2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Synthesize TLB refill handlers at runtime.
8 * Copyright (C) 2004, 2005, 2006, 2008 Thiemo Seufer
9 * Copyright (C) 2005, 2007, 2008, 2009 Maciej W. Rozycki
10 * Copyright (C) 2006 Ralf Baechle (ralf@linux-mips.org)
11 * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12 * Copyright (C) 2011 MIPS Technologies, Inc.
14 * ... and the days got worse and worse and now you see
15 * I've gone completly out of my mind.
17 * They're coming to take me a away haha
18 * they're coming to take me a away hoho hihi haha
19 * to the funny farm where code is beautiful all the time ...
21 * (Condolences to Napoleon XIV)
24 #include <linux/bug.h>
25 #include <linux/kernel.h>
26 #include <linux/types.h>
27 #include <linux/smp.h>
28 #include <linux/string.h>
29 #include <linux/cache.h>
31 #include <asm/cacheflush.h>
32 #include <asm/cpu-type.h>
33 #include <asm/pgtable.h>
36 #include <asm/setup.h>
39 * TLB load/store/modify handlers.
41 * Only the fastpath gets synthesized at runtime, the slowpath for
42 * do_page_fault remains normal asm.
44 extern void tlb_do_page_fault_0(void);
45 extern void tlb_do_page_fault_1(void);
47 struct work_registers {
56 } ____cacheline_aligned_in_smp;
58 static struct tlb_reg_save handler_reg_save[NR_CPUS];
60 static inline int r45k_bvahwbug(void)
62 /* XXX: We should probe for the presence of this bug, but we don't. */
66 static inline int r4k_250MHZhwbug(void)
68 /* XXX: We should probe for the presence of this bug, but we don't. */
72 static inline int __maybe_unused bcm1250_m3_war(void)
74 return BCM1250_M3_WAR;
77 static inline int __maybe_unused r10000_llsc_war(void)
79 return R10000_LLSC_WAR;
82 static int use_bbit_insns(void)
84 switch (current_cpu_type()) {
85 case CPU_CAVIUM_OCTEON:
86 case CPU_CAVIUM_OCTEON_PLUS:
87 case CPU_CAVIUM_OCTEON2:
88 case CPU_CAVIUM_OCTEON3:
95 static int use_lwx_insns(void)
97 switch (current_cpu_type()) {
98 case CPU_CAVIUM_OCTEON2:
99 case CPU_CAVIUM_OCTEON3:
105 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
106 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
107 static bool scratchpad_available(void)
111 static int scratchpad_offset(int i)
114 * CVMSEG starts at address -32768 and extends for
115 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
117 i += 1; /* Kernel use starts at the top and works down. */
118 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
121 static bool scratchpad_available(void)
125 static int scratchpad_offset(int i)
128 /* Really unreachable, but evidently some GCC want this. */
133 * Found by experiment: At least some revisions of the 4kc throw under
134 * some circumstances a machine check exception, triggered by invalid
135 * values in the index register. Delaying the tlbp instruction until
136 * after the next branch, plus adding an additional nop in front of
137 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
138 * why; it's not an issue caused by the core RTL.
141 static int m4kc_tlbp_war(void)
143 return (current_cpu_data.processor_id & 0xffff00) ==
144 (PRID_COMP_MIPS | PRID_IMP_4KC);
147 /* Handle labels (which must be positive integers). */
149 label_second_part = 1,
154 label_split = label_tlbw_hazard_0 + 8,
155 label_tlbl_goaround1,
156 label_tlbl_goaround2,
160 label_smp_pgtable_change,
161 label_r3000_write_probe_fail,
162 label_large_segbits_fault,
163 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
164 label_tlb_huge_update,
168 UASM_L_LA(_second_part)
171 UASM_L_LA(_vmalloc_done)
172 /* _tlbw_hazard_x is handled differently. */
174 UASM_L_LA(_tlbl_goaround1)
175 UASM_L_LA(_tlbl_goaround2)
176 UASM_L_LA(_nopage_tlbl)
177 UASM_L_LA(_nopage_tlbs)
178 UASM_L_LA(_nopage_tlbm)
179 UASM_L_LA(_smp_pgtable_change)
180 UASM_L_LA(_r3000_write_probe_fail)
181 UASM_L_LA(_large_segbits_fault)
182 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
183 UASM_L_LA(_tlb_huge_update)
186 static int hazard_instance;
188 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
192 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
199 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
203 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
211 * pgtable bits are assigned dynamically depending on processor feature
212 * and statically based on kernel configuration. This spits out the actual
213 * values the kernel is using. Required to make sense from disassembled
214 * TLB exception handlers.
216 static void output_pgtable_bits_defines(void)
218 #define pr_define(fmt, ...) \
219 pr_debug("#define " fmt, ##__VA_ARGS__)
221 pr_debug("#include <asm/asm.h>\n");
222 pr_debug("#include <asm/regdef.h>\n");
225 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
226 pr_define("_PAGE_READ_SHIFT %d\n", _PAGE_READ_SHIFT);
227 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
228 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
229 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
230 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
231 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
232 pr_define("_PAGE_SPLITTING_SHIFT %d\n", _PAGE_SPLITTING_SHIFT);
235 #ifdef _PAGE_NO_EXEC_SHIFT
236 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
238 #ifdef _PAGE_NO_READ_SHIFT
239 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
242 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
243 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
244 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
245 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
249 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
253 pr_debug("LEAF(%s)\n", symbol);
255 pr_debug("\t.set push\n");
256 pr_debug("\t.set noreorder\n");
258 for (i = 0; i < count; i++)
259 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
261 pr_debug("\t.set\tpop\n");
263 pr_debug("\tEND(%s)\n", symbol);
266 /* The only general purpose registers allowed in TLB handlers. */
270 /* Some CP0 registers */
271 #define C0_INDEX 0, 0
272 #define C0_ENTRYLO0 2, 0
273 #define C0_TCBIND 2, 2
274 #define C0_ENTRYLO1 3, 0
275 #define C0_CONTEXT 4, 0
276 #define C0_PAGEMASK 5, 0
277 #define C0_BADVADDR 8, 0
278 #define C0_ENTRYHI 10, 0
280 #define C0_XCONTEXT 20, 0
283 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
285 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
288 /* The worst case length of the handler is around 18 instructions for
289 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
290 * Maximum space available is 32 instructions for R3000 and 64
291 * instructions for R4000.
