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>
38 static int mips_xpa_disabled;
40 static int __init xpa_disable(char *s)
42 mips_xpa_disabled = 1;
47 __setup("noxpa", xpa_disable);
50 * TLB load/store/modify handlers.
52 * Only the fastpath gets synthesized at runtime, the slowpath for
53 * do_page_fault remains normal asm.
55 extern void tlb_do_page_fault_0(void);
56 extern void tlb_do_page_fault_1(void);
58 struct work_registers {
67 } ____cacheline_aligned_in_smp;
69 static struct tlb_reg_save handler_reg_save[NR_CPUS];
71 static inline int r45k_bvahwbug(void)
73 /* XXX: We should probe for the presence of this bug, but we don't. */
77 static inline int r4k_250MHZhwbug(void)
79 /* XXX: We should probe for the presence of this bug, but we don't. */
83 static inline int __maybe_unused bcm1250_m3_war(void)
85 return BCM1250_M3_WAR;
88 static inline int __maybe_unused r10000_llsc_war(void)
90 return R10000_LLSC_WAR;
93 static int use_bbit_insns(void)
95 switch (current_cpu_type()) {
96 case CPU_CAVIUM_OCTEON:
97 case CPU_CAVIUM_OCTEON_PLUS:
98 case CPU_CAVIUM_OCTEON2:
99 case CPU_CAVIUM_OCTEON3:
106 static int use_lwx_insns(void)
108 switch (current_cpu_type()) {
109 case CPU_CAVIUM_OCTEON2:
110 case CPU_CAVIUM_OCTEON3:
116 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
117 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
118 static bool scratchpad_available(void)
122 static int scratchpad_offset(int i)
125 * CVMSEG starts at address -32768 and extends for
126 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
128 i += 1; /* Kernel use starts at the top and works down. */
129 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
132 static bool scratchpad_available(void)
136 static int scratchpad_offset(int i)
139 /* Really unreachable, but evidently some GCC want this. */
144 * Found by experiment: At least some revisions of the 4kc throw under
145 * some circumstances a machine check exception, triggered by invalid
146 * values in the index register. Delaying the tlbp instruction until
147 * after the next branch, plus adding an additional nop in front of
148 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
149 * why; it's not an issue caused by the core RTL.
152 static int m4kc_tlbp_war(void)
154 return (current_cpu_data.processor_id & 0xffff00) ==
155 (PRID_COMP_MIPS | PRID_IMP_4KC);
158 /* Handle labels (which must be positive integers). */
160 label_second_part = 1,
165 label_split = label_tlbw_hazard_0 + 8,
166 label_tlbl_goaround1,
167 label_tlbl_goaround2,
171 label_smp_pgtable_change,
172 label_r3000_write_probe_fail,
173 label_large_segbits_fault,
174 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
175 label_tlb_huge_update,
179 UASM_L_LA(_second_part)
182 UASM_L_LA(_vmalloc_done)
183 /* _tlbw_hazard_x is handled differently. */
185 UASM_L_LA(_tlbl_goaround1)
186 UASM_L_LA(_tlbl_goaround2)
187 UASM_L_LA(_nopage_tlbl)
188 UASM_L_LA(_nopage_tlbs)
189 UASM_L_LA(_nopage_tlbm)
190 UASM_L_LA(_smp_pgtable_change)
191 UASM_L_LA(_r3000_write_probe_fail)
192 UASM_L_LA(_large_segbits_fault)
193 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
194 UASM_L_LA(_tlb_huge_update)
197 static int hazard_instance;
199 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
203 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
210 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
214 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
222 * pgtable bits are assigned dynamically depending on processor feature
223 * and statically based on kernel configuration. This spits out the actual
224 * values the kernel is using. Required to make sense from disassembled
225 * TLB exception handlers.
227 static void output_pgtable_bits_defines(void)
229 #define pr_define(fmt, ...) \
230 pr_debug("#define " fmt, ##__VA_ARGS__)
232 pr_debug("#include <asm/asm.h>\n");
233 pr_debug("#include <asm/regdef.h>\n");
236 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
237 pr_define("_PAGE_READ_SHIFT %d\n", _PAGE_READ_SHIFT);
238 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
239 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
240 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
241 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
242 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
244 #ifdef CONFIG_CPU_MIPSR2
246 #ifdef _PAGE_NO_EXEC_SHIFT
247 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
248 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
252 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
253 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
254 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
255 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
259 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
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_BADVADDR 8, 0
288 #define C0_ENTRYHI 10, 0
290 #define C0_XCONTEXT 20, 0
293 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
295 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
298 /* The worst case length of the handler is around 18 instructions for
299 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
300 * Maximum space available is 32 instructions for R3000 and 64
301 * instructions for R4000.
303 * We deliberately chose a buffer size of 128, so we won't scribble
304 * over anything important on overflow before we panic.
306 static u32 tlb_handler[128];
308 /* simply assume worst case size for labels and relocs */
309 static struct uasm_label labels[128];
310 static struct uasm_reloc relocs[128];
312 static int check_for_high_segbits;
314 static unsigned int kscratch_used_mask;
316 static inline int __maybe_unused c0_kscratch(void)
318 switch (current_cpu_type()) {
327 static int allocate_kscratch(void)
330 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
337 r--; /* make it zero based */
339 kscratch_used_mask |= (1 << r);
344 static int scratch_reg;
346 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
348 static struct work_registers build_get_work_registers(u32 **p)
350 struct work_registers r;
352 if (scratch_reg >= 0) {
353 /* Save in CPU local C0_KScratch? */
354 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
361 if (num_possible_cpus() > 1) {
362 /* Get smp_processor_id */
363 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
364 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
366 /* handler_reg_save index in K0 */
367 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
369 UASM_i_LA(p, K1, (long)&handler_reg_save);
370 UASM_i_ADDU(p, K0, K0, K1);
372 UASM_i_LA(p, K0, (long)&handler_reg_save);
374 /* K0 now points to save area, save $1 and $2 */
375 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
376 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
384 static void build_restore_work_registers(u32 **p)
386 if (scratch_reg >= 0) {
387 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
390 /* K0 already points to save area, restore $1 and $2 */
391 UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
392 UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
395 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
398 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
399 * we cannot do r3000 under these circumstances.
401 * Declare pgd_current here instead of including mmu_context.h to avoid type
402 * conflicts for tlbmiss_handler_setup_pgd
404 extern unsigned long pgd_current[];
407 * The R3000 TLB handler is simple.
