions for
304 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
305 * Maximum space available is 32 instructions for R3000 and 64
306 * instructions for R4000.
308 * We deliberately chose a buffer size of 128, so we won't scribble
309 * over anything important on overflow before we panic.
311 static u32 tlb_handler[128];
313 /* simply assume worst case size for labels and relocs */
314 static struct uasm_label labels[128];
315 static struct uasm_reloc relocs[128];
317 static int check_for_high_segbits;
318 static bool fill_includes_sw_bits;
320 static unsigned int kscratch_used_mask;
322 static inline int __maybe_unused c0_kscratch(void)
324 switch (current_cpu_type()) {
333 static int allocate_kscratch(void)
336 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
343 r--; /* make it zero based */
345 kscratch_used_mask |= (1 << r);
350 static int scratch_reg;
352 EXPORT_SYMBOL_GPL(pgd_reg);
353 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
355 static struct work_registers build_get_work_registers(u32 **p)
357 struct work_registers r;
359 if (scratch_reg >= 0) {
360 /* Save in CPU local C0_KScratch? */
361 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
368 if (num_possible_cpus() > 1) {
369 /* Get smp_processor_id */
370 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
371 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
373 /* handler_reg_save index in K0 */
374 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
376 UASM_i_LA(p, K1, (long)&handler_reg_save);
377 UASM_i_ADDU(p, K0, K0, K1);
379 UASM_i_LA(p, K0, (long)&handler_reg_save);
381 /* K0 now points to save area, save $1 and $2 */
382 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
383 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
391 static void build_restore_work_registers(u32 **p)
393 if (scratch_reg >= 0) {
395 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
398 /* K0 already points to save area, restore $1 and $2 */
399 UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
400 UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
403 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
406 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
407 * we cannot do r3000 under these circumstances.
409 * The R3000 TLB handler is simple.
411 static void build_r3000_tlb_refill_handler(void)
413 long pgdc = (long)pgd_current;
416 memset(tlb_handler, 0, sizeof(tlb_handler));
419 uasm_i_mfc0(&p, K0, C0_BADVADDR);
420 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
421 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
422 uasm_i_srl(&p, K0, K0, 22); /* load delay */
423 uasm_i_sll(&p, K0, K0, 2);
424 uasm_i_addu(&p, K1, K1, K0);
425 uasm_i_mfc0(&p, K0, C0_CONTEXT);
426 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
427 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
428 uasm_i_addu(&p, K1, K1, K0);
429 uasm_i_lw(&p, K0, 0, K1);
430 uasm_i_nop(&p); /* load delay */
431 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
432 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
433 uasm_i_tlbwr(&p); /* cp0 delay */
435 uasm_i_rfe(&p); /* branch delay */
437 if (p > tlb_handler + 32)
438 panic("TLB refill handler space exceeded");
440 pr_debug("Wrote TLB refill handler (%u instructions).\n",
441 (unsigned int)(p - tlb_handler));
443 memcpy((void *)ebase, tlb_handler, 0x80);
444 local_flush_icache_range(ebase, ebase + 0x80);
445 dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
447 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
450 * The R4000 TLB handler is much more complicated. We have two
451 * consecutive handler areas with 32 instructions space each.
452 * Since they aren't used at the same time, we can overflow in the
453 * other one.To keep things simple, we first assume linear space,
454 * then we relocate it to the final handler layout as needed.
456 static u32 final_handler[64];
461 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
462 * 2. A timing hazard exists for the TLBP instruction.
464 * stalling_instruction
467 * The JTLB is being read for the TLBP throughout the stall generated by the
468 * previous instruction. This is not really correct as the stalling instruction
469 * can modify the address used to access the JTLB. The failure symptom is that
470 * the TLBP instruction will use an address created for the stalling instruction
471 * and not the address held in C0_ENHI and thus report the wrong results.
473 * The software work-around is to not allow the instruction preceding the TLBP
474 * to stall - make it an NOP or some other instruction guaranteed not to stall.
476 * Errata 2 will not be fixed. This errata is also on the R5000.
478 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
480 static void __maybe_unused build_tlb_probe_entry(u32 **p)
482 switch (current_cpu_type()) {
483 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
498 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
499 struct uasm_reloc **r,
500 enum tlb_write_entry wmode)
502 void(*tlbw)(u32 **) = NULL;
505 case tlb_random: tlbw = uasm_i_tlbwr; break;
506 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
509 if (cpu_has_mips_r2_r6) {
510 if (cpu_has_mips_r2_exec_hazard)
516 switch (current_cpu_type()) {
524 * This branch uses up a mtc0 hazard nop slot and saves
525 * two nops after the tlbw instruction.
527 uasm_bgezl_hazard(p, r, hazard_instance);
529 uasm_bgezl_label(l, p, hazard_instance);
543 uasm_i_nop(p); /* QED specifies 2 nops hazard */
544 uasm_i_nop(p); /* QED specifies 2 nops hazard */
618 panic("No TLB refill handler yet (CPU type: %d)",
623 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
625 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
628 if (_PAGE_GLOBAL_SHIFT == 0) {
629 /* pte_t is already in EntryLo format */
633 if (cpu_has_rixi && _PAGE_NO_EXEC != 0) {
634 if (fill_includes_sw_bits) {
635 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
637 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
638 UASM_i_ROTR(p, reg, reg,
639 ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
642 #ifdef CONFIG_PHYS_ADDR_T_64BIT
643 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
645 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
650 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
652 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
653 unsigned int tmp, enum label_id lid,
656 if (restore_scratch) {
658 * Ensure the MFC0 below observes the value written to the
659 * KScratch register by the prior MTC0.
661 if (scratch_reg >= 0)
664 /* Reset default page size */
665 if (PM_DEFAULT_MASK >> 16) {
666 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
667 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
668 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
669 uasm_il_b(p, r, lid);
670 } else if (PM_DEFAULT_MASK) {
671 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
672 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
673 uasm_il_b(p, r, lid);
675 uasm_i_mtc0(p, 0, C0_PAGEMASK);
676 uasm_il_b(p, r, lid);
678 if (scratch_reg >= 0)
679 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
681 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
683 /* Reset default page size */
684 if (PM_DEFAULT_MASK >> 16) {
685 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
686 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
687 uasm_il_b(p, r, lid);
688 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
689 } else if (PM_DEFAULT_MASK) {
690 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
691 uasm_il_b(p, r, lid);
692 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
694 uasm_il_b(p, r, lid);
695 uasm_i_mtc0(p, 0, C0_PAGEMASK);
700 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
701 struct uasm_reloc **r,
703 enum tlb_write_entry wmode,
706 /* Set huge page tlb entry size */
707 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
708 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
709 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
711 build_tlb_write_entry(p, l, r, wmode);
713 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
717 * Check if Huge PTE is present, if so then jump to LABEL.