293 * We deliberately chose a buffer size of 128, so we won't scribble
294 * over anything important on overflow before we panic.
296 static u32 tlb_handler[128];
298 /* simply assume worst case size for labels and relocs */
299 static struct uasm_label labels[128];
300 static struct uasm_reloc relocs[128];
302 static int check_for_high_segbits;
304 static unsigned int kscratch_used_mask;
306 static inline int __maybe_unused c0_kscratch(void)
308 switch (current_cpu_type()) {
317 static int allocate_kscratch(void)
320 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
327 r--; /* make it zero based */
329 kscratch_used_mask |= (1 << r);
334 static int scratch_reg;
336 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
338 static struct work_registers build_get_work_registers(u32 **p)
340 struct work_registers r;
342 if (scratch_reg >= 0) {
343 /* Save in CPU local C0_KScratch? */
344 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
351 if (num_possible_cpus() > 1) {
352 /* Get smp_processor_id */
353 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
354 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
356 /* handler_reg_save index in K0 */
357 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
359 UASM_i_LA(p, K1, (long)&handler_reg_save);
360 UASM_i_ADDU(p, K0, K0, K1);
362 UASM_i_LA(p, K0, (long)&handler_reg_save);
364 /* K0 now points to save area, save $1 and $2 */
365 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
366 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
374 static void build_restore_work_registers(u32 **p)
376 if (scratch_reg >= 0) {
377 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
380 /* K0 already points to save area, restore $1 and $2 */
381 UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
382 UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
385 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
388 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
389 * we cannot do r3000 under these circumstances.
391 * Declare pgd_current here instead of including mmu_context.h to avoid type
392 * conflicts for tlbmiss_handler_setup_pgd
394 extern unsigned long pgd_current[];
397 * The R3000 TLB handler is simple.
399 static void build_r3000_tlb_refill_handler(void)
401 long pgdc = (long)pgd_current;
404 memset(tlb_handler, 0, sizeof(tlb_handler));
407 uasm_i_mfc0(&p, K0, C0_BADVADDR);
408 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
409 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
410 uasm_i_srl(&p, K0, K0, 22); /* load delay */
411 uasm_i_sll(&p, K0, K0, 2);
412 uasm_i_addu(&p, K1, K1, K0);
413 uasm_i_mfc0(&p, K0, C0_CONTEXT);
414 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
415 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
416 uasm_i_addu(&p, K1, K1, K0);
417 uasm_i_lw(&p, K0, 0, K1);
418 uasm_i_nop(&p); /* load delay */
419 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
420 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
421 uasm_i_tlbwr(&p); /* cp0 delay */
423 uasm_i_rfe(&p); /* branch delay */
425 if (p > tlb_handler + 32)
426 panic("TLB refill handler space exceeded");
428 pr_debug("Wrote TLB refill handler (%u instructions).\n",
429 (unsigned int)(p - tlb_handler));
431 memcpy((void *)ebase, tlb_handler, 0x80);
432 local_flush_icache_range(ebase, ebase + 0x80);
434 dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
436 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
439 * The R4000 TLB handler is much more complicated. We have two
440 * consecutive handler areas with 32 instructions space each.
441 * Since they aren't used at the same time, we can overflow in the
442 * other one.To keep things simple, we first assume linear space,
443 * then we relocate it to the final handler layout as needed.
445 static u32 final_handler[64];
450 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
451 * 2. A timing hazard exists for the TLBP instruction.
453 * stalling_instruction
456 * The JTLB is being read for the TLBP throughout the stall generated by the
457 * previous instruction. This is not really correct as the stalling instruction
458 * can modify the address used to access the JTLB. The failure symptom is that
459 * the TLBP instruction will use an address created for the stalling instruction
460 * and not the address held in C0_ENHI and thus report the wrong results.
462 * The software work-around is to not allow the instruction preceding the TLBP
463 * to stall - make it an NOP or some other instruction guaranteed not to stall.
465 * Errata 2 will not be fixed. This errata is also on the R5000.
467 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
469 static void __maybe_unused build_tlb_probe_entry(u32 **p)
471 switch (current_cpu_type()) {
472 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
488 * Write random or indexed TLB entry, and care about the hazards from
489 * the preceding mtc0 and for the following eret.
491 enum tlb_write_entry { tlb_random, tlb_indexed };
493 static void build_tlb_write_entry(u32 **p, struct uasm_label **l,
494 struct uasm_reloc **r,
495 enum tlb_write_entry wmode)
497 void(*tlbw)(u32 **) = NULL;
500 case tlb_random: tlbw = uasm_i_tlbwr; break;
501 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
504 if (cpu_has_mips_r2_exec_hazard) {
506 * The architecture spec says an ehb is required here,
507 * but a number of cores do not have the hazard and
508 * using an ehb causes an expensive pipeline stall.
510 switch (current_cpu_type()) {
517 case CPU_QEMU_GENERIC:
528 switch (current_cpu_type()) {
536 * This branch uses up a mtc0 hazard nop slot and saves
537 * two nops after the tlbw instruction.
539 uasm_bgezl_hazard(p, r, hazard_instance);
541 uasm_bgezl_label(l, p, hazard_instance);
555 uasm_i_nop(p); /* QED specifies 2 nops hazard */
556 uasm_i_nop(p); /* QED specifies 2 nops hazard */
629 panic("No TLB refill handler yet (CPU type: %d)",
635 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
639 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
641 #ifdef CONFIG_PHYS_ADDR_T_64BIT
642 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
644 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
649 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
651 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
652 unsigned int tmp, enum label_id lid,
655 if (restore_scratch) {
656 /* Reset default page size */
657 if (PM_DEFAULT_MASK >> 16) {
658 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
659 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
660 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
661 uasm_il_b(p, r, lid);
662 } else if (PM_DEFAULT_MASK) {
663 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
664 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
665 uasm_il_b(p, r, lid);
667 uasm_i_mtc0(p, 0, C0_PAGEMASK);
668 uasm_il_b(p, r, lid);
670 if (scratch_reg >= 0)
671 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
673 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
675 /* Reset default page size */
676 if (PM_DEFAULT_MASK >> 16) {
677 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
678 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
679 uasm_il_b(p, r, lid);
680 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
681 } else if (PM_DEFAULT_MASK) {
682 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
683 uasm_il_b(p, r, lid);
684 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
686 uasm_il_b(p, r, lid);
687 uasm_i_mtc0(p, 0, C0_PAGEMASK);
692 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
693 struct uasm_reloc **r,
695 enum tlb_write_entry wmode,
698 /* Set huge page tlb entry size */
699 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
700 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
701 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
703 build_tlb_write_entry(p, l, r, wmode);
705 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
709 * Check if Huge PTE is present, if so then jump to LABEL.
