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.
13 * ... and the days got worse and worse and now you see
14 * I've gone completly out of my mind.
16 * They're coming to take me a away haha
17 * they're coming to take me a away hoho hihi haha
18 * to the funny farm where code is beautiful all the time ...
20 * (Condolences to Napoleon XIV)
23 #include <linux/bug.h>
24 #include <linux/kernel.h>
25 #include <linux/types.h>
26 #include <linux/smp.h>
27 #include <linux/string.h>
28 #include <linux/init.h>
30 #include <asm/mmu_context.h>
35 static inline int r45k_bvahwbug(void)
37 /* XXX: We should probe for the presence of this bug, but we don't. */
41 static inline int r4k_250MHZhwbug(void)
43 /* XXX: We should probe for the presence of this bug, but we don't. */
47 static inline int __maybe_unused bcm1250_m3_war(void)
49 return BCM1250_M3_WAR;
52 static inline int __maybe_unused r10000_llsc_war(void)
54 return R10000_LLSC_WAR;
58 * Found by experiment: At least some revisions of the 4kc throw under
59 * some circumstances a machine check exception, triggered by invalid
60 * values in the index register. Delaying the tlbp instruction until
61 * after the next branch, plus adding an additional nop in front of
62 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
63 * why; it's not an issue caused by the core RTL.
66 static int __cpuinit m4kc_tlbp_war(void)
68 return (current_cpu_data.processor_id & 0xffff00) ==
69 (PRID_COMP_MIPS | PRID_IMP_4KC);
72 /* Handle labels (which must be positive integers). */
74 label_second_part = 1,
83 label_smp_pgtable_change,
84 label_r3000_write_probe_fail,
85 #ifdef CONFIG_HUGETLB_PAGE
86 label_tlb_huge_update,
90 UASM_L_LA(_second_part)
93 UASM_L_LA(_vmalloc_done)
94 UASM_L_LA(_tlbw_hazard)
96 UASM_L_LA(_nopage_tlbl)
97 UASM_L_LA(_nopage_tlbs)
98 UASM_L_LA(_nopage_tlbm)
99 UASM_L_LA(_smp_pgtable_change)
100 UASM_L_LA(_r3000_write_probe_fail)
101 #ifdef CONFIG_HUGETLB_PAGE
102 UASM_L_LA(_tlb_huge_update)
106 * For debug purposes.
108 static inline void dump_handler(const u32 *handler, int count)
112 pr_debug("\t.set push\n");
113 pr_debug("\t.set noreorder\n");
115 for (i = 0; i < count; i++)
116 pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
118 pr_debug("\t.set pop\n");
121 /* The only general purpose registers allowed in TLB handlers. */
125 /* Some CP0 registers */
126 #define C0_INDEX 0, 0
127 #define C0_ENTRYLO0 2, 0
128 #define C0_TCBIND 2, 2
129 #define C0_ENTRYLO1 3, 0
130 #define C0_CONTEXT 4, 0
131 #define C0_PAGEMASK 5, 0
132 #define C0_BADVADDR 8, 0
133 #define C0_ENTRYHI 10, 0
135 #define C0_XCONTEXT 20, 0
138 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
140 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
143 /* The worst case length of the handler is around 18 instructions for
144 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
145 * Maximum space available is 32 instructions for R3000 and 64
146 * instructions for R4000.
148 * We deliberately chose a buffer size of 128, so we won't scribble
149 * over anything important on overflow before we panic.
151 static u32 tlb_handler[128] __cpuinitdata;
153 /* simply assume worst case size for labels and relocs */
154 static struct uasm_label labels[128] __cpuinitdata;
155 static struct uasm_reloc relocs[128] __cpuinitdata;
157 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
159 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
160 * we cannot do r3000 under these circumstances.
164 * The R3000 TLB handler is simple.