720 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
721 unsigned int pmd, int lid)
723 UASM_i_LW(p, tmp, 0, pmd);
724 if (use_bbit_insns()) {
725 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
727 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
728 uasm_il_bnez(p, r, tmp, lid);
732 static void build_huge_update_entries(u32 **p, unsigned int pte,
738 * A huge PTE describes an area the size of the
739 * configured huge page size. This is twice the
740 * of the large TLB entry size we intend to use.
741 * A TLB entry half the size of the configured
742 * huge page size is configured into entrylo0
743 * and entrylo1 to cover the contiguous huge PTE
746 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
748 /* We can clobber tmp. It isn't used after this.*/
750 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
752 build_convert_pte_to_entrylo(p, pte);
753 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
754 /* convert to entrylo1 */
756 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
758 UASM_i_ADDU(p, pte, pte, tmp);
760 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
763 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
764 struct uasm_label **l,
770 UASM_i_SC(p, pte, 0, ptr);
771 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
772 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
774 UASM_i_SW(p, pte, 0, ptr);
776 if (cpu_has_ftlb && flush) {
777 BUG_ON(!cpu_has_tlbinv);
779 UASM_i_MFC0(p, ptr, C0_ENTRYHI);
780 uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
781 UASM_i_MTC0(p, ptr, C0_ENTRYHI);
782 build_tlb_write_entry(p, l, r, tlb_indexed);
784 uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
785 UASM_i_MTC0(p, ptr, C0_ENTRYHI);
786 build_huge_update_entries(p, pte, ptr);
787 build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
792 build_huge_update_entries(p, pte, ptr);
793 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
795 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
799 * TMP and PTR are scratch.
800 * TMP will be clobbered, PTR will hold the pmd entry.
802 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
803 unsigned int tmp, unsigned int ptr)
805 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
806 long pgdc = (long)pgd_current;
809 * The vmalloc handling is not in the hotpath.
811 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
813 if (check_for_high_segbits) {
815 * The kernel currently implicitely assumes that the
816 * MIPS SEGBITS parameter for the processor is
817 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
818 * allocate virtual addresses outside the maximum
819 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
820 * that doesn't prevent user code from accessing the
821 * higher xuseg addresses. Here, we make sure that
822 * everything but the lower xuseg addresses goes down
823 * the module_alloc/vmalloc path.
825 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
826 uasm_il_bnez(p, r, ptr, label_vmalloc);
828 uasm_il_bltz(p, r, tmp, label_vmalloc);
830 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
833 /* pgd is in pgd_reg */
835 UASM_i_MFC0(p, ptr, C0_PWBASE);
837 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
839 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
841 * &pgd << 11 stored in CONTEXT [23..63].
843 UASM_i_MFC0(p, ptr, C0_CONTEXT);
845 /* Clear lower 23 bits of context. */
846 uasm_i_dins(p, ptr, 0, 0, 23);
848 /* 1 0 1 0 1 << 6 xkphys cached */
849 uasm_i_ori(p, ptr, ptr, 0x540);
850 uasm_i_drotr(p, ptr, ptr, 11);
851 #elif defined(CONFIG_SMP)
852 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
853 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
854 UASM_i_LA_mostly(p, tmp, pgdc);
855 uasm_i_daddu(p, ptr, ptr, tmp);
856 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
857 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
859 UASM_i_LA_mostly(p, ptr, pgdc);
860 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
864 uasm_l_vmalloc_done(l, *p);
866 /* get pgd offset in bytes */
867 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
869 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
870 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
871 #ifndef __PAGETABLE_PUD_FOLDED
872 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
873 uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
874 uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
875 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
876 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
878 #ifndef __PAGETABLE_PMD_FOLDED
879 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
880 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
881 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
882 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
883 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
886 EXPORT_SYMBOL_GPL(build_get_pmde64);
889 * BVADDR is the faulting address, PTR is scratch.
890 * PTR will hold the pgd for vmalloc.
893 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
894 unsigned int bvaddr, unsigned int ptr,
895 enum vmalloc64_mode mode)
897 long swpd = (long)swapper_pg_dir;
898 int single_insn_swpd;
899 int did_vmalloc_branch = 0;
901 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
903 uasm_l_vmalloc(l, *p);
905 if (mode != not_refill && check_for_high_segbits) {
906 if (single_insn_swpd) {
907 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
908 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
909 did_vmalloc_branch = 1;
912 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
915 if (!did_vmalloc_branch) {
916 if (single_insn_swpd) {
917 uasm_il_b(p, r, label_vmalloc_done);
918 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
920 UASM_i_LA_mostly(p, ptr, swpd);
921 uasm_il_b(p, r, label_vmalloc_done);
922 if (uasm_in_compat_space_p(swpd))
923 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
925 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
928 if (mode != not_refill && check_for_high_segbits) {
929 uasm_l_large_segbits_fault(l, *p);
931 if (mode == refill_scratch && scratch_reg >= 0)
935 * We get here if we are an xsseg address, or if we are
936 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
938 * Ignoring xsseg (assume disabled so would generate
939 * (address errors?), the only remaining possibility
940 * is the upper xuseg addresses. On processors with
941 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
942 * addresses would have taken an address error. We try
943 * to mimic that here by taking a load/istream page
946 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
949 if (mode == refill_scratch) {
950 if (scratch_reg >= 0)
951 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
953 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
960 #else /* !CONFIG_64BIT */
963 * TMP and PTR are scratch.
964 * TMP will be clobbered, PTR will hold the pgd entry.
966 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
969 /* pgd is in pgd_reg */
970 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
971 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
973 long pgdc = (long)pgd_current;
975 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
977 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
978 UASM_i_LA_mostly(p, tmp, pgdc);
979 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
980 uasm_i_addu(p, ptr, tmp, ptr);
982 UASM_i_LA_mostly(p, ptr, pgdc);
984 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
985 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
987 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
988 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
989 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
991 EXPORT_SYMBOL_GPL(build_get_pgde32);
993 #endif /* !CONFIG_64BIT */
995 static void build_adjust_context(u32 **p, unsigned int ctx)
997 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
998 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
1000 switch (current_cpu_type()) {
1017 UASM_i_SRL(p, ctx, ctx, shift);
1018 uasm_i_andi(p, ctx, ctx, mask);
1021 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1024 * Bug workaround for the Nevada. It seems as if under certain
1025 * circumstances the move from cp0_context might produce a
1026 * bogus result when the mfc0 instruction and its consumer are
1027 * in a different cacheline or a load instruction, probably any
1028 * memory reference, is between them.