712 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
713 unsigned int pmd, int lid)
715 UASM_i_LW(p, tmp, 0, pmd);
716 if (use_bbit_insns()) {
717 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
719 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
720 uasm_il_bnez(p, r, tmp, lid);
724 static void build_huge_update_entries(u32 **p, unsigned int pte,
730 * A huge PTE describes an area the size of the
731 * configured huge page size. This is twice the
732 * of the large TLB entry size we intend to use.
733 * A TLB entry half the size of the configured
734 * huge page size is configured into entrylo0
735 * and entrylo1 to cover the contiguous huge PTE
738 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
740 /* We can clobber tmp. It isn't used after this.*/
742 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
744 build_convert_pte_to_entrylo(p, pte);
745 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
746 /* convert to entrylo1 */
748 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
750 UASM_i_ADDU(p, pte, pte, tmp);
752 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
755 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
756 struct uasm_label **l,
761 UASM_i_SC(p, pte, 0, ptr);
762 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
763 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
765 UASM_i_SW(p, pte, 0, ptr);
767 build_huge_update_entries(p, pte, ptr);
768 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
770 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
774 * TMP and PTR are scratch.
775 * TMP will be clobbered, PTR will hold the pmd entry.
778 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
779 unsigned int tmp, unsigned int ptr)
781 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
782 long pgdc = (long)pgd_current;
785 * The vmalloc handling is not in the hotpath.
787 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
789 if (check_for_high_segbits) {
791 * The kernel currently implicitely assumes that the
792 * MIPS SEGBITS parameter for the processor is
793 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
794 * allocate virtual addresses outside the maximum
795 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
796 * that doesn't prevent user code from accessing the
797 * higher xuseg addresses. Here, we make sure that
798 * everything but the lower xuseg addresses goes down
799 * the module_alloc/vmalloc path.
801 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
802 uasm_il_bnez(p, r, ptr, label_vmalloc);
804 uasm_il_bltz(p, r, tmp, label_vmalloc);
806 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
809 /* pgd is in pgd_reg */
810 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
812 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
814 * &pgd << 11 stored in CONTEXT [23..63].
816 UASM_i_MFC0(p, ptr, C0_CONTEXT);
818 /* Clear lower 23 bits of context. */
819 uasm_i_dins(p, ptr, 0, 0, 23);
821 /* 1 0 1 0 1 << 6 xkphys cached */
822 uasm_i_ori(p, ptr, ptr, 0x540);
823 uasm_i_drotr(p, ptr, ptr, 11);
824 #elif defined(CONFIG_SMP)
825 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
826 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
827 UASM_i_LA_mostly(p, tmp, pgdc);
828 uasm_i_daddu(p, ptr, ptr, tmp);
829 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
830 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
832 UASM_i_LA_mostly(p, ptr, pgdc);
833 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
837 uasm_l_vmalloc_done(l, *p);
839 /* get pgd offset in bytes */
840 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
842 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
843 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
844 #ifndef __PAGETABLE_PMD_FOLDED
845 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
846 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
847 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
848 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
849 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
854 * BVADDR is the faulting address, PTR is scratch.
855 * PTR will hold the pgd for vmalloc.
858 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
859 unsigned int bvaddr, unsigned int ptr,
860 enum vmalloc64_mode mode)
862 long swpd = (long)swapper_pg_dir;
863 int single_insn_swpd;
864 int did_vmalloc_branch = 0;
866 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
868 uasm_l_vmalloc(l, *p);
870 if (mode != not_refill && check_for_high_segbits) {
871 if (single_insn_swpd) {
872 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
873 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
874 did_vmalloc_branch = 1;
877 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
880 if (!did_vmalloc_branch) {
881 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
882 uasm_il_b(p, r, label_vmalloc_done);
883 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
885 UASM_i_LA_mostly(p, ptr, swpd);
886 uasm_il_b(p, r, label_vmalloc_done);
887 if (uasm_in_compat_space_p(swpd))
888 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
890 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
893 if (mode != not_refill && check_for_high_segbits) {
894 uasm_l_large_segbits_fault(l, *p);
896 * We get here if we are an xsseg address, or if we are
897 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
899 * Ignoring xsseg (assume disabled so would generate
900 * (address errors?), the only remaining possibility
901 * is the upper xuseg addresses. On processors with
902 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
903 * addresses would have taken an address error. We try
904 * to mimic that here by taking a load/istream page
907 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
910 if (mode == refill_scratch) {
911 if (scratch_reg >= 0)
912 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
914 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
921 #else /* !CONFIG_64BIT */
924 * TMP and PTR are scratch.
925 * TMP will be clobbered, PTR will hold the pgd entry.
927 static void __maybe_unused
928 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
931 /* pgd is in pgd_reg */
932 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
933 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
935 long pgdc = (long)pgd_current;
937 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
939 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
940 UASM_i_LA_mostly(p, tmp, pgdc);
941 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
942 uasm_i_addu(p, ptr, tmp, ptr);
944 UASM_i_LA_mostly(p, ptr, pgdc);
946 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
947 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
949 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
950 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
951 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
954 #endif /* !CONFIG_64BIT */
956 static void build_adjust_context(u32 **p, unsigned int ctx)
958 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
959 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
961 switch (current_cpu_type()) {
978 UASM_i_SRL(p, ctx, ctx, shift);
979 uasm_i_andi(p, ctx, ctx, mask);
982 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
985 * Bug workaround for the Nevada. It seems as if under certain
986 * circumstances the move from cp0_context might produce a
987 * bogus result when the mfc0 instruction and its consumer are
988 * in a different cacheline or a load instruction, probably any
989 * memory reference, is between them.
991 switch (current_cpu_type()) {
993 UASM_i_LW(p, ptr, 0, ptr);
994 GET_CONTEXT(p, tmp); /* get context reg */
998 GET_CONTEXT(p, tmp); /* get context reg */
999 UASM_i_LW(p, ptr, 0, ptr);
1003 build_adjust_context(p, tmp);
1004 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1007 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1010 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1011 * Kernel is a special case. Only a few CPUs use it.