409 static void build_r3000_tlb_refill_handler(void)
411 long pgdc = (long)pgd_current;
414 memset(tlb_handler, 0, sizeof(tlb_handler));
417 uasm_i_mfc0(&p, K0, C0_BADVADDR);
418 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
419 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
420 uasm_i_srl(&p, K0, K0, 22); /* load delay */
421 uasm_i_sll(&p, K0, K0, 2);
422 uasm_i_addu(&p, K1, K1, K0);
423 uasm_i_mfc0(&p, K0, C0_CONTEXT);
424 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
425 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
426 uasm_i_addu(&p, K1, K1, K0);
427 uasm_i_lw(&p, K0, 0, K1);
428 uasm_i_nop(&p); /* load delay */
429 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
430 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
431 uasm_i_tlbwr(&p); /* cp0 delay */
433 uasm_i_rfe(&p); /* branch delay */
435 if (p > tlb_handler + 32)
436 panic("TLB refill handler space exceeded");
438 pr_debug("Wrote TLB refill handler (%u instructions).\n",
439 (unsigned int)(p - tlb_handler));
441 memcpy((void *)ebase, tlb_handler, 0x80);
442 local_flush_icache_range(ebase, ebase + 0x80);
444 dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
446 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
449 * The R4000 TLB handler is much more complicated. We have two
450 * consecutive handler areas with 32 instructions space each.
451 * Since they aren't used at the same time, we can overflow in the
452 * other one.To keep things simple, we first assume linear space,
453 * then we relocate it to the final handler layout as needed.
455 static u32 final_handler[64];
460 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
461 * 2. A timing hazard exists for the TLBP instruction.
463 * stalling_instruction
466 * The JTLB is being read for the TLBP throughout the stall generated by the
467 * previous instruction. This is not really correct as the stalling instruction
468 * can modify the address used to access the JTLB. The failure symptom is that
469 * the TLBP instruction will use an address created for the stalling instruction
470 * and not the address held in C0_ENHI and thus report the wrong results.
472 * The software work-around is to not allow the instruction preceding the TLBP
473 * to stall - make it an NOP or some other instruction guaranteed not to stall.
475 * Errata 2 will not be fixed. This errata is also on the R5000.
477 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
479 static void __maybe_unused build_tlb_probe_entry(u32 **p)
481 switch (current_cpu_type()) {
482 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
498 * Write random or indexed TLB entry, and care about the hazards from
499 * the preceding mtc0 and for the following eret.
501 enum tlb_write_entry { tlb_random, tlb_indexed };
503 static void build_tlb_write_entry(u32 **p, struct uasm_label **l,
504 struct uasm_reloc **r,
505 enum tlb_write_entry wmode)
507 void(*tlbw)(u32 **) = NULL;
510 case tlb_random: tlbw = uasm_i_tlbwr; break;
511 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
514 if (cpu_has_mips_r2_r6) {
515 if (cpu_has_mips_r2_exec_hazard)
521 switch (current_cpu_type()) {
529 * This branch uses up a mtc0 hazard nop slot and saves
530 * two nops after the tlbw instruction.
532 uasm_bgezl_hazard(p, r, hazard_instance);
534 uasm_bgezl_label(l, p, hazard_instance);
548 uasm_i_nop(p); /* QED specifies 2 nops hazard */
549 uasm_i_nop(p); /* QED specifies 2 nops hazard */
623 panic("No TLB refill handler yet (CPU type: %d)",
629 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
633 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
635 #ifdef CONFIG_PHYS_ADDR_T_64BIT
636 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
638 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
643 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
645 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
646 unsigned int tmp, enum label_id lid,
649 if (restore_scratch) {
650 /* Reset default page size */
651 if (PM_DEFAULT_MASK >> 16) {
652 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
653 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
654 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
655 uasm_il_b(p, r, lid);
656 } else if (PM_DEFAULT_MASK) {
657 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
658 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
659 uasm_il_b(p, r, lid);
661 uasm_i_mtc0(p, 0, C0_PAGEMASK);
662 uasm_il_b(p, r, lid);
664 if (scratch_reg >= 0)
665 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
667 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
669 /* Reset default page size */
670 if (PM_DEFAULT_MASK >> 16) {
671 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
672 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
673 uasm_il_b(p, r, lid);
674 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
675 } else if (PM_DEFAULT_MASK) {
676 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
677 uasm_il_b(p, r, lid);
678 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
680 uasm_il_b(p, r, lid);
681 uasm_i_mtc0(p, 0, C0_PAGEMASK);
686 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
687 struct uasm_reloc **r,
689 enum tlb_write_entry wmode,
692 /* Set huge page tlb entry size */
693 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
694 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
695 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
697 build_tlb_write_entry(p, l, r, wmode);
699 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
703 * Check if Huge PTE is present, if so then jump to LABEL.
706 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
707 unsigned int pmd, int lid)
709 UASM_i_LW(p, tmp, 0, pmd);
710 if (use_bbit_insns()) {
711 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
713 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
714 uasm_il_bnez(p, r, tmp, lid);
718 static void build_huge_update_entries(u32 **p, unsigned int pte,
724 * A huge PTE describes an area the size of the
725 * configured huge page size. This is twice the
726 * of the large TLB entry size we intend to use.
727 * A TLB entry half the size of the configured
728 * huge page size is configured into entrylo0
729 * and entrylo1 to cover the contiguous huge PTE
732 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
734 /* We can clobber tmp. It isn't used after this.*/
736 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
738 build_convert_pte_to_entrylo(p, pte);
739 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
740 /* convert to entrylo1 */
742 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
744 UASM_i_ADDU(p, pte, pte, tmp);
746 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
749 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
750 struct uasm_label **l,
755 UASM_i_SC(p, pte, 0, ptr);
756 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
757 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
759 UASM_i_SW(p, pte, 0, ptr);
761 build_huge_update_entries(p, pte, ptr);
762 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
764 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
768 * TMP and PTR are scratch.
769 * TMP will be clobbered, PTR will hold the pmd entry.
772 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
773 unsigned int tmp, unsigned int ptr)
775 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
776 long pgdc = (long)pgd_current;
779 * The vmalloc handling is not in the hotpath.