1030 switch (current_cpu_type()) {
1032 UASM_i_LW(p, ptr, 0, ptr);
1033 GET_CONTEXT(p, tmp); /* get context reg */
1037 GET_CONTEXT(p, tmp); /* get context reg */
1038 UASM_i_LW(p, ptr, 0, ptr);
1042 build_adjust_context(p, tmp);
1043 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1045 EXPORT_SYMBOL_GPL(build_get_ptep);
1047 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1049 int pte_off_even = 0;
1050 int pte_off_odd = sizeof(pte_t);
1052 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1053 /* The low 32 bits of EntryLo is stored in pte_high */
1054 pte_off_even += offsetof(pte_t, pte_high);
1055 pte_off_odd += offsetof(pte_t, pte_high);
1058 if (IS_ENABLED(CONFIG_XPA)) {
1059 uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1060 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1061 UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1063 if (cpu_has_xpa && !mips_xpa_disabled) {
1064 uasm_i_lw(p, tmp, 0, ptep);
1065 uasm_i_ext(p, tmp, tmp, 0, 24);
1066 uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1069 uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1070 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1071 UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1073 if (cpu_has_xpa && !mips_xpa_disabled) {
1074 uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1075 uasm_i_ext(p, tmp, tmp, 0, 24);
1076 uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1081 UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1082 UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1083 if (r45k_bvahwbug())
1084 build_tlb_probe_entry(p);
1085 build_convert_pte_to_entrylo(p, tmp);
1086 if (r4k_250MHZhwbug())
1087 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1088 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1089 build_convert_pte_to_entrylo(p, ptep);
1090 if (r45k_bvahwbug())
1091 uasm_i_mfc0(p, tmp, C0_INDEX);
1092 if (r4k_250MHZhwbug())
1093 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1094 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1096 EXPORT_SYMBOL_GPL(build_update_entries);
1098 struct mips_huge_tlb_info {
1100 int restore_scratch;
1101 bool need_reload_pte;
1104 static struct mips_huge_tlb_info
1105 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1106 struct uasm_reloc **r, unsigned int tmp,
1107 unsigned int ptr, int c0_scratch_reg)
1109 struct mips_huge_tlb_info rv;
1110 unsigned int even, odd;
1111 int vmalloc_branch_delay_filled = 0;
1112 const int scratch = 1; /* Our extra working register */
1114 rv.huge_pte = scratch;
1115 rv.restore_scratch = 0;
1116 rv.need_reload_pte = false;
1118 if (check_for_high_segbits) {
1119 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1122 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1124 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1126 if (c0_scratch_reg >= 0)
1127 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1129 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1131 uasm_i_dsrl_safe(p, scratch, tmp,
1132 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1133 uasm_il_bnez(p, r, scratch, label_vmalloc);
1135 if (pgd_reg == -1) {
1136 vmalloc_branch_delay_filled = 1;
1137 /* Clear lower 23 bits of context. */
1138 uasm_i_dins(p, ptr, 0, 0, 23);
1142 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1144 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1146 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1148 if (c0_scratch_reg >= 0)
1149 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1151 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1154 /* Clear lower 23 bits of context. */
1155 uasm_i_dins(p, ptr, 0, 0, 23);
1157 uasm_il_bltz(p, r, tmp, label_vmalloc);
1160 if (pgd_reg == -1) {
1161 vmalloc_branch_delay_filled = 1;
1162 /* 1 0 1 0 1 << 6 xkphys cached */
1163 uasm_i_ori(p, ptr, ptr, 0x540);
1164 uasm_i_drotr(p, ptr, ptr, 11);
1167 #ifdef __PAGETABLE_PMD_FOLDED
1168 #define LOC_PTEP scratch
1170 #define LOC_PTEP ptr
1173 if (!vmalloc_branch_delay_filled)
1174 /* get pgd offset in bytes */
1175 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1177 uasm_l_vmalloc_done(l, *p);
1181 * fall-through case = badvaddr *pgd_current
1182 * vmalloc case = badvaddr swapper_pg_dir
1185 if (vmalloc_branch_delay_filled)
1186 /* get pgd offset in bytes */
1187 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1189 #ifdef __PAGETABLE_PMD_FOLDED
1190 GET_CONTEXT(p, tmp); /* get context reg */
1192 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1194 if (use_lwx_insns()) {
1195 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1197 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1198 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1201 #ifndef __PAGETABLE_PUD_FOLDED
1202 /* get pud offset in bytes */
1203 uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1204 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1206 if (use_lwx_insns()) {
1207 UASM_i_LWX(p, ptr, scratch, ptr);
1209 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1210 UASM_i_LW(p, ptr, 0, ptr);
1212 /* ptr contains a pointer to PMD entry */
1213 /* tmp contains the address */
1216 #ifndef __PAGETABLE_PMD_FOLDED
1217 /* get pmd offset in bytes */
1218 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1219 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1220 GET_CONTEXT(p, tmp); /* get context reg */
1222 if (use_lwx_insns()) {
1223 UASM_i_LWX(p, scratch, scratch, ptr);
1225 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1226 UASM_i_LW(p, scratch, 0, ptr);
1229 /* Adjust the context during the load latency. */
1230 build_adjust_context(p, tmp);
1232 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1233 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1235 * The in the LWX case we don't want to do the load in the
1236 * delay slot. It cannot issue in the same cycle and may be
1237 * speculative and unneeded.
1239 if (use_lwx_insns())
1241 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1244 /* build_update_entries */
1245 if (use_lwx_insns()) {
1248 UASM_i_LWX(p, even, scratch, tmp);
1249 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1250 UASM_i_LWX(p, odd, scratch, tmp);
1252 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1255 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1256 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1259 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1260 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1261 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1263 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1264 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1265 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1267 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1269 if (c0_scratch_reg >= 0) {
1271 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1272 build_tlb_write_entry(p, l, r, tlb_random);
1273 uasm_l_leave(l, *p);
1274 rv.restore_scratch = 1;
1275 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1276 build_tlb_write_entry(p, l, r, tlb_random);
1277 uasm_l_leave(l, *p);
1278 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1280 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1281 build_tlb_write_entry(p, l, r, tlb_random);
1282 uasm_l_leave(l, *p);
1283 rv.restore_scratch = 1;
1286 uasm_i_eret(p); /* return from trap */
1292 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1293 * because EXL == 0. If we wrap, we can also use the 32 instruction
1294 * slots before the XTLB refill exception handler which belong to the
1295 * unused TLB refill exception.