1013 #ifdef CONFIG_PHYS_ADDR_T_64BIT
1014 if (cpu_has_64bits) {
1015 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1016 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1018 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1019 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1020 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1022 uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1023 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1024 uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1026 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1028 int pte_off_even = sizeof(pte_t) / 2;
1029 int pte_off_odd = pte_off_even + sizeof(pte_t);
1031 /* The pte entries are pre-shifted */
1032 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1033 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1034 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1035 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1038 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1039 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1040 if (r45k_bvahwbug())
1041 build_tlb_probe_entry(p);
1043 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1044 if (r4k_250MHZhwbug())
1045 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1046 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1047 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1049 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1050 if (r4k_250MHZhwbug())
1051 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1052 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1053 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1054 if (r45k_bvahwbug())
1055 uasm_i_mfc0(p, tmp, C0_INDEX);
1057 if (r4k_250MHZhwbug())
1058 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1059 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1063 struct mips_huge_tlb_info {
1065 int restore_scratch;
1066 bool need_reload_pte;
1069 static struct mips_huge_tlb_info
1070 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1071 struct uasm_reloc **r, unsigned int tmp,
1072 unsigned int ptr, int c0_scratch_reg)
1074 struct mips_huge_tlb_info rv;
1075 unsigned int even, odd;
1076 int vmalloc_branch_delay_filled = 0;
1077 const int scratch = 1; /* Our extra working register */
1079 rv.huge_pte = scratch;
1080 rv.restore_scratch = 0;
1081 rv.need_reload_pte = false;
1083 if (check_for_high_segbits) {
1084 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1087 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1089 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1091 if (c0_scratch_reg >= 0)
1092 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1094 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1096 uasm_i_dsrl_safe(p, scratch, tmp,
1097 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1098 uasm_il_bnez(p, r, scratch, label_vmalloc);
1100 if (pgd_reg == -1) {
1101 vmalloc_branch_delay_filled = 1;
1102 /* Clear lower 23 bits of context. */
1103 uasm_i_dins(p, ptr, 0, 0, 23);
1107 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1109 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1111 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1113 if (c0_scratch_reg >= 0)
1114 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1116 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1119 /* Clear lower 23 bits of context. */
1120 uasm_i_dins(p, ptr, 0, 0, 23);
1122 uasm_il_bltz(p, r, tmp, label_vmalloc);
1125 if (pgd_reg == -1) {
1126 vmalloc_branch_delay_filled = 1;
1127 /* 1 0 1 0 1 << 6 xkphys cached */
1128 uasm_i_ori(p, ptr, ptr, 0x540);
1129 uasm_i_drotr(p, ptr, ptr, 11);
1132 #ifdef __PAGETABLE_PMD_FOLDED
1133 #define LOC_PTEP scratch
1135 #define LOC_PTEP ptr
1138 if (!vmalloc_branch_delay_filled)
1139 /* get pgd offset in bytes */
1140 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1142 uasm_l_vmalloc_done(l, *p);
1146 * fall-through case = badvaddr *pgd_current
1147 * vmalloc case = badvaddr swapper_pg_dir
1150 if (vmalloc_branch_delay_filled)
1151 /* get pgd offset in bytes */
1152 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1154 #ifdef __PAGETABLE_PMD_FOLDED
1155 GET_CONTEXT(p, tmp); /* get context reg */
1157 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1159 if (use_lwx_insns()) {
1160 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1162 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1163 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1166 #ifndef __PAGETABLE_PMD_FOLDED
1167 /* get pmd offset in bytes */
1168 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1169 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1170 GET_CONTEXT(p, tmp); /* get context reg */
1172 if (use_lwx_insns()) {
1173 UASM_i_LWX(p, scratch, scratch, ptr);
1175 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1176 UASM_i_LW(p, scratch, 0, ptr);
1179 /* Adjust the context during the load latency. */
1180 build_adjust_context(p, tmp);
1182 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1183 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1185 * The in the LWX case we don't want to do the load in the
1186 * delay slot. It cannot issue in the same cycle and may be
1187 * speculative and unneeded.
1189 if (use_lwx_insns())
1191 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1194 /* build_update_entries */
1195 if (use_lwx_insns()) {
1198 UASM_i_LWX(p, even, scratch, tmp);
1199 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1200 UASM_i_LWX(p, odd, scratch, tmp);
1202 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1205 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1206 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1209 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1210 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1211 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1213 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1214 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1215 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1217 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1219 if (c0_scratch_reg >= 0) {
1220 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1221 build_tlb_write_entry(p, l, r, tlb_random);
1222 uasm_l_leave(l, *p);
1223 rv.restore_scratch = 1;
1224 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1225 build_tlb_write_entry(p, l, r, tlb_random);
1226 uasm_l_leave(l, *p);
1227 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1229 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1230 build_tlb_write_entry(p, l, r, tlb_random);
1231 uasm_l_leave(l, *p);
1232 rv.restore_scratch = 1;
1235 uasm_i_eret(p); /* return from trap */
1241 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1242 * because EXL == 0. If we wrap, we can also use the 32 instruction
1243 * slots before the XTLB refill exception handler which belong to the
1244 * unused TLB refill exception.