781 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
783 if (check_for_high_segbits) {
785 * The kernel currently implicitely assumes that the
786 * MIPS SEGBITS parameter for the processor is
787 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
788 * allocate virtual addresses outside the maximum
789 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
790 * that doesn't prevent user code from accessing the
791 * higher xuseg addresses. Here, we make sure that
792 * everything but the lower xuseg addresses goes down
793 * the module_alloc/vmalloc path.
795 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
796 uasm_il_bnez(p, r, ptr, label_vmalloc);
798 uasm_il_bltz(p, r, tmp, label_vmalloc);
800 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
803 /* pgd is in pgd_reg */
804 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
806 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
808 * &pgd << 11 stored in CONTEXT [23..63].
810 UASM_i_MFC0(p, ptr, C0_CONTEXT);
812 /* Clear lower 23 bits of context. */
813 uasm_i_dins(p, ptr, 0, 0, 23);
815 /* 1 0 1 0 1 << 6 xkphys cached */
816 uasm_i_ori(p, ptr, ptr, 0x540);
817 uasm_i_drotr(p, ptr, ptr, 11);
818 #elif defined(CONFIG_SMP)
819 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
820 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
821 UASM_i_LA_mostly(p, tmp, pgdc);
822 uasm_i_daddu(p, ptr, ptr, tmp);
823 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
824 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
826 UASM_i_LA_mostly(p, ptr, pgdc);
827 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
831 uasm_l_vmalloc_done(l, *p);
833 /* get pgd offset in bytes */
834 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
836 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
837 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
838 #ifndef __PAGETABLE_PMD_FOLDED
839 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
840 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
841 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
842 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
843 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
848 * BVADDR is the faulting address, PTR is scratch.
849 * PTR will hold the pgd for vmalloc.
852 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
853 unsigned int bvaddr, unsigned int ptr,
854 enum vmalloc64_mode mode)
856 long swpd = (long)swapper_pg_dir;
857 int single_insn_swpd;
858 int did_vmalloc_branch = 0;
860 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
862 uasm_l_vmalloc(l, *p);
864 if (mode != not_refill && check_for_high_segbits) {
865 if (single_insn_swpd) {
866 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
867 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
868 did_vmalloc_branch = 1;
871 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
874 if (!did_vmalloc_branch) {
875 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
876 uasm_il_b(p, r, label_vmalloc_done);
877 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
879 UASM_i_LA_mostly(p, ptr, swpd);
880 uasm_il_b(p, r, label_vmalloc_done);
881 if (uasm_in_compat_space_p(swpd))
882 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
884 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
887 if (mode != not_refill && check_for_high_segbits) {
888 uasm_l_large_segbits_fault(l, *p);
890 * We get here if we are an xsseg address, or if we are
891 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
893 * Ignoring xsseg (assume disabled so would generate
894 * (address errors?), the only remaining possibility
895 * is the upper xuseg addresses. On processors with
896 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
897 * addresses would have taken an address error. We try
898 * to mimic that here by taking a load/istream page
901 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
904 if (mode == refill_scratch) {
905 if (scratch_reg >= 0)
906 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
908 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
915 #else /* !CONFIG_64BIT */
918 * TMP and PTR are scratch.
919 * TMP will be clobbered, PTR will hold the pgd entry.
921 static void __maybe_unused
922 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
925 /* pgd is in pgd_reg */
926 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
927 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
929 long pgdc = (long)pgd_current;
931 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
933 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
934 UASM_i_LA_mostly(p, tmp, pgdc);
935 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
936 uasm_i_addu(p, ptr, tmp, ptr);
938 UASM_i_LA_mostly(p, ptr, pgdc);
940 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
941 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
943 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
944 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
945 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
948 #endif /* !CONFIG_64BIT */
950 static void build_adjust_context(u32 **p, unsigned int ctx)
952 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
953 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
955 switch (current_cpu_type()) {
972 UASM_i_SRL(p, ctx, ctx, shift);
973 uasm_i_andi(p, ctx, ctx, mask);
976 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
979 * Bug workaround for the Nevada. It seems as if under certain
980 * circumstances the move from cp0_context might produce a
981 * bogus result when the mfc0 instruction and its consumer are
982 * in a different cacheline or a load instruction, probably any
983 * memory reference, is between them.
985 switch (current_cpu_type()) {
987 UASM_i_LW(p, ptr, 0, ptr);
988 GET_CONTEXT(p, tmp); /* get context reg */
992 GET_CONTEXT(p, tmp); /* get context reg */
993 UASM_i_LW(p, ptr, 0, ptr);
997 build_adjust_context(p, tmp);
998 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1001 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1004 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1005 * Kernel is a special case. Only a few CPUs use it.