1297 #define MIPS64_REFILL_INSNS 32
1299 static void build_r4000_tlb_refill_handler(void)
1301 u32 *p = tlb_handler;
1302 struct uasm_label *l = labels;
1303 struct uasm_reloc *r = relocs;
1305 unsigned int final_len;
1306 struct mips_huge_tlb_info htlb_info __maybe_unused;
1307 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1309 memset(tlb_handler, 0, sizeof(tlb_handler));
1310 memset(labels, 0, sizeof(labels));
1311 memset(relocs, 0, sizeof(relocs));
1312 memset(final_handler, 0, sizeof(final_handler));
1314 if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1315 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1317 vmalloc_mode = refill_scratch;
1319 htlb_info.huge_pte = K0;
1320 htlb_info.restore_scratch = 0;
1321 htlb_info.need_reload_pte = true;
1322 vmalloc_mode = refill_noscratch;
1324 * create the plain linear handler
1326 if (bcm1250_m3_war()) {
1327 unsigned int segbits = 44;
1329 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1330 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1331 uasm_i_xor(&p, K0, K0, K1);
1332 uasm_i_dsrl_safe(&p, K1, K0, 62);
1333 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1334 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1335 uasm_i_or(&p, K0, K0, K1);
1336 uasm_il_bnez(&p, &r, K0, label_leave);
1337 /* No need for uasm_i_nop */
1341 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1343 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1346 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1347 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1350 build_get_ptep(&p, K0, K1);
1351 build_update_entries(&p, K0, K1);
1352 build_tlb_write_entry(&p, &l, &r, tlb_random);
1353 uasm_l_leave(&l, p);
1354 uasm_i_eret(&p); /* return from trap */
1356 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1357 uasm_l_tlb_huge_update(&l, p);
1358 if (htlb_info.need_reload_pte)
1359 UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1360 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1361 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1362 htlb_info.restore_scratch);
1366 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1370 * Overflow check: For the 64bit handler, we need at least one
1371 * free instruction slot for the wrap-around branch. In worst
1372 * case, if the intended insertion point is a delay slot, we
1373 * need three, with the second nop'ed and the third being
1376 switch (boot_cpu_type()) {
1378 if (sizeof(long) == 4) {
1380 /* Loongson2 ebase is different than r4k, we have more space */
1381 if ((p - tlb_handler) > 64)
1382 panic("TLB refill handler space exceeded");
1384 * Now fold the handler in the TLB refill handler space.
1387 /* Simplest case, just copy the handler. */
1388 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1389 final_len = p - tlb_handler;
1392 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1393 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1394 && uasm_insn_has_bdelay(relocs,
1395 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1396 panic("TLB refill handler space exceeded");
1398 * Now fold the handler in the TLB refill handler space.
1400 f = final_handler + MIPS64_REFILL_INSNS;
1401 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1402 /* Just copy the handler. */
1403 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1404 final_len = p - tlb_handler;
1406 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1407 const enum label_id ls = label_tlb_huge_update;
1409 const enum label_id ls = label_vmalloc;
1415 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1417 BUG_ON(i == ARRAY_SIZE(labels));
1418 split = labels[i].addr;
1421 * See if we have overflown one way or the other.
1423 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1424 split < p - MIPS64_REFILL_INSNS)
1429 * Split two instructions before the end. One
1430 * for the branch and one for the instruction
1431 * in the delay slot.
1433 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1436 * If the branch would fall in a delay slot,
1437 * we must back up an additional instruction
1438 * so that it is no longer in a delay slot.
1440 if (uasm_insn_has_bdelay(relocs, split - 1))
1443 /* Copy first part of the handler. */
1444 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1445 f += split - tlb_handler;
1448 /* Insert branch. */
1449 uasm_l_split(&l, final_handler);
1450 uasm_il_b(&f, &r, label_split);
1451 if (uasm_insn_has_bdelay(relocs, split))
1454 uasm_copy_handler(relocs, labels,
1455 split, split + 1, f);
1456 uasm_move_labels(labels, f, f + 1, -1);
1462 /* Copy the rest of the handler. */
1463 uasm_copy_handler(relocs, labels, split, p, final_handler);
1464 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1471 uasm_resolve_relocs(relocs, labels);
1472 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1475 memcpy((void *)ebase, final_handler, 0x100);
1476 local_flush_icache_range(ebase, ebase + 0x100);
1477 dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1480 static void setup_pw(void)
1483 unsigned long pgd_i, pgd_w;
1484 #ifndef __PAGETABLE_PMD_FOLDED
1485 unsigned long pmd_i, pmd_w;
1487 unsigned long pt_i, pt_w;
1488 unsigned long pte_i, pte_w;
1489 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1492 psn = ilog2(_PAGE_HUGE); /* bit used to indicate huge page */
1494 pgd_i = PGDIR_SHIFT; /* 1st level PGD */
1495 #ifndef __PAGETABLE_PMD_FOLDED
1496 pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1498 pmd_i = PMD_SHIFT; /* 2nd level PMD */
1499 pmd_w = PMD_SHIFT - PAGE_SHIFT;
1501 pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1504 pt_i = PAGE_SHIFT; /* 3rd level PTE */
1505 pt_w = PAGE_SHIFT - 3;
1507 pte_i = ilog2(_PAGE_GLOBAL);
1509 pwctl = 1 << 30; /* Set PWDirExt */
1511 #ifndef __PAGETABLE_PMD_FOLDED
1512 write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1513 write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1515 write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1516 write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1519 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1520 pwctl |= (1 << 6 | psn);
1522 write_c0_pwctl(pwctl);
1523 write_c0_kpgd((long)swapper_pg_dir);
1524 kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1527 static void build_loongson3_tlb_refill_handler(void)
1529 u32 *p = tlb_handler;
1530 struct uasm_label *l = labels;
1531 struct uasm_reloc *r = relocs;
1533 memset(labels, 0, sizeof(labels));
1534 memset(relocs, 0, sizeof(relocs));
1535 memset(tlb_handler, 0, sizeof(tlb_handler));
1537 if (check_for_high_segbits) {
1538 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1539 uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1540 uasm_il_beqz(&p, &r, K1, label_vmalloc);
1543 uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1545 uasm_l_vmalloc(&l, p);
1548 uasm_i_dmfc0(&p, K1, C0_PGD);
1550 uasm_i_lddir(&p, K0, K1, 3); /* global page dir */
1551 #ifndef __PAGETABLE_PMD_FOLDED
1552 uasm_i_lddir(&p, K1, K0, 1); /* middle page dir */
1554 uasm_i_ldpte(&p, K1, 0); /* even */
1555 uasm_i_ldpte(&p, K1, 1); /* odd */
1558 /* restore page mask */
1559 if (PM_DEFAULT_MASK >> 16) {
1560 uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1561 uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1562 uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1563 } else if (PM_DEFAULT_MASK) {
1564 uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1565 uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1567 uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1572 if (check_for_high_segbits) {
1573 uasm_l_large_segbits_fault(&l, p);
1574 UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1579 uasm_resolve_relocs(relocs, labels);
1580 memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1581 local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1582 dump_handler("loongson3_tlb_refill",
1583 (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1586 static void build_setup_pgd(void)
1589 const int __maybe_unused a1 = 5;
1590 const int __maybe_unused a2 = 6;
1591 u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1592 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1593 long pgdc = (long)pgd_current;
1596 memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1597 memset(labels, 0, sizeof(labels));
1598 memset(relocs, 0, sizeof(relocs));
1599 pgd_reg = allocate_kscratch();
1600 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1601 if (pgd_reg == -1) {
1602 struct uasm_label *l = labels;
1603 struct uasm_reloc *r = relocs;
1605 /* PGD << 11 in c0_Context */
1607 * If it is a ckseg0 address, convert to a physical
1608 * address. Shifting right by 29 and adding 4 will
1609 * result in zero for these addresses.