1246 #define MIPS64_REFILL_INSNS 32
1248 static void build_r4000_tlb_refill_handler(void)
1250 u32 *p = tlb_handler;
1251 struct uasm_label *l = labels;
1252 struct uasm_reloc *r = relocs;
1254 unsigned int final_len;
1255 struct mips_huge_tlb_info htlb_info __maybe_unused;
1256 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1258 memset(tlb_handler, 0, sizeof(tlb_handler));
1259 memset(labels, 0, sizeof(labels));
1260 memset(relocs, 0, sizeof(relocs));
1261 memset(final_handler, 0, sizeof(final_handler));
1263 if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1264 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1266 vmalloc_mode = refill_scratch;
1268 htlb_info.huge_pte = K0;
1269 htlb_info.restore_scratch = 0;
1270 htlb_info.need_reload_pte = true;
1271 vmalloc_mode = refill_noscratch;
1273 * create the plain linear handler
1275 if (bcm1250_m3_war()) {
1276 unsigned int segbits = 44;
1278 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1279 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1280 uasm_i_xor(&p, K0, K0, K1);
1281 uasm_i_dsrl_safe(&p, K1, K0, 62);
1282 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1283 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1284 uasm_i_or(&p, K0, K0, K1);
1285 uasm_il_bnez(&p, &r, K0, label_leave);
1286 /* No need for uasm_i_nop */
1290 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1292 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1295 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1296 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1299 build_get_ptep(&p, K0, K1);
1300 build_update_entries(&p, K0, K1);
1301 build_tlb_write_entry(&p, &l, &r, tlb_random);
1302 uasm_l_leave(&l, p);
1303 uasm_i_eret(&p); /* return from trap */
1305 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1306 uasm_l_tlb_huge_update(&l, p);
1307 if (htlb_info.need_reload_pte)
1308 UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1309 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1310 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1311 htlb_info.restore_scratch);
1315 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1319 * Overflow check: For the 64bit handler, we need at least one
1320 * free instruction slot for the wrap-around branch. In worst
1321 * case, if the intended insertion point is a delay slot, we
1322 * need three, with the second nop'ed and the third being
1325 switch (boot_cpu_type()) {
1327 if (sizeof(long) == 4) {
1329 /* Loongson2 ebase is different than r4k, we have more space */
1330 if ((p - tlb_handler) > 64)
1331 panic("TLB refill handler space exceeded");
1333 * Now fold the handler in the TLB refill handler space.
1336 /* Simplest case, just copy the handler. */
1337 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1338 final_len = p - tlb_handler;
1341 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1342 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1343 && uasm_insn_has_bdelay(relocs,
1344 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1345 panic("TLB refill handler space exceeded");
1347 * Now fold the handler in the TLB refill handler space.
1349 f = final_handler + MIPS64_REFILL_INSNS;
1350 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1351 /* Just copy the handler. */
1352 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1353 final_len = p - tlb_handler;
1355 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1356 const enum label_id ls = label_tlb_huge_update;
1358 const enum label_id ls = label_vmalloc;
1364 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1366 BUG_ON(i == ARRAY_SIZE(labels));
1367 split = labels[i].addr;
1370 * See if we have overflown one way or the other.
1372 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1373 split < p - MIPS64_REFILL_INSNS)
1378 * Split two instructions before the end. One
1379 * for the branch and one for the instruction
1380 * in the delay slot.
1382 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1385 * If the branch would fall in a delay slot,
1386 * we must back up an additional instruction
1387 * so that it is no longer in a delay slot.
1389 if (uasm_insn_has_bdelay(relocs, split - 1))
1392 /* Copy first part of the handler. */
1393 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1394 f += split - tlb_handler;
1397 /* Insert branch. */
1398 uasm_l_split(&l, final_handler);
1399 uasm_il_b(&f, &r, label_split);
1400 if (uasm_insn_has_bdelay(relocs, split))
1403 uasm_copy_handler(relocs, labels,
1404 split, split + 1, f);
1405 uasm_move_labels(labels, f, f + 1, -1);
1411 /* Copy the rest of the handler. */
1412 uasm_copy_handler(relocs, labels, split, p, final_handler);
1413 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1420 uasm_resolve_relocs(relocs, labels);
1421 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1424 memcpy((void *)ebase, final_handler, 0x100);
1425 local_flush_icache_range(ebase, ebase + 0x100);
1427 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1430 extern u32 handle_tlbl[], handle_tlbl_end[];
1431 extern u32 handle_tlbs[], handle_tlbs_end[];
1432 extern u32 handle_tlbm[], handle_tlbm_end[];
1433 extern u32 tlbmiss_handler_setup_pgd_start[], tlbmiss_handler_setup_pgd[];
1434 extern u32 tlbmiss_handler_setup_pgd_end[];
1436 static void build_setup_pgd(void)
1439 const int __maybe_unused a1 = 5;
1440 const int __maybe_unused a2 = 6;
1441 u32 *p = tlbmiss_handler_setup_pgd_start;
1442 const int tlbmiss_handler_setup_pgd_size =
1443 tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1444 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1445 long pgdc = (long)pgd_current;
1448 memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1449 sizeof(tlbmiss_handler_setup_pgd[0]));
1450 memset(labels, 0, sizeof(labels));
1451 memset(relocs, 0, sizeof(relocs));
1452 pgd_reg = allocate_kscratch();
1453 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1454 if (pgd_reg == -1) {
1455 struct uasm_label *l = labels;
1456 struct uasm_reloc *r = relocs;
1458 /* PGD << 11 in c0_Context */
1460 * If it is a ckseg0 address, convert to a physical
1461 * address. Shifting right by 29 and adding 4 will
1462 * result in zero for these addresses.
1465 UASM_i_SRA(&p, a1, a0, 29);
1466 UASM_i_ADDIU(&p, a1, a1, 4);
1467 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1469 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1470 uasm_l_tlbl_goaround1(&l, p);
1471 UASM_i_SLL(&p, a0, a0, 11);
1473 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1475 /* PGD in c0_KScratch */
1477 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1481 /* Save PGD to pgd_current[smp_processor_id()] */
1482 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1483 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1484 UASM_i_LA_mostly(&p, a2, pgdc);
1485 UASM_i_ADDU(&p, a2, a2, a1);
1486 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1488 UASM_i_LA_mostly(&p, a2, pgdc);
1489 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1493 /* if pgd_reg is allocated, save PGD also to scratch register */
1495 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1499 if (p >= tlbmiss_handler_setup_pgd_end)
1500 panic("tlbmiss_handler_setup_pgd space exceeded");
1502 uasm_resolve_relocs(relocs, labels);
1503 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1504 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1506 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1507 tlbmiss_handler_setup_pgd_size);
1511 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1514 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1516 uasm_i_lld(p, pte, 0, ptr);
1519 UASM_i_LL(p, pte, 0, ptr);
1521 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1523 uasm_i_ld(p, pte, 0, ptr);
1526 UASM_i_LW(p, pte, 0, ptr);
1531 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1534 #ifdef CONFIG_PHYS_ADDR_T_64BIT
1535 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1538 uasm_i_ori(p, pte, pte, mode);
1540 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1542 uasm_i_scd(p, pte, 0, ptr);
1545 UASM_i_SC(p, pte, 0, ptr);
1547 if (r10000_llsc_war())
1548 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1550 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1552 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1553 if (!cpu_has_64bits) {
1554 /* no uasm_i_nop needed */
1555 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1556 uasm_i_ori(p, pte, pte, hwmode);
1557 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1558 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1559 /* no uasm_i_nop needed */
1560 uasm_i_lw(p, pte, 0, ptr);
1567 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1569 uasm_i_sd(p, pte, 0, ptr);
1572 UASM_i_SW(p, pte, 0, ptr);
1574 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1575 if (!cpu_has_64bits) {
1576 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1577 uasm_i_ori(p, pte, pte, hwmode);
1578 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1579 uasm_i_lw(p, pte, 0, ptr);
1586 * Check if PTE is present, if not then jump to LABEL. PTR points to
1587 * the page table where this PTE is located, PTE will be re-loaded
1588 * with it's original value.