1007 #ifdef CONFIG_PHYS_ADDR_T_64BIT
1008 if (cpu_has_64bits) {
1009 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1010 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1012 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1013 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1014 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1016 uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1017 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1018 uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1020 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1022 int pte_off_even = sizeof(pte_t) / 2;
1023 int pte_off_odd = pte_off_even + sizeof(pte_t);
1025 const int scratch = 1; /* Our extra working register */
1027 uasm_i_addu(p, scratch, 0, ptep);
1029 uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1030 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* odd pte */
1031 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1032 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1033 UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1034 UASM_i_MTC0(p, ptep, C0_ENTRYLO1);
1036 uasm_i_lw(p, tmp, 0, scratch);
1037 uasm_i_lw(p, ptep, sizeof(pte_t), scratch);
1038 uasm_i_lui(p, scratch, 0xff);
1039 uasm_i_ori(p, scratch, scratch, 0xffff);
1040 uasm_i_and(p, tmp, scratch, tmp);
1041 uasm_i_and(p, ptep, scratch, ptep);
1042 uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1043 uasm_i_mthc0(p, ptep, C0_ENTRYLO1);
1047 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1048 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1049 if (r45k_bvahwbug())
1050 build_tlb_probe_entry(p);
1052 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1053 if (r4k_250MHZhwbug())
1054 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1055 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1056 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1058 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1059 if (r4k_250MHZhwbug())
1060 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1061 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1062 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1063 if (r45k_bvahwbug())
1064 uasm_i_mfc0(p, tmp, C0_INDEX);
1066 if (r4k_250MHZhwbug())
1067 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1068 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1072 struct mips_huge_tlb_info {
1074 int restore_scratch;
1075 bool need_reload_pte;
1078 static struct mips_huge_tlb_info
1079 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1080 struct uasm_reloc **r, unsigned int tmp,
1081 unsigned int ptr, int c0_scratch_reg)
1083 struct mips_huge_tlb_info rv;
1084 unsigned int even, odd;
1085 int vmalloc_branch_delay_filled = 0;
1086 const int scratch = 1; /* Our extra working register */
1088 rv.huge_pte = scratch;
1089 rv.restore_scratch = 0;
1090 rv.need_reload_pte = false;
1092 if (check_for_high_segbits) {
1093 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1096 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1098 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1100 if (c0_scratch_reg >= 0)
1101 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1103 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1105 uasm_i_dsrl_safe(p, scratch, tmp,
1106 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1107 uasm_il_bnez(p, r, scratch, label_vmalloc);
1109 if (pgd_reg == -1) {
1110 vmalloc_branch_delay_filled = 1;
1111 /* Clear lower 23 bits of context. */
1112 uasm_i_dins(p, ptr, 0, 0, 23);
1116 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1118 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1120 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1122 if (c0_scratch_reg >= 0)
1123 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1125 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1128 /* Clear lower 23 bits of context. */
1129 uasm_i_dins(p, ptr, 0, 0, 23);
1131 uasm_il_bltz(p, r, tmp, label_vmalloc);
1134 if (pgd_reg == -1) {
1135 vmalloc_branch_delay_filled = 1;
1136 /* 1 0 1 0 1 << 6 xkphys cached */
1137 uasm_i_ori(p, ptr, ptr, 0x540);
1138 uasm_i_drotr(p, ptr, ptr, 11);
1141 #ifdef __PAGETABLE_PMD_FOLDED
1142 #define LOC_PTEP scratch
1144 #define LOC_PTEP ptr
1147 if (!vmalloc_branch_delay_filled)
1148 /* get pgd offset in bytes */
1149 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1151 uasm_l_vmalloc_done(l, *p);
1155 * fall-through case = badvaddr *pgd_current
1156 * vmalloc case = badvaddr swapper_pg_dir
1159 if (vmalloc_branch_delay_filled)
1160 /* get pgd offset in bytes */
1161 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1163 #ifdef __PAGETABLE_PMD_FOLDED
1164 GET_CONTEXT(p, tmp); /* get context reg */
1166 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1168 if (use_lwx_insns()) {
1169 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1171 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1172 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1175 #ifndef __PAGETABLE_PMD_FOLDED
1176 /* get pmd offset in bytes */
1177 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1178 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1179 GET_CONTEXT(p, tmp); /* get context reg */
1181 if (use_lwx_insns()) {
1182 UASM_i_LWX(p, scratch, scratch, ptr);
1184 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1185 UASM_i_LW(p, scratch, 0, ptr);
1188 /* Adjust the context during the load latency. */
1189 build_adjust_context(p, tmp);
1191 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1192 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1194 * The in the LWX case we don't want to do the load in the
1195 * delay slot. It cannot issue in the same cycle and may be
1196 * speculative and unneeded.
1198 if (use_lwx_insns())
1200 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1203 /* build_update_entries */
1204 if (use_lwx_insns()) {
1207 UASM_i_LWX(p, even, scratch, tmp);
1208 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1209 UASM_i_LWX(p, odd, scratch, tmp);
1211 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1214 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1215 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1218 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1219 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1220 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1222 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1223 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1224 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1226 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1228 if (c0_scratch_reg >= 0) {
1229 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1230 build_tlb_write_entry(p, l, r, tlb_random);
1231 uasm_l_leave(l, *p);
1232 rv.restore_scratch = 1;
1233 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1234 build_tlb_write_entry(p, l, r, tlb_random);
1235 uasm_l_leave(l, *p);
1236 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1238 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1239 build_tlb_write_entry(p, l, r, tlb_random);
1240 uasm_l_leave(l, *p);
1241 rv.restore_scratch = 1;
1244 uasm_i_eret(p); /* return from trap */
1250 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1251 * because EXL == 0. If we wrap, we can also use the 32 instruction
1252 * slots before the XTLB refill exception handler which belong to the
1253 * unused TLB refill exception.
1255 #define MIPS64_REFILL_INSNS 32
1257 static void build_r4000_tlb_refill_handler(void)
1259 u32 *p = tlb_handler;
1260 struct uasm_label *l = labels;
1261 struct uasm_reloc *r = relocs;
1263 unsigned int final_len;
1264 struct mips_huge_tlb_info htlb_info __maybe_unused;
1265 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1267 memset(tlb_handler, 0, sizeof(tlb_handler));
1268 memset(labels, 0, sizeof(labels));
1269 memset(relocs, 0, sizeof(relocs));
1270 memset(final_handler, 0, sizeof(final_handler));
1272 if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1273 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1275 vmalloc_mode = refill_scratch;
1277 htlb_info.huge_pte = K0;
1278 htlb_info.restore_scratch = 0;
1279 htlb_info.need_reload_pte = true;
1280 vmalloc_mode = refill_noscratch;
1282 * create the plain linear handler
1284 if (bcm1250_m3_war()) {
1285 unsigned int segbits = 44;
1287 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1288 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1289 uasm_i_xor(&p, K0, K0, K1);
1290 uasm_i_dsrl_safe(&p, K1, K0, 62);
1291 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1292 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1293 uasm_i_or(&p, K0, K0, K1);
1294 uasm_il_bnez(&p, &r, K0, label_leave);
1295 /* No need for uasm_i_nop */
1299 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1301 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1304 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1305 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1308 build_get_ptep(&p, K0, K1);
1309 build_update_entries(&p, K0, K1);
1310 build_tlb_write_entry(&p, &l, &r, tlb_random);
1311 uasm_l_leave(&l, p);
1312 uasm_i_eret(&p); /* return from trap */
1314 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1315 uasm_l_tlb_huge_update(&l, p);
1316 if (htlb_info.need_reload_pte)
1317 UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1318 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1319 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1320 htlb_info.restore_scratch);
1324 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1328 * Overflow check: For the 64bit handler, we need at least one
1329 * free instruction slot for the wrap-around branch. In worst
1330 * case, if the intended insertion point is a delay slot, we
1331 * need three, with the second nop'ed and the third being
1334 switch (boot_cpu_type()) {
1336 if (sizeof(long) == 4) {
1338 /* Loongson2 ebase is different than r4k, we have more space */
1339 if ((p - tlb_handler) > 64)
1340 panic("TLB refill handler space exceeded");
1342 * Now fold the handler in the TLB refill handler space.