1612 UASM_i_SRA(&p, a1, a0, 29);
1613 UASM_i_ADDIU(&p, a1, a1, 4);
1614 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1616 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1617 uasm_l_tlbl_goaround1(&l, p);
1618 UASM_i_SLL(&p, a0, a0, 11);
1619 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1623 /* PGD in c0_KScratch */
1625 UASM_i_MTC0(&p, a0, C0_PWBASE);
1627 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1633 /* Save PGD to pgd_current[smp_processor_id()] */
1634 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1635 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1636 UASM_i_LA_mostly(&p, a2, pgdc);
1637 UASM_i_ADDU(&p, a2, a2, a1);
1638 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1640 UASM_i_LA_mostly(&p, a2, pgdc);
1641 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1644 /* if pgd_reg is allocated, save PGD also to scratch register */
1645 if (pgd_reg != -1) {
1646 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1654 if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1655 panic("tlbmiss_handler_setup_pgd space exceeded");
1657 uasm_resolve_relocs(relocs, labels);
1658 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1659 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1661 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1662 tlbmiss_handler_setup_pgd_end);
1666 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1669 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1671 uasm_i_lld(p, pte, 0, ptr);
1674 UASM_i_LL(p, pte, 0, ptr);
1676 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1678 uasm_i_ld(p, pte, 0, ptr);
1681 UASM_i_LW(p, pte, 0, ptr);
1686 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1687 unsigned int mode, unsigned int scratch)
1689 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1690 unsigned int swmode = mode & ~hwmode;
1692 if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1693 uasm_i_lui(p, scratch, swmode >> 16);
1694 uasm_i_or(p, pte, pte, scratch);
1695 BUG_ON(swmode & 0xffff);
1697 uasm_i_ori(p, pte, pte, mode);
1701 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1703 uasm_i_scd(p, pte, 0, ptr);
1706 UASM_i_SC(p, pte, 0, ptr);
1708 if (r10000_llsc_war())
1709 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1711 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1713 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1714 if (!cpu_has_64bits) {
1715 /* no uasm_i_nop needed */
1716 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1717 uasm_i_ori(p, pte, pte, hwmode);
1718 BUG_ON(hwmode & ~0xffff);
1719 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1720 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1721 /* no uasm_i_nop needed */
1722 uasm_i_lw(p, pte, 0, ptr);
1729 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1731 uasm_i_sd(p, pte, 0, ptr);
1734 UASM_i_SW(p, pte, 0, ptr);
1736 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1737 if (!cpu_has_64bits) {
1738 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1739 uasm_i_ori(p, pte, pte, hwmode);
1740 BUG_ON(hwmode & ~0xffff);
1741 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1742 uasm_i_lw(p, pte, 0, ptr);
1749 * Check if PTE is present, if not then jump to LABEL. PTR points to
1750 * the page table where this PTE is located, PTE will be re-loaded
1751 * with it's original value.
1754 build_pte_present(u32 **p, struct uasm_reloc **r,
1755 int pte, int ptr, int scratch, enum label_id lid)
1757 int t = scratch >= 0 ? scratch : pte;
1761 if (use_bbit_insns()) {
1762 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1765 if (_PAGE_PRESENT_SHIFT) {
1766 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1769 uasm_i_andi(p, t, cur, 1);
1770 uasm_il_beqz(p, r, t, lid);
1772 /* You lose the SMP race :-(*/
1773 iPTE_LW(p, pte, ptr);
1776 if (_PAGE_PRESENT_SHIFT) {
1777 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1780 uasm_i_andi(p, t, cur,
1781 (_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1782 uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1783 uasm_il_bnez(p, r, t, lid);
1785 /* You lose the SMP race :-(*/
1786 iPTE_LW(p, pte, ptr);
1790 /* Make PTE valid, store result in PTR. */
1792 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1793 unsigned int ptr, unsigned int scratch)
1795 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1797 iPTE_SW(p, r, pte, ptr, mode, scratch);
1801 * Check if PTE can be written to, if not branch to LABEL. Regardless
1802 * restore PTE with value from PTR when done.
1805 build_pte_writable(u32 **p, struct uasm_reloc **r,
1806 unsigned int pte, unsigned int ptr, int scratch,
1809 int t = scratch >= 0 ? scratch : pte;
1812 if (_PAGE_PRESENT_SHIFT) {
1813 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1816 uasm_i_andi(p, t, cur,
1817 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1818 uasm_i_xori(p, t, t,
1819 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1820 uasm_il_bnez(p, r, t, lid);
1822 /* You lose the SMP race :-(*/
1823 iPTE_LW(p, pte, ptr);
1828 /* Make PTE writable, update software status bits as well, then store
1832 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1833 unsigned int ptr, unsigned int scratch)
1835 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1838 iPTE_SW(p, r, pte, ptr, mode, scratch);
1842 * Check if PTE can be modified, if not branch to LABEL. Regardless
1843 * restore PTE with value from PTR when done.