1591 build_pte_present(u32 **p, struct uasm_reloc **r,
1592 int pte, int ptr, int scratch, enum label_id lid)
1594 int t = scratch >= 0 ? scratch : pte;
1597 if (use_bbit_insns()) {
1598 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1601 uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1602 uasm_il_beqz(p, r, t, lid);
1604 /* You lose the SMP race :-(*/
1605 iPTE_LW(p, pte, ptr);
1608 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1609 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1610 uasm_il_bnez(p, r, t, lid);
1612 /* You lose the SMP race :-(*/
1613 iPTE_LW(p, pte, ptr);
1617 /* Make PTE valid, store result in PTR. */
1619 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1622 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1624 iPTE_SW(p, r, pte, ptr, mode);
1628 * Check if PTE can be written to, if not branch to LABEL. Regardless
1629 * restore PTE with value from PTR when done.
1632 build_pte_writable(u32 **p, struct uasm_reloc **r,
1633 unsigned int pte, unsigned int ptr, int scratch,
1636 int t = scratch >= 0 ? scratch : pte;
1638 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1639 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1640 uasm_il_bnez(p, r, t, lid);
1642 /* You lose the SMP race :-(*/
1643 iPTE_LW(p, pte, ptr);
1648 /* Make PTE writable, update software status bits as well, then store
1652 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1655 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1658 iPTE_SW(p, r, pte, ptr, mode);
1662 * Check if PTE can be modified, if not branch to LABEL. Regardless
1663 * restore PTE with value from PTR when done.
1666 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1667 unsigned int pte, unsigned int ptr, int scratch,
1670 if (use_bbit_insns()) {
1671 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1674 int t = scratch >= 0 ? scratch : pte;
1675 uasm_i_andi(p, t, pte, _PAGE_WRITE);
1676 uasm_il_beqz(p, r, t, lid);
1678 /* You lose the SMP race :-(*/
1679 iPTE_LW(p, pte, ptr);
1683 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1687 * R3000 style TLB load/store/modify handlers.
1691 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1695 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1697 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1698 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1701 uasm_i_rfe(p); /* branch delay */
1705 * This places the pte into ENTRYLO0 and writes it with tlbwi
1706 * or tlbwr as appropriate. This is because the index register
1707 * may have the probe fail bit set as a result of a trap on a
1708 * kseg2 access, i.e. without refill. Then it returns.
1711 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1712 struct uasm_reloc **r, unsigned int pte,
1715 uasm_i_mfc0(p, tmp, C0_INDEX);
1716 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1717 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1718 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1719 uasm_i_tlbwi(p); /* cp0 delay */
1721 uasm_i_rfe(p); /* branch delay */
1722 uasm_l_r3000_write_probe_fail(l, *p);
1723 uasm_i_tlbwr(p); /* cp0 delay */
1725 uasm_i_rfe(p); /* branch delay */
1729 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1732 long pgdc = (long)pgd_current;
1734 uasm_i_mfc0(p, pte, C0_BADVADDR);
1735 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1736 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1737 uasm_i_srl(p, pte, pte, 22); /* load delay */
1738 uasm_i_sll(p, pte, pte, 2);
1739 uasm_i_addu(p, ptr, ptr, pte);
1740 uasm_i_mfc0(p, pte, C0_CONTEXT);
1741 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1742 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1743 uasm_i_addu(p, ptr, ptr, pte);
1744 uasm_i_lw(p, pte, 0, ptr);
1745 uasm_i_tlbp(p); /* load delay */
1748 static void build_r3000_tlb_load_handler(void)
1750 u32 *p = handle_tlbl;
1751 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1752 struct uasm_label *l = labels;
1753 struct uasm_reloc *r = relocs;
1755 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1756 memset(labels, 0, sizeof(labels));
1757 memset(relocs, 0, sizeof(relocs));
1759 build_r3000_tlbchange_handler_head(&p, K0, K1);
1760 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1761 uasm_i_nop(&p); /* load delay */
1762 build_make_valid(&p, &r, K0, K1);
1763 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1765 uasm_l_nopage_tlbl(&l, p);
1766 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1769 if (p >= handle_tlbl_end)
1770 panic("TLB load handler fastpath space exceeded");
1772 uasm_resolve_relocs(relocs, labels);
1773 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1774 (unsigned int)(p - handle_tlbl));
1776 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1779 static void build_r3000_tlb_store_handler(void)
1781 u32 *p = handle_tlbs;
1782 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1783 struct uasm_label *l = labels;
1784 struct uasm_reloc *r = relocs;
1786 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1787 memset(labels, 0, sizeof(labels));
1788 memset(relocs, 0, sizeof(relocs));
1790 build_r3000_tlbchange_handler_head(&p, K0, K1);
1791 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1792 uasm_i_nop(&p); /* load delay */
1793 build_make_write(&p, &r, K0, K1);
1794 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1796 uasm_l_nopage_tlbs(&l, p);
1797 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1800 if (p >= handle_tlbs_end)
1801 panic("TLB store handler fastpath space exceeded");
1803 uasm_resolve_relocs(relocs, labels);
1804 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1805 (unsigned int)(p - handle_tlbs));
1807 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1810 static void build_r3000_tlb_modify_handler(void)
1812 u32 *p = handle_tlbm;
1813 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1814 struct uasm_label *l = labels;
1815 struct uasm_reloc *r = relocs;
1817 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1818 memset(labels, 0, sizeof(labels));
1819 memset(relocs, 0, sizeof(relocs));
1821 build_r3000_tlbchange_handler_head(&p, K0, K1);
1822 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1823 uasm_i_nop(&p); /* load delay */
1824 build_make_write(&p, &r, K0, K1);
1825 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1827 uasm_l_nopage_tlbm(&l, p);
1828 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1831 if (p >= handle_tlbm_end)
1832 panic("TLB modify handler fastpath space exceeded");
1834 uasm_resolve_relocs(relocs, labels);
1835 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1836 (unsigned int)(p - handle_tlbm));
1838 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1840 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1843 * R4000 style TLB load/store/modify handlers.