1345 /* Simplest case, just copy the handler. */
1346 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1347 final_len = p - tlb_handler;
1350 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1351 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1352 && uasm_insn_has_bdelay(relocs,
1353 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1354 panic("TLB refill handler space exceeded");
1356 * Now fold the handler in the TLB refill handler space.
1358 f = final_handler + MIPS64_REFILL_INSNS;
1359 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1360 /* Just copy the handler. */
1361 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1362 final_len = p - tlb_handler;
1364 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1365 const enum label_id ls = label_tlb_huge_update;
1367 const enum label_id ls = label_vmalloc;
1373 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1375 BUG_ON(i == ARRAY_SIZE(labels));
1376 split = labels[i].addr;
1379 * See if we have overflown one way or the other.
1381 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1382 split < p - MIPS64_REFILL_INSNS)
1387 * Split two instructions before the end. One
1388 * for the branch and one for the instruction
1389 * in the delay slot.
1391 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1394 * If the branch would fall in a delay slot,
1395 * we must back up an additional instruction
1396 * so that it is no longer in a delay slot.
1398 if (uasm_insn_has_bdelay(relocs, split - 1))
1401 /* Copy first part of the handler. */
1402 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1403 f += split - tlb_handler;
1406 /* Insert branch. */
1407 uasm_l_split(&l, final_handler);
1408 uasm_il_b(&f, &r, label_split);
1409 if (uasm_insn_has_bdelay(relocs, split))
1412 uasm_copy_handler(relocs, labels,
1413 split, split + 1, f);
1414 uasm_move_labels(labels, f, f + 1, -1);
1420 /* Copy the rest of the handler. */
1421 uasm_copy_handler(relocs, labels, split, p, final_handler);
1422 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1429 uasm_resolve_relocs(relocs, labels);
1430 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1433 memcpy((void *)ebase, final_handler, 0x100);
1434 local_flush_icache_range(ebase, ebase + 0x100);
1436 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1439 extern u32 handle_tlbl[], handle_tlbl_end[];
1440 extern u32 handle_tlbs[], handle_tlbs_end[];
1441 extern u32 handle_tlbm[], handle_tlbm_end[];
1442 extern u32 tlbmiss_handler_setup_pgd_start[], tlbmiss_handler_setup_pgd[];
1443 extern u32 tlbmiss_handler_setup_pgd_end[];
1445 static void build_setup_pgd(void)
1448 const int __maybe_unused a1 = 5;
1449 const int __maybe_unused a2 = 6;
1450 u32 *p = tlbmiss_handler_setup_pgd_start;
1451 const int tlbmiss_handler_setup_pgd_size =
1452 tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1453 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1454 long pgdc = (long)pgd_current;
1457 memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1458 sizeof(tlbmiss_handler_setup_pgd[0]));
1459 memset(labels, 0, sizeof(labels));
1460 memset(relocs, 0, sizeof(relocs));
1461 pgd_reg = allocate_kscratch();
1462 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1463 if (pgd_reg == -1) {
1464 struct uasm_label *l = labels;
1465 struct uasm_reloc *r = relocs;
1467 /* PGD << 11 in c0_Context */
1469 * If it is a ckseg0 address, convert to a physical
1470 * address. Shifting right by 29 and adding 4 will
1471 * result in zero for these addresses.
1474 UASM_i_SRA(&p, a1, a0, 29);
1475 UASM_i_ADDIU(&p, a1, a1, 4);
1476 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1478 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1479 uasm_l_tlbl_goaround1(&l, p);
1480 UASM_i_SLL(&p, a0, a0, 11);
1482 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1484 /* PGD in c0_KScratch */
1486 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1490 /* Save PGD to pgd_current[smp_processor_id()] */
1491 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1492 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1493 UASM_i_LA_mostly(&p, a2, pgdc);
1494 UASM_i_ADDU(&p, a2, a2, a1);
1495 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1497 UASM_i_LA_mostly(&p, a2, pgdc);
1498 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1502 /* if pgd_reg is allocated, save PGD also to scratch register */
1504 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1508 if (p >= tlbmiss_handler_setup_pgd_end)
1509 panic("tlbmiss_handler_setup_pgd space exceeded");
1511 uasm_resolve_relocs(relocs, labels);
1512 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1513 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1515 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1516 tlbmiss_handler_setup_pgd_size);
1520 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1523 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1525 uasm_i_lld(p, pte, 0, ptr);
1528 UASM_i_LL(p, pte, 0, ptr);
1530 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1532 uasm_i_ld(p, pte, 0, ptr);
1535 UASM_i_LW(p, pte, 0, ptr);
1540 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1543 #ifdef CONFIG_PHYS_ADDR_T_64BIT
1544 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1546 if (!cpu_has_64bits) {
1547 const int scratch = 1; /* Our extra working register */
1549 uasm_i_lui(p, scratch, (mode >> 16));
1550 uasm_i_or(p, pte, pte, scratch);
1553 uasm_i_ori(p, pte, pte, mode);
1555 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1557 uasm_i_scd(p, pte, 0, ptr);
1560 UASM_i_SC(p, pte, 0, ptr);
1562 if (r10000_llsc_war())
1563 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1565 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1567 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1568 if (!cpu_has_64bits) {
1569 /* no uasm_i_nop needed */
1570 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1571 uasm_i_ori(p, pte, pte, hwmode);
1572 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1573 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1574 /* no uasm_i_nop needed */
1575 uasm_i_lw(p, pte, 0, ptr);
1582 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1584 uasm_i_sd(p, pte, 0, ptr);
1587 UASM_i_SW(p, pte, 0, ptr);
1589 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1590 if (!cpu_has_64bits) {
1591 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1592 uasm_i_ori(p, pte, pte, hwmode);
1593 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1594 uasm_i_lw(p, pte, 0, ptr);
1601 * Check if PTE is present, if not then jump to LABEL. PTR points to
1602 * the page table where this PTE is located, PTE will be re-loaded
1603 * with it's original value.