1846 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1847 unsigned int pte, unsigned int ptr, int scratch,
1850 if (use_bbit_insns()) {
1851 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1854 int t = scratch >= 0 ? scratch : pte;
1855 uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1856 uasm_i_andi(p, t, t, 1);
1857 uasm_il_beqz(p, r, t, lid);
1859 /* You lose the SMP race :-(*/
1860 iPTE_LW(p, pte, ptr);
1864 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1868 * R3000 style TLB load/store/modify handlers.
1872 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1876 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1878 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1879 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1882 uasm_i_rfe(p); /* branch delay */
1886 * This places the pte into ENTRYLO0 and writes it with tlbwi
1887 * or tlbwr as appropriate. This is because the index register
1888 * may have the probe fail bit set as a result of a trap on a
1889 * kseg2 access, i.e. without refill. Then it returns.
1892 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1893 struct uasm_reloc **r, unsigned int pte,
1896 uasm_i_mfc0(p, tmp, C0_INDEX);
1897 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1898 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1899 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1900 uasm_i_tlbwi(p); /* cp0 delay */
1902 uasm_i_rfe(p); /* branch delay */
1903 uasm_l_r3000_write_probe_fail(l, *p);
1904 uasm_i_tlbwr(p); /* cp0 delay */
1906 uasm_i_rfe(p); /* branch delay */
1910 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1913 long pgdc = (long)pgd_current;
1915 uasm_i_mfc0(p, pte, C0_BADVADDR);
1916 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1917 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1918 uasm_i_srl(p, pte, pte, 22); /* load delay */
1919 uasm_i_sll(p, pte, pte, 2);
1920 uasm_i_addu(p, ptr, ptr, pte);
1921 uasm_i_mfc0(p, pte, C0_CONTEXT);
1922 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1923 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1924 uasm_i_addu(p, ptr, ptr, pte);
1925 uasm_i_lw(p, pte, 0, ptr);
1926 uasm_i_tlbp(p); /* load delay */
1929 static void build_r3000_tlb_load_handler(void)
1931 u32 *p = (u32 *)handle_tlbl;
1932 struct uasm_label *l = labels;
1933 struct uasm_reloc *r = relocs;
1935 memset(p, 0, handle_tlbl_end - (char *)p);
1936 memset(labels, 0, sizeof(labels));
1937 memset(relocs, 0, sizeof(relocs));
1939 build_r3000_tlbchange_handler_head(&p, K0, K1);
1940 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1941 uasm_i_nop(&p); /* load delay */
1942 build_make_valid(&p, &r, K0, K1, -1);
1943 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1945 uasm_l_nopage_tlbl(&l, p);
1946 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1949 if (p >= (u32 *)handle_tlbl_end)
1950 panic("TLB load handler fastpath space exceeded");
1952 uasm_resolve_relocs(relocs, labels);
1953 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1954 (unsigned int)(p - (u32 *)handle_tlbl));
1956 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1959 static void build_r3000_tlb_store_handler(void)
1961 u32 *p = (u32 *)handle_tlbs;
1962 struct uasm_label *l = labels;
1963 struct uasm_reloc *r = relocs;
1965 memset(p, 0, handle_tlbs_end - (char *)p);
1966 memset(labels, 0, sizeof(labels));
1967 memset(relocs, 0, sizeof(relocs));
1969 build_r3000_tlbchange_handler_head(&p, K0, K1);
1970 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1971 uasm_i_nop(&p); /* load delay */
1972 build_make_write(&p, &r, K0, K1, -1);
1973 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1975 uasm_l_nopage_tlbs(&l, p);
1976 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1979 if (p >= (u32 *)handle_tlbs_end)
1980 panic("TLB store handler fastpath space exceeded");
1982 uasm_resolve_relocs(relocs, labels);
1983 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1984 (unsigned int)(p - (u32 *)handle_tlbs));
1986 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1989 static void build_r3000_tlb_modify_handler(void)
1991 u32 *p = (u32 *)handle_tlbm;
1992 struct uasm_label *l = labels;
1993 struct uasm_reloc *r = relocs;
1995 memset(p, 0, handle_tlbm_end - (char *)p);
1996 memset(labels, 0, sizeof(labels));
1997 memset(relocs, 0, sizeof(relocs));
1999 build_r3000_tlbchange_handler_head(&p, K0, K1);
2000 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
2001 uasm_i_nop(&p); /* load delay */
2002 build_make_write(&p, &r, K0, K1, -1);
2003 build_r3000_pte_reload_tlbwi(&p, K0, K1);
2005 uasm_l_nopage_tlbm(&l, p);
2006 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2009 if (p >= (u32 *)handle_tlbm_end)
2010 panic("TLB modify handler fastpath space exceeded");
2012 uasm_resolve_relocs(relocs, labels);
2013 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2014 (unsigned int)(p - (u32 *)handle_tlbm));
2016 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
2018 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2020 static bool cpu_has_tlbex_tlbp_race(void)
2023 * When a Hardware Table Walker is running it can replace TLB entries
2024 * at any time, leading to a race between it & the CPU.
2030 * If the CPU shares FTLB RAM with its siblings then our entry may be
2031 * replaced at any time by a sibling performing a write to the FTLB.
2033 if (cpu_has_shared_ftlb_ram)
2036 /* In all other cases there ought to be no race condition to handle */
2041 * R4000 style TLB load/store/modify handlers.
2043 static struct work_registers
2044 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2045 struct uasm_reloc **r)
2047 struct work_registers wr = build_get_work_registers(p);
2050 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2052 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2055 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2057 * For huge tlb entries, pmd doesn't contain an address but
2058 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2059 * see if we need to jump to huge tlb processing.
2061 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2064 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2065 UASM_i_LW(p, wr.r2, 0, wr.r2);
2066 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2067 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2068 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2071 uasm_l_smp_pgtable_change(l, *p);
2073 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2074 if (!m4kc_tlbp_war()) {
2075 build_tlb_probe_entry(p);
2076 if (cpu_has_tlbex_tlbp_race()) {
2077 /* race condition happens, leaving */
2079 uasm_i_mfc0(p, wr.r3, C0_INDEX);
2080 uasm_il_bltz(p, r, wr.r3, label_leave);
2088 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2089 struct uasm_reloc **r, unsigned int tmp,
2092 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2093 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2094 build_update_entries(p, tmp, ptr);
2095 build_tlb_write_entry(p, l, r, tlb_indexed);
2096 uasm_l_leave(l, *p);
2097 build_restore_work_registers(p);
2098 uasm_i_eret(p); /* return from trap */
2101 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2105 static void build_r4000_tlb_load_handler(void)
2107 u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2108 struct uasm_label *l = labels;
2109 struct uasm_reloc *r = relocs;
2110 struct work_registers wr;
2112 memset(p, 0, handle_tlbl_end - (char *)p);
2113 memset(labels, 0, sizeof(labels));
2114 memset(relocs, 0, sizeof(relocs));
2116 if (bcm1250_m3_war()) {
2117 unsigned int segbits = 44;
2119 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2120 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2121 uasm_i_xor(&p, K0, K0, K1);
2122 uasm_i_dsrl_safe(&p, K1, K0, 62);
2123 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2124 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2125 uasm_i_or(&p, K0, K0, K1);
2126 uasm_il_bnez(&p, &r, K0, label_leave);
2127 /* No need for uasm_i_nop */
2130 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2131 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2132 if (m4kc_tlbp_war())
2133 build_tlb_probe_entry(&p);
2135 if (cpu_has_rixi && !cpu_has_rixiex) {
2137 * If the page is not _PAGE_VALID, RI or XI could not
2138 * have triggered it. Skip the expensive test..