1845 static struct work_registers
1846 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1847 struct uasm_reloc **r)
1849 struct work_registers wr = build_get_work_registers(p);
1852 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1854 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1857 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1859 * For huge tlb entries, pmd doesn't contain an address but
1860 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1861 * see if we need to jump to huge tlb processing.
1863 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1866 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1867 UASM_i_LW(p, wr.r2, 0, wr.r2);
1868 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1869 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1870 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1873 uasm_l_smp_pgtable_change(l, *p);
1875 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1876 if (!m4kc_tlbp_war()) {
1877 build_tlb_probe_entry(p);
1879 /* race condition happens, leaving */
1881 uasm_i_mfc0(p, wr.r3, C0_INDEX);
1882 uasm_il_bltz(p, r, wr.r3, label_leave);
1890 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1891 struct uasm_reloc **r, unsigned int tmp,
1894 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1895 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1896 build_update_entries(p, tmp, ptr);
1897 build_tlb_write_entry(p, l, r, tlb_indexed);
1898 uasm_l_leave(l, *p);
1899 build_restore_work_registers(p);
1900 uasm_i_eret(p); /* return from trap */
1903 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1907 static void build_r4000_tlb_load_handler(void)
1909 u32 *p = handle_tlbl;
1910 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1911 struct uasm_label *l = labels;
1912 struct uasm_reloc *r = relocs;
1913 struct work_registers wr;
1915 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1916 memset(labels, 0, sizeof(labels));
1917 memset(relocs, 0, sizeof(relocs));
1919 if (bcm1250_m3_war()) {
1920 unsigned int segbits = 44;
1922 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1923 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1924 uasm_i_xor(&p, K0, K0, K1);
1925 uasm_i_dsrl_safe(&p, K1, K0, 62);
1926 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1927 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1928 uasm_i_or(&p, K0, K0, K1);
1929 uasm_il_bnez(&p, &r, K0, label_leave);
1930 /* No need for uasm_i_nop */
1933 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1934 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1935 if (m4kc_tlbp_war())
1936 build_tlb_probe_entry(&p);
1938 if (cpu_has_rixi && !cpu_has_rixiex) {
1940 * If the page is not _PAGE_VALID, RI or XI could not
1941 * have triggered it. Skip the expensive test..
1943 if (use_bbit_insns()) {
1944 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1945 label_tlbl_goaround1);
1947 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1948 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1954 switch (current_cpu_type()) {
1956 if (cpu_has_mips_r2_exec_hazard) {
1959 case CPU_CAVIUM_OCTEON:
1960 case CPU_CAVIUM_OCTEON_PLUS:
1961 case CPU_CAVIUM_OCTEON2:
1966 /* Examine entrylo 0 or 1 based on ptr. */
1967 if (use_bbit_insns()) {
1968 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1970 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1971 uasm_i_beqz(&p, wr.r3, 8);
1973 /* load it in the delay slot*/
1974 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1975 /* load it if ptr is odd */
1976 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1978 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1979 * XI must have triggered it.
1981 if (use_bbit_insns()) {
1982 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1984 uasm_l_tlbl_goaround1(&l, p);
1986 uasm_i_andi(&p, wr.r3, wr.r3, 2);
1987 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1990 uasm_l_tlbl_goaround1(&l, p);
1992 build_make_valid(&p, &r, wr.r1, wr.r2);
1993 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1995 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1997 * This is the entry point when build_r4000_tlbchange_handler_head
1998 * spots a huge page.
2000 uasm_l_tlb_huge_update(&l, p);
2001 iPTE_LW(&p, wr.r1, wr.r2);
2002 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2003 build_tlb_probe_entry(&p);
2005 if (cpu_has_rixi && !cpu_has_rixiex) {
2007 * If the page is not _PAGE_VALID, RI or XI could not
2008 * have triggered it. Skip the expensive test..
2010 if (use_bbit_insns()) {
2011 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2012 label_tlbl_goaround2);
2014 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2015 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2021 switch (current_cpu_type()) {
2023 if (cpu_has_mips_r2_exec_hazard) {
2026 case CPU_CAVIUM_OCTEON:
2027 case CPU_CAVIUM_OCTEON_PLUS:
2028 case CPU_CAVIUM_OCTEON2:
2033 /* Examine entrylo 0 or 1 based on ptr. */
2034 if (use_bbit_insns()) {
2035 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2037 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2038 uasm_i_beqz(&p, wr.r3, 8);
2040 /* load it in the delay slot*/
2041 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2042 /* load it if ptr is odd */
2043 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2045 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2046 * XI must have triggered it.
2048 if (use_bbit_insns()) {
2049 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2051 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2052 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2054 if (PM_DEFAULT_MASK == 0)
2057 * We clobbered C0_PAGEMASK, restore it. On the other branch
2058 * it is restored in build_huge_tlb_write_entry.
2060 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2062 uasm_l_tlbl_goaround2(&l, p);
2064 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2065 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2068 uasm_l_nopage_tlbl(&l, p);
2069 build_restore_work_registers(&p);
2070 #ifdef CONFIG_CPU_MICROMIPS
2071 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2072 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2073 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2077 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2080 if (p >= handle_tlbl_end)
2081 panic("TLB load handler fastpath space exceeded");
2083 uasm_resolve_relocs(relocs, labels);
2084 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2085 (unsigned int)(p - handle_tlbl));
2087 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2090 static void build_r4000_tlb_store_handler(void)
2092 u32 *p = handle_tlbs;
2093 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2094 struct uasm_label *l = labels;
2095 struct uasm_reloc *r = relocs;
2096 struct work_registers wr;
2098 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2099 memset(labels, 0, sizeof(labels));
2100 memset(relocs, 0, sizeof(relocs));
2102 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2103 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2104 if (m4kc_tlbp_war())
2105 build_tlb_probe_entry(&p);
2106 build_make_write(&p, &r, wr.r1, wr.r2);
2107 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2109 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2111 * This is the entry point when
2112 * build_r4000_tlbchange_handler_head spots a huge page.