1606 build_pte_present(u32 **p, struct uasm_reloc **r,
1607 int pte, int ptr, int scratch, enum label_id lid)
1609 int t = scratch >= 0 ? scratch : pte;
1613 if (use_bbit_insns()) {
1614 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1617 if (_PAGE_PRESENT_SHIFT) {
1618 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1621 uasm_i_andi(p, t, cur, 1);
1622 uasm_il_beqz(p, r, t, lid);
1624 /* You lose the SMP race :-(*/
1625 iPTE_LW(p, pte, ptr);
1628 if (_PAGE_PRESENT_SHIFT) {
1629 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1632 uasm_i_andi(p, t, cur,
1633 (_PAGE_PRESENT | _PAGE_READ) >> _PAGE_PRESENT_SHIFT);
1634 uasm_i_xori(p, t, t,
1635 (_PAGE_PRESENT | _PAGE_READ) >> _PAGE_PRESENT_SHIFT);
1636 uasm_il_bnez(p, r, t, lid);
1638 /* You lose the SMP race :-(*/
1639 iPTE_LW(p, pte, ptr);
1643 /* Make PTE valid, store result in PTR. */
1645 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1648 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1650 iPTE_SW(p, r, pte, ptr, mode);
1654 * Check if PTE can be written to, if not branch to LABEL. Regardless
1655 * restore PTE with value from PTR when done.
1658 build_pte_writable(u32 **p, struct uasm_reloc **r,
1659 unsigned int pte, unsigned int ptr, int scratch,
1662 int t = scratch >= 0 ? scratch : pte;
1665 if (_PAGE_PRESENT_SHIFT) {
1666 uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1669 uasm_i_andi(p, t, cur,
1670 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1671 uasm_i_xori(p, t, t,
1672 (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1673 uasm_il_bnez(p, r, t, lid);
1675 /* You lose the SMP race :-(*/
1676 iPTE_LW(p, pte, ptr);
1681 /* Make PTE writable, update software status bits as well, then store
1685 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1688 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1691 iPTE_SW(p, r, pte, ptr, mode);
1695 * Check if PTE can be modified, if not branch to LABEL. Regardless
1696 * restore PTE with value from PTR when done.
1699 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1700 unsigned int pte, unsigned int ptr, int scratch,
1703 if (use_bbit_insns()) {
1704 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1707 int t = scratch >= 0 ? scratch : pte;
1708 uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1709 uasm_i_andi(p, t, t, 1);
1710 uasm_il_beqz(p, r, t, lid);
1712 /* You lose the SMP race :-(*/
1713 iPTE_LW(p, pte, ptr);
1717 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1721 * R3000 style TLB load/store/modify handlers.
1725 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1729 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1731 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1732 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1735 uasm_i_rfe(p); /* branch delay */
1739 * This places the pte into ENTRYLO0 and writes it with tlbwi
1740 * or tlbwr as appropriate. This is because the index register
1741 * may have the probe fail bit set as a result of a trap on a
1742 * kseg2 access, i.e. without refill. Then it returns.
1745 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1746 struct uasm_reloc **r, unsigned int pte,
1749 uasm_i_mfc0(p, tmp, C0_INDEX);
1750 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1751 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1752 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1753 uasm_i_tlbwi(p); /* cp0 delay */
1755 uasm_i_rfe(p); /* branch delay */
1756 uasm_l_r3000_write_probe_fail(l, *p);
1757 uasm_i_tlbwr(p); /* cp0 delay */
1759 uasm_i_rfe(p); /* branch delay */
1763 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1766 long pgdc = (long)pgd_current;
1768 uasm_i_mfc0(p, pte, C0_BADVADDR);
1769 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1770 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1771 uasm_i_srl(p, pte, pte, 22); /* load delay */
1772 uasm_i_sll(p, pte, pte, 2);
1773 uasm_i_addu(p, ptr, ptr, pte);
1774 uasm_i_mfc0(p, pte, C0_CONTEXT);
1775 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1776 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1777 uasm_i_addu(p, ptr, ptr, pte);
1778 uasm_i_lw(p, pte, 0, ptr);
1779 uasm_i_tlbp(p); /* load delay */
1782 static void build_r3000_tlb_load_handler(void)
1784 u32 *p = handle_tlbl;
1785 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1786 struct uasm_label *l = labels;
1787 struct uasm_reloc *r = relocs;
1789 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1790 memset(labels, 0, sizeof(labels));
1791 memset(relocs, 0, sizeof(relocs));
1793 build_r3000_tlbchange_handler_head(&p, K0, K1);
1794 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1795 uasm_i_nop(&p); /* load delay */
1796 build_make_valid(&p, &r, K0, K1);
1797 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1799 uasm_l_nopage_tlbl(&l, p);
1800 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1803 if (p >= handle_tlbl_end)
1804 panic("TLB load handler fastpath space exceeded");
1806 uasm_resolve_relocs(relocs, labels);
1807 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1808 (unsigned int)(p - handle_tlbl));
1810 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1813 static void build_r3000_tlb_store_handler(void)
1815 u32 *p = handle_tlbs;
1816 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1817 struct uasm_label *l = labels;
1818 struct uasm_reloc *r = relocs;
1820 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1821 memset(labels, 0, sizeof(labels));
1822 memset(relocs, 0, sizeof(relocs));
1824 build_r3000_tlbchange_handler_head(&p, K0, K1);
1825 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1826 uasm_i_nop(&p); /* load delay */
1827 build_make_write(&p, &r, K0, K1);
1828 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1830 uasm_l_nopage_tlbs(&l, p);
1831 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1834 if (p >= handle_tlbs_end)
1835 panic("TLB store handler fastpath space exceeded");
1837 uasm_resolve_relocs(relocs, labels);
1838 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1839 (unsigned int)(p - handle_tlbs));
1841 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1844 static void build_r3000_tlb_modify_handler(void)
1846 u32 *p = handle_tlbm;
1847 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1848 struct uasm_label *l = labels;
1849 struct uasm_reloc *r = relocs;
1851 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1852 memset(labels, 0, sizeof(labels));
1853 memset(relocs, 0, sizeof(relocs));
1855 build_r3000_tlbchange_handler_head(&p, K0, K1);
1856 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1857 uasm_i_nop(&p); /* load delay */
1858 build_make_write(&p, &r, K0, K1);
1859 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1861 uasm_l_nopage_tlbm(&l, p);
1862 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1865 if (p >= handle_tlbm_end)
1866 panic("TLB modify handler fastpath space exceeded");
1868 uasm_resolve_relocs(relocs, labels);
1869 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1870 (unsigned int)(p - handle_tlbm));
1872 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1874 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1877 * R4000 style TLB load/store/modify handlers.
1879 static struct work_registers
1880 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1881 struct uasm_reloc **r)
1883 struct work_registers wr = build_get_work_registers(p);
1886 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1888 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1891 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1893 * For huge tlb entries, pmd doesn't contain an address but
1894 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1895 * see if we need to jump to huge tlb processing.