2140 if (use_bbit_insns()) {
2141 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2142 label_tlbl_goaround1);
2144 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2145 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2150 * Warn if something may race with us & replace the TLB entry
2151 * before we read it here. Everything with such races should
2152 * also have dedicated RiXi exception handlers, so this
2155 WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2159 switch (current_cpu_type()) {
2161 if (cpu_has_mips_r2_exec_hazard) {
2164 case CPU_CAVIUM_OCTEON:
2165 case CPU_CAVIUM_OCTEON_PLUS:
2166 case CPU_CAVIUM_OCTEON2:
2171 /* Examine entrylo 0 or 1 based on ptr. */
2172 if (use_bbit_insns()) {
2173 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2175 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2176 uasm_i_beqz(&p, wr.r3, 8);
2178 /* load it in the delay slot*/
2179 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2180 /* load it if ptr is odd */
2181 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2183 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2184 * XI must have triggered it.
2186 if (use_bbit_insns()) {
2187 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2189 uasm_l_tlbl_goaround1(&l, p);
2191 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2192 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2195 uasm_l_tlbl_goaround1(&l, p);
2197 build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2198 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2200 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2202 * This is the entry point when build_r4000_tlbchange_handler_head
2203 * spots a huge page.
2205 uasm_l_tlb_huge_update(&l, p);
2206 iPTE_LW(&p, wr.r1, wr.r2);
2207 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2208 build_tlb_probe_entry(&p);
2210 if (cpu_has_rixi && !cpu_has_rixiex) {
2212 * If the page is not _PAGE_VALID, RI or XI could not
2213 * have triggered it. Skip the expensive test..
2215 if (use_bbit_insns()) {
2216 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2217 label_tlbl_goaround2);
2219 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2220 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2225 * Warn if something may race with us & replace the TLB entry
2226 * before we read it here. Everything with such races should
2227 * also have dedicated RiXi exception handlers, so this
2230 WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2234 switch (current_cpu_type()) {
2236 if (cpu_has_mips_r2_exec_hazard) {
2239 case CPU_CAVIUM_OCTEON:
2240 case CPU_CAVIUM_OCTEON_PLUS:
2241 case CPU_CAVIUM_OCTEON2:
2246 /* Examine entrylo 0 or 1 based on ptr. */
2247 if (use_bbit_insns()) {
2248 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2250 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2251 uasm_i_beqz(&p, wr.r3, 8);
2253 /* load it in the delay slot*/
2254 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2255 /* load it if ptr is odd */
2256 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2258 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2259 * XI must have triggered it.
2261 if (use_bbit_insns()) {
2262 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2264 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2265 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2267 if (PM_DEFAULT_MASK == 0)
2270 * We clobbered C0_PAGEMASK, restore it. On the other branch
2271 * it is restored in build_huge_tlb_write_entry.
2273 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2275 uasm_l_tlbl_goaround2(&l, p);
2277 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2278 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2281 uasm_l_nopage_tlbl(&l, p);
2282 build_restore_work_registers(&p);
2283 #ifdef CONFIG_CPU_MICROMIPS
2284 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2285 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2286 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2290 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2293 if (p >= (u32 *)handle_tlbl_end)
2294 panic("TLB load handler fastpath space exceeded");
2296 uasm_resolve_relocs(relocs, labels);
2297 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2298 (unsigned int)(p - (u32 *)handle_tlbl));
2300 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2303 static void build_r4000_tlb_store_handler(void)
2305 u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2306 struct uasm_label *l = labels;
2307 struct uasm_reloc *r = relocs;
2308 struct work_registers wr;
2310 memset(p, 0, handle_tlbs_end - (char *)p);
2311 memset(labels, 0, sizeof(labels));
2312 memset(relocs, 0, sizeof(relocs));
2314 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2315 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2316 if (m4kc_tlbp_war())
2317 build_tlb_probe_entry(&p);
2318 build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2319 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2321 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2323 * This is the entry point when
2324 * build_r4000_tlbchange_handler_head spots a huge page.
2326 uasm_l_tlb_huge_update(&l, p);
2327 iPTE_LW(&p, wr.r1, wr.r2);
2328 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2329 build_tlb_probe_entry(&p);
2330 uasm_i_ori(&p, wr.r1, wr.r1,
2331 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2332 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2335 uasm_l_nopage_tlbs(&l, p);
2336 build_restore_work_registers(&p);
2337 #ifdef CONFIG_CPU_MICROMIPS
2338 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2339 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2340 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2344 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2347 if (p >= (u32 *)handle_tlbs_end)
2348 panic("TLB store handler fastpath space exceeded");
2350 uasm_resolve_relocs(relocs, labels);
2351 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2352 (unsigned int)(p - (u32 *)handle_tlbs));
2354 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2357 static void build_r4000_tlb_modify_handler(void)
2359 u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2360 struct uasm_label *l = labels;
2361 struct uasm_reloc *r = relocs;
2362 struct work_registers wr;
2364 memset(p, 0, handle_tlbm_end - (char *)p);
2365 memset(labels, 0, sizeof(labels));
2366 memset(relocs, 0, sizeof(relocs));
2368 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2369 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2370 if (m4kc_tlbp_war())
2371 build_tlb_probe_entry(&p);
2372 /* Present and writable bits set, set accessed and dirty bits. */
2373 build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2374 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2376 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2378 * This is the entry point when
2379 * build_r4000_tlbchange_handler_head spots a huge page.