166 static void __cpuinit build_r3000_tlb_refill_handler(void)
168 long pgdc = (long)pgd_current;
171 memset(tlb_handler, 0, sizeof(tlb_handler));
174 uasm_i_mfc0(&p, K0, C0_BADVADDR);
175 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
176 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
177 uasm_i_srl(&p, K0, K0, 22); /* load delay */
178 uasm_i_sll(&p, K0, K0, 2);
179 uasm_i_addu(&p, K1, K1, K0);
180 uasm_i_mfc0(&p, K0, C0_CONTEXT);
181 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
182 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
183 uasm_i_addu(&p, K1, K1, K0);
184 uasm_i_lw(&p, K0, 0, K1);
185 uasm_i_nop(&p); /* load delay */
186 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
187 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
188 uasm_i_tlbwr(&p); /* cp0 delay */
190 uasm_i_rfe(&p); /* branch delay */
192 if (p > tlb_handler + 32)
193 panic("TLB refill handler space exceeded");
195 pr_debug("Wrote TLB refill handler (%u instructions).\n",
196 (unsigned int)(p - tlb_handler));
198 memcpy((void *)ebase, tlb_handler, 0x80);
200 dump_handler((u32 *)ebase, 32);
202 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
205 * The R4000 TLB handler is much more complicated. We have two
206 * consecutive handler areas with 32 instructions space each.
207 * Since they aren't used at the same time, we can overflow in the
208 * other one.To keep things simple, we first assume linear space,
209 * then we relocate it to the final handler layout as needed.
211 static u32 final_handler[64] __cpuinitdata;
216 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
217 * 2. A timing hazard exists for the TLBP instruction.
219 * stalling_instruction
222 * The JTLB is being read for the TLBP throughout the stall generated by the
223 * previous instruction. This is not really correct as the stalling instruction
224 * can modify the address used to access the JTLB. The failure symptom is that
225 * the TLBP instruction will use an address created for the stalling instruction
226 * and not the address held in C0_ENHI and thus report the wrong results.
228 * The software work-around is to not allow the instruction preceding the TLBP
229 * to stall - make it an NOP or some other instruction guaranteed not to stall.
231 * Errata 2 will not be fixed. This errata is also on the R5000.
233 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
235 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
237 switch (current_cpu_type()) {
238 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
255 * Write random or indexed TLB entry, and care about the hazards from
256 * the preceeding mtc0 and for the following eret.
258 enum tlb_write_entry { tlb_random, tlb_indexed };
260 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
261 struct uasm_reloc **r,
262 enum tlb_write_entry wmode)
264 void(*tlbw)(u32 **) = NULL;
267 case tlb_random: tlbw = uasm_i_tlbwr; break;
268 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
271 if (cpu_has_mips_r2) {
272 if (cpu_has_mips_r2_exec_hazard)
278 switch (current_cpu_type()) {
286 * This branch uses up a mtc0 hazard nop slot and saves
287 * two nops after the tlbw instruction.
289 uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
291 uasm_l_tlbw_hazard(l, *p);
337 uasm_i_nop(p); /* QED specifies 2 nops hazard */
339 * This branch uses up a mtc0 hazard nop slot and saves
340 * a nop after the tlbw instruction.
342 uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
344 uasm_l_tlbw_hazard(l, *p);
357 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
358 * use of the JTLB for instructions should not occur for 4
359 * cpu cycles and use for data translations should not occur
394 panic("No TLB refill handler yet (CPU type: %d)",
395 current_cpu_data.cputype);
400 #ifdef CONFIG_HUGETLB_PAGE
401 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
402 struct uasm_label **l,
403 struct uasm_reloc **r,
405 enum tlb_write_entry wmode)
407 /* Set huge page tlb entry size */
408 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
409 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
410 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
412 build_tlb_write_entry(p, l, r, wmode);
414 /* Reset default page size */
415 if (PM_DEFAULT_MASK >> 16) {
416 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
417 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
418 uasm_il_b(p, r, label_leave);
419 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
420 } else if (PM_DEFAULT_MASK) {
421 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
422 uasm_il_b(p, r, label_leave);
423 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
425 uasm_il_b(p, r, label_leave);
426 uasm_i_mtc0(p, 0, C0_PAGEMASK);
431 * Check if Huge PTE is present, if so then jump to LABEL.
433 static void __cpuinit
434 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
435 unsigned int pmd, int lid)
437 UASM_i_LW(p, tmp, 0, pmd);
438 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
439 uasm_il_bnez(p, r, tmp, lid);
442 static __cpuinit void build_huge_update_entries(u32 **p,
449 * A huge PTE describes an area the size of the
450 * configured huge page size. This is twice the
451 * of the large TLB entry size we intend to use.
452 * A TLB entry half the size of the configured
453 * huge page size is configured into entrylo0
454 * and entrylo1 to cover the contiguous huge PTE
457 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
459 /* We can clobber tmp. It isn't used after this.*/
461 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
463 UASM_i_SRL(p, pte, pte, 