1897 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1900 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1901 UASM_i_LW(p, wr.r2, 0, wr.r2);
1902 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1903 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1904 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1907 uasm_l_smp_pgtable_change(l, *p);
1909 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1910 if (!m4kc_tlbp_war()) {
1911 build_tlb_probe_entry(p);
1913 /* race condition happens, leaving */
1915 uasm_i_mfc0(p, wr.r3, C0_INDEX);
1916 uasm_il_bltz(p, r, wr.r3, label_leave);
1924 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1925 struct uasm_reloc **r, unsigned int tmp,
1928 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1929 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1930 build_update_entries(p, tmp, ptr);
1931 build_tlb_write_entry(p, l, r, tlb_indexed);
1932 uasm_l_leave(l, *p);
1933 build_restore_work_registers(p);
1934 uasm_i_eret(p); /* return from trap */
1937 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1941 static void build_r4000_tlb_load_handler(void)
1943 u32 *p = handle_tlbl;
1944 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1945 struct uasm_label *l = labels;
1946 struct uasm_reloc *r = relocs;
1947 struct work_registers wr;
1949 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1950 memset(labels, 0, sizeof(labels));
1951 memset(relocs, 0, sizeof(relocs));
1953 if (bcm1250_m3_war()) {
1954 unsigned int segbits = 44;
1956 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1957 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1958 uasm_i_xor(&p, K0, K0, K1);
1959 uasm_i_dsrl_safe(&p, K1, K0, 62);
1960 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1961 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1962 uasm_i_or(&p, K0, K0, K1);
1963 uasm_il_bnez(&p, &r, K0, label_leave);
1964 /* No need for uasm_i_nop */
1967 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1968 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1969 if (m4kc_tlbp_war())
1970 build_tlb_probe_entry(&p);
1972 if (cpu_has_rixi && !cpu_has_rixiex) {
1974 * If the page is not _PAGE_VALID, RI or XI could not
1975 * have triggered it. Skip the expensive test..
1977 if (use_bbit_insns()) {
1978 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1979 label_tlbl_goaround1);
1981 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1982 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1988 switch (current_cpu_type()) {
1990 if (cpu_has_mips_r2_exec_hazard) {
1993 case CPU_CAVIUM_OCTEON:
1994 case CPU_CAVIUM_OCTEON_PLUS:
1995 case CPU_CAVIUM_OCTEON2:
2000 /* Examine entrylo 0 or 1 based on ptr. */
2001 if (use_bbit_insns()) {
2002 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2004 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2005 uasm_i_beqz(&p, wr.r3, 8);
2007 /* load it in the delay slot*/
2008 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2009 /* load it if ptr is odd */
2010 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2012 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2013 * XI must have triggered it.
2015 if (use_bbit_insns()) {
2016 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2018 uasm_l_tlbl_goaround1(&l, p);
2020 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2021 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2024 uasm_l_tlbl_goaround1(&l, p);
2026 build_make_valid(&p, &r, wr.r1, wr.r2);
2027 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2029 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2031 * This is the entry point when build_r4000_tlbchange_handler_head
2032 * spots a huge page.
2034 uasm_l_tlb_huge_update(&l, p);
2035 iPTE_LW(&p, wr.r1, wr.r2);
2036 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2037 build_tlb_probe_entry(&p);
2039 if (cpu_has_rixi && !cpu_has_rixiex) {
2041 * If the page is not _PAGE_VALID, RI or XI could not
2042 * have triggered it. Skip the expensive test..
2044 if (use_bbit_insns()) {
2045 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2046 label_tlbl_goaround2);
2048 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2049 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2055 switch (current_cpu_type()) {
2057 if (cpu_has_mips_r2_exec_hazard) {
2060 case CPU_CAVIUM_OCTEON:
2061 case CPU_CAVIUM_OCTEON_PLUS:
2062 case CPU_CAVIUM_OCTEON2:
2067 /* Examine entrylo 0 or 1 based on ptr. */
2068 if (use_bbit_insns()) {
2069 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2071 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2072 uasm_i_beqz(&p, wr.r3, 8);
2074 /* load it in the delay slot*/
2075 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2076 /* load it if ptr is odd */
2077 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2079 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2080 * XI must have triggered it.
2082 if (use_bbit_insns()) {
2083 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2085 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2086 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2088 if (PM_DEFAULT_MASK == 0)
2091 * We clobbered C0_PAGEMASK, restore it. On the other branch
2092 * it is restored in build_huge_tlb_write_entry.
2094 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2096 uasm_l_tlbl_goaround2(&l, p);
2098 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2099 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2102 uasm_l_nopage_tlbl(&l, p);
2103 build_restore_work_registers(&p);
2104 #ifdef CONFIG_CPU_MICROMIPS
2105 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2106 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2107 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2111 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2114 if (p >= handle_tlbl_end)
2115 panic("TLB load handler fastpath space exceeded");
2117 uasm_resolve_relocs(relocs, labels);
2118 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2119 (unsigned int)(p - handle_tlbl));
2121 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2124 static void build_r4000_tlb_store_handler(void)
2126 u32 *p = handle_tlbs;
2127 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2128 struct uasm_label *l = labels;
2129 struct uasm_reloc *r = relocs;
2130 struct work_registers wr;
2132 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2133 memset(labels, 0, sizeof(labels));
2134 memset(relocs, 0, sizeof(relocs));
2136 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2137 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2138 if (m4kc_tlbp_war())
2139 build_tlb_probe_entry(&p);
2140 build_make_write(&p, &r, wr.r1, wr.r2);
2141 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2143 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2145 * This is the entry point when
2146 * build_r4000_tlbchange_handler_head spots a huge page.