2381 uasm_l_tlb_huge_update(&l, p);
2382 iPTE_LW(&p, wr.r1, wr.r2);
2383 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2384 build_tlb_probe_entry(&p);
2385 uasm_i_ori(&p, wr.r1, wr.r1,
2386 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2387 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2390 uasm_l_nopage_tlbm(&l, p);
2391 build_restore_work_registers(&p);
2392 #ifdef CONFIG_CPU_MICROMIPS
2393 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2394 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2395 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2399 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2402 if (p >= (u32 *)handle_tlbm_end)
2403 panic("TLB modify handler fastpath space exceeded");
2405 uasm_resolve_relocs(relocs, labels);
2406 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2407 (unsigned int)(p - (u32 *)handle_tlbm));
2409 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2412 static void flush_tlb_handlers(void)
2414 local_flush_icache_range((unsigned long)handle_tlbl,
2415 (unsigned long)handle_tlbl_end);
2416 local_flush_icache_range((unsigned long)handle_tlbs,
2417 (unsigned long)handle_tlbs_end);
2418 local_flush_icache_range((unsigned long)handle_tlbm,
2419 (unsigned long)handle_tlbm_end);
2420 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2421 (unsigned long)tlbmiss_handler_setup_pgd_end);
2424 static void print_htw_config(void)
2426 unsigned long config;
2428 const int field = 2 * sizeof(unsigned long);
2430 config = read_c0_pwfield();
2431 pr_debug("PWField (0x%0*lx): GDI: 0x%02lx UDI: 0x%02lx MDI: 0x%02lx PTI: 0x%02lx PTEI: 0x%02lx\n",
2433 (config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2434 (config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2435 (config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2436 (config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2437 (config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2439 config = read_c0_pwsize();
2440 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",
2442 (config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2443 (config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2444 (config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2445 (config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2446 (config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2447 (config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2449 pwctl = read_c0_pwctl();
2450 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",
2452 (pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2453 (pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2454 (pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2455 (pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2456 (pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2457 (pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2458 (pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2461 static void config_htw_params(void)
2463 unsigned long pwfield, pwsize, ptei;
2464 unsigned int config;
2467 * We are using 2-level page tables, so we only need to
2468 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2469 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2470 * write values less than 0xc in these fields because the entire
2471 * write will be dropped. As a result of which, we must preserve
2472 * the original reset values and overwrite only what we really want.
2475 pwfield = read_c0_pwfield();
2476 /* re-initialize the GDI field */
2477 pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2478 pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2479 /* re-initialize the PTI field including the even/odd bit */
2480 pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2481 pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2482 if (CONFIG_PGTABLE_LEVELS >= 3) {
2483 pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2484 pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2486 /* Set the PTEI right shift */
2487 ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2489 write_c0_pwfield(pwfield);
2490 /* Check whether the PTEI value is supported */
2491 back_to_back_c0_hazard();
2492 pwfield = read_c0_pwfield();
2493 if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2495 pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2498 * Drop option to avoid HTW being enabled via another path
2501 current_cpu_data.options &= ~MIPS_CPU_HTW;
2505 pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2506 pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2507 if (CONFIG_PGTABLE_LEVELS >= 3)
2508 pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2510 /* Set pointer size to size of directory pointers */
2511 if (IS_ENABLED(CONFIG_64BIT))
2512 pwsize |= MIPS_PWSIZE_PS_MASK;
2513 /* PTEs may be multiple pointers long (e.g. with XPA) */
2514 pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2515 & MIPS_PWSIZE_PTEW_MASK;
2517 write_c0_pwsize(pwsize);
2519 /* Make sure everything is set before we enable the HTW */
2520 back_to_back_c0_hazard();
2523 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2526 config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2527 if (IS_ENABLED(CONFIG_64BIT))
2528 config |= MIPS_PWCTL_XU_MASK;
2529 write_c0_pwctl(config);
2530 pr_info("Hardware Page Table Walker enabled\n");
2535 static void config_xpa_params(void)
2538 unsigned int pagegrain;
2540 if (mips_xpa_disabled) {
2541 pr_info("Extended Physical Addressing (XPA) disabled\n");
2545 pagegrain = read_c0_pagegrain();
2546 write_c0_pagegrain(pagegrain | PG_ELPA);
2547 back_to_back_c0_hazard();
2548 pagegrain = read_c0_pagegrain();
2550 if (pagegrain & PG_ELPA)
2551 pr_info("Extended Physical Addressing (XPA) enabled\n");
2553 panic("Extended Physical Addressing (XPA) disabled");
2557 static void check_pabits(void)
2559 unsigned long entry;
2560 unsigned pabits, fillbits;
2562 if (!cpu_has_rixi || _PAGE_NO_EXEC == 0) {
2564 * We'll only be making use of the fact that we can rotate bits
2565 * into the fill if the CPU supports RIXI, so don't bother
2566 * probing this for CPUs which don't.
2571 write_c0_entrylo0(~0ul);
2572 back_to_back_c0_hazard();
2573 entry = read_c0_entrylo0();
2575 /* clear all non-PFN bits */
2576 entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2577 entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2579 /* find a lower bound on PABITS, and upper bound on fill bits */
2580 pabits = fls_long(entry) + 6;
2581 fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2583 /* minus the RI & XI bits */
2584 fillbits -= min_t(unsigned, fillbits, 2);
2586 if (fillbits >= ilog2(_PAGE_NO_EXEC))
2587 fill_includes_sw_bits = true;
2589 pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2592 void build_tlb_refill_handler(void)
2595 * The refill handler is generated per-CPU, multi-node systems
2596 * may have local storage for it. The other handlers are only
2599 static int run_once = 0;
2601 if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2602 panic("Kernels supporting XPA currently require CPUs with RIXI");
2604 output_pgtable_bits_defines();
2608 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2611 switch (current_cpu_type()) {
2619 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2620 if (cpu_has_local_ebase)
2621 build_r3000_tlb_refill_handler();
2623 if (!cpu_has_local_ebase)
2624 build_r3000_tlb_refill_handler();
2626 build_r3000_tlb_load_handler();
2627 build_r3000_tlb_store_handler();
2628 build_r3000_tlb_modify_handler();
2629 flush_tlb_handlers();
2633 panic("No R3000 TLB refill handler");
2638 panic("No R8000 TLB refill handler yet");
2646 scratch_reg = allocate_kscratch();
2648 build_r4000_tlb_load_handler();
2649 build_r4000_tlb_store_handler();
2650 build_r4000_tlb_modify_handler();
2652 build_loongson3_tlb_refill_handler();
2653 else if (!cpu_has_local_ebase)
2654 build_r4000_tlb_refill_handler();
2655 flush_tlb_handlers();
2658 if (cpu_has_local_ebase)
2659 build_r4000_tlb_refill_handler();
2661 config_xpa_params();
2663 config_htw_params();