2114 uasm_l_tlb_huge_update(&l, p);
2115 iPTE_LW(&p, wr.r1, wr.r2);
2116 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2117 build_tlb_probe_entry(&p);
2118 uasm_i_ori(&p, wr.r1, wr.r1,
2119 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2120 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2123 uasm_l_nopage_tlbs(&l, p);
2124 build_restore_work_registers(&p);
2125 #ifdef CONFIG_CPU_MICROMIPS
2126 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2127 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2128 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2132 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2135 if (p >= handle_tlbs_end)
2136 panic("TLB store handler fastpath space exceeded");
2138 uasm_resolve_relocs(relocs, labels);
2139 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2140 (unsigned int)(p - handle_tlbs));
2142 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2145 static void build_r4000_tlb_modify_handler(void)
2147 u32 *p = handle_tlbm;
2148 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2149 struct uasm_label *l = labels;
2150 struct uasm_reloc *r = relocs;
2151 struct work_registers wr;
2153 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2154 memset(labels, 0, sizeof(labels));
2155 memset(relocs, 0, sizeof(relocs));
2157 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2158 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2159 if (m4kc_tlbp_war())
2160 build_tlb_probe_entry(&p);
2161 /* Present and writable bits set, set accessed and dirty bits. */
2162 build_make_write(&p, &r, wr.r1, wr.r2);
2163 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2165 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2167 * This is the entry point when
2168 * build_r4000_tlbchange_handler_head spots a huge page.
2170 uasm_l_tlb_huge_update(&l, p);
2171 iPTE_LW(&p, wr.r1, wr.r2);
2172 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2173 build_tlb_probe_entry(&p);
2174 uasm_i_ori(&p, wr.r1, wr.r1,
2175 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2176 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2179 uasm_l_nopage_tlbm(&l, p);
2180 build_restore_work_registers(&p);
2181 #ifdef CONFIG_CPU_MICROMIPS
2182 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2183 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2184 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2188 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2191 if (p >= handle_tlbm_end)
2192 panic("TLB modify handler fastpath space exceeded");
2194 uasm_resolve_relocs(relocs, labels);
2195 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2196 (unsigned int)(p - handle_tlbm));
2198 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2201 static void flush_tlb_handlers(void)
2203 local_flush_icache_range((unsigned long)handle_tlbl,
2204 (unsigned long)handle_tlbl_end);
2205 local_flush_icache_range((unsigned long)handle_tlbs,
2206 (unsigned long)handle_tlbs_end);
2207 local_flush_icache_range((unsigned long)handle_tlbm,
2208 (unsigned long)handle_tlbm_end);
2209 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2210 (unsigned long)tlbmiss_handler_setup_pgd_end);
2213 static void print_htw_config(void)
2215 unsigned long config;
2217 const int field = 2 * sizeof(unsigned long);
2219 config = read_c0_pwfield();
2220 pr_debug("PWField (0x%0*lx): GDI: 0x%02lx UDI: 0x%02lx MDI: 0x%02lx PTI: 0x%02lx PTEI: 0x%02lx\n",
2222 (config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2223 (config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2224 (config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2225 (config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2226 (config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2228 config = read_c0_pwsize();
2229 pr_debug("PWSize (0x%0*lx): GDW: 0x%02lx UDW: 0x%02lx MDW: 0x%02lx PTW: 0x%02lx PTEW: 0x%02lx\n",
2231 (config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2232 (config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2233 (config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2234 (config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2235 (config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2237 pwctl = read_c0_pwctl();
2238 pr_debug("PWCtl (0x%x): PWEn: 0x%x DPH: 0x%x HugePg: 0x%x Psn: 0x%x\n",
2240 (pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2241 (pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2242 (pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2243 (pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2246 static void config_htw_params(void)
2248 unsigned long pwfield, pwsize, ptei;
2249 unsigned int config;
2252 * We are using 2-level page tables, so we only need to
2253 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2254 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2255 * write values less than 0xc in these fields because the entire
2256 * write will be dropped. As a result of which, we must preserve
2257 * the original reset values and overwrite only what we really want.
2260 pwfield = read_c0_pwfield();
2261 /* re-initialize the GDI field */
2262 pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2263 pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2264 /* re-initialize the PTI field including the even/odd bit */
2265 pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2266 pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2267 /* Set the PTEI right shift */
2268 ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2270 write_c0_pwfield(pwfield);
2271 /* Check whether the PTEI value is supported */
2272 back_to_back_c0_hazard();
2273 pwfield = read_c0_pwfield();
2274 if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2276 pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2279 * Drop option to avoid HTW being enabled via another path
2282 current_cpu_data.options &= ~MIPS_CPU_HTW;
2286 pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2287 pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2288 write_c0_pwsize(pwsize);
2290 /* Make sure everything is set before we enable the HTW */
2291 back_to_back_c0_hazard();
2293 /* Enable HTW and disable the rest of the pwctl fields */
2294 config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2295 write_c0_pwctl(config);
2296 pr_info("Hardware Page Table Walker enabled\n");
2301 void build_tlb_refill_handler(void)
2304 * The refill handler is generated per-CPU, multi-node systems
2305 * may have local storage for it. The other handlers are only
2308 static int run_once = 0;
2310 output_pgtable_bits_defines();
2313 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2316 switch (current_cpu_type()) {
2324 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2325 if (cpu_has_local_ebase)
2326 build_r3000_tlb_refill_handler();
2328 if (!cpu_has_local_ebase)
2329 build_r3000_tlb_refill_handler();
2331 build_r3000_tlb_load_handler();
2332 build_r3000_tlb_store_handler();
2333 build_r3000_tlb_modify_handler();
2334 flush_tlb_handlers();
2338 panic("No R3000 TLB refill handler");
2344 panic("No R6000 TLB refill handler yet");
2348 panic("No R8000 TLB refill handler yet");
2353 scratch_reg = allocate_kscratch();
2355 build_r4000_tlb_load_handler();
2356 build_r4000_tlb_store_handler();
2357 build_r4000_tlb_modify_handler();
2358 if (!cpu_has_local_ebase)
2359 build_r4000_tlb_refill_handler();
2360 flush_tlb_handlers();
2363 if (cpu_has_local_ebase)
2364 build_r4000_tlb_refill_handler();
2366 config_htw_params();