2148 uasm_l_tlb_huge_update(&l, p);
2149 iPTE_LW(&p, wr.r1, wr.r2);
2150 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2151 build_tlb_probe_entry(&p);
2152 uasm_i_ori(&p, wr.r1, wr.r1,
2153 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2154 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2157 uasm_l_nopage_tlbs(&l, p);
2158 build_restore_work_registers(&p);
2159 #ifdef CONFIG_CPU_MICROMIPS
2160 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2161 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2162 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2166 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2169 if (p >= handle_tlbs_end)
2170 panic("TLB store handler fastpath space exceeded");
2172 uasm_resolve_relocs(relocs, labels);
2173 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2174 (unsigned int)(p - handle_tlbs));
2176 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2179 static void build_r4000_tlb_modify_handler(void)
2181 u32 *p = handle_tlbm;
2182 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2183 struct uasm_label *l = labels;
2184 struct uasm_reloc *r = relocs;
2185 struct work_registers wr;
2187 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2188 memset(labels, 0, sizeof(labels));
2189 memset(relocs, 0, sizeof(relocs));
2191 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2192 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2193 if (m4kc_tlbp_war())
2194 build_tlb_probe_entry(&p);
2195 /* Present and writable bits set, set accessed and dirty bits. */
2196 build_make_write(&p, &r, wr.r1, wr.r2);
2197 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2199 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2201 * This is the entry point when
2202 * build_r4000_tlbchange_handler_head spots a huge page.
2204 uasm_l_tlb_huge_update(&l, p);
2205 iPTE_LW(&p, wr.r1, wr.r2);
2206 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2207 build_tlb_probe_entry(&p);
2208 uasm_i_ori(&p, wr.r1, wr.r1,
2209 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2210 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2213 uasm_l_nopage_tlbm(&l, p);
2214 build_restore_work_registers(&p);
2215 #ifdef CONFIG_CPU_MICROMIPS
2216 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2217 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2218 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2222 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2225 if (p >= handle_tlbm_end)
2226 panic("TLB modify handler fastpath space exceeded");
2228 uasm_resolve_relocs(relocs, labels);
2229 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2230 (unsigned int)(p - handle_tlbm));
2232 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2235 static void flush_tlb_handlers(void)
2237 local_flush_icache_range((unsigned long)handle_tlbl,
2238 (unsigned long)handle_tlbl_end);
2239 local_flush_icache_range((unsigned long)handle_tlbs,
2240 (unsigned long)handle_tlbs_end);
2241 local_flush_icache_range((unsigned long)handle_tlbm,
2242 (unsigned long)handle_tlbm_end);
2243 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2244 (unsigned long)tlbmiss_handler_setup_pgd_end);
2247 static void print_htw_config(void)
2249 unsigned long config;
2251 const int field = 2 * sizeof(unsigned long);
2253 config = read_c0_pwfield();
2254 pr_debug("PWField (0x%0*lx): GDI: 0x%02lx UDI: 0x%02lx MDI: 0x%02lx PTI: 0x%02lx PTEI: 0x%02lx\n",
2256 (config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2257 (config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2258 (config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2259 (config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2260 (config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2262 config = read_c0_pwsize();
2263 pr_debug("PWSize (0x%0*lx): GDW: 0x%02lx UDW: 0x%02lx MDW: 0x%02lx PTW: 0x%02lx PTEW: 0x%02lx\n",
2265 (config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2266 (config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2267 (config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2268 (config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2269 (config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2271 pwctl = read_c0_pwctl();
2272 pr_debug("PWCtl (0x%x): PWEn: 0x%x DPH: 0x%x HugePg: 0x%x Psn: 0x%x\n",
2274 (pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2275 (pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2276 (pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2277 (pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2280 static void config_htw_params(void)
2282 unsigned long pwfield, pwsize, ptei;
2283 unsigned int config;
2286 * We are using 2-level page tables, so we only need to
2287 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2288 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2289 * write values less than 0xc in these fields because the entire
2290 * write will be dropped. As a result of which, we must preserve
2291 * the original reset values and overwrite only what we really want.
2294 pwfield = read_c0_pwfield();
2295 /* re-initialize the GDI field */
2296 pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2297 pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2298 /* re-initialize the PTI field including the even/odd bit */
2299 pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2300 pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2301 /* Set the PTEI right shift */
2302 ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2304 write_c0_pwfield(pwfield);
2305 /* Check whether the PTEI value is supported */
2306 back_to_back_c0_hazard();
2307 pwfield = read_c0_pwfield();
2308 if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2310 pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2313 * Drop option to avoid HTW being enabled via another path
2316 current_cpu_data.options &= ~MIPS_CPU_HTW;
2320 pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2321 pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2323 /* If XPA has been enabled, PTEs are 64-bit in size. */
2324 if (read_c0_pagegrain() & PG_ELPA)
2327 write_c0_pwsize(pwsize);
2329 /* Make sure everything is set before we enable the HTW */
2330 back_to_back_c0_hazard();
2332 /* Enable HTW and disable the rest of the pwctl fields */
2333 config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2334 write_c0_pwctl(config);
2335 pr_info("Hardware Page Table Walker enabled\n");
2340 static void config_xpa_params(void)
2343 unsigned int pagegrain;
2345 if (mips_xpa_disabled) {
2346 pr_info("Extended Physical Addressing (XPA) disabled\n");
2350 pagegrain = read_c0_pagegrain();
2351 write_c0_pagegrain(pagegrain | PG_ELPA);
2352 back_to_back_c0_hazard();
2353 pagegrain = read_c0_pagegrain();
2355 if (pagegrain & PG_ELPA)
2356 pr_info("Extended Physical Addressing (XPA) enabled\n");
2358 panic("Extended Physical Addressing (XPA) disabled");
2362 void build_tlb_refill_handler(void)
2365 * The refill handler is generated per-CPU, multi-node systems
2366 * may have local storage for it. The other handlers are only
2369 static int run_once = 0;
2371 output_pgtable_bits_defines();
2374 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2377 switch (current_cpu_type()) {
2385 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2386 if (cpu_has_local_ebase)
2387 build_r3000_tlb_refill_handler();
2389 if (!cpu_has_local_ebase)
2390 build_r3000_tlb_refill_handler();
2392 build_r3000_tlb_load_handler();
2393 build_r3000_tlb_store_handler();
2394 build_r3000_tlb_modify_handler();
2395 flush_tlb_handlers();
2399 panic("No R3000 TLB refill handler");
2405 panic("No R6000 TLB refill handler yet");
2409 panic("No R8000 TLB refill handler yet");
2414 scratch_reg = allocate_kscratch();
2416 build_r4000_tlb_load_handler();
2417 build_r4000_tlb_store_handler();
2418 build_r4000_tlb_modify_handler();
2419 if (!cpu_has_local_ebase)
2420 build_r4000_tlb_refill_handler();
2421 flush_tlb_handlers();
2424 if (cpu_has_local_ebase)
2425 build_r4000_tlb_refill_handler();
2427 config_xpa_params();
2429 config_htw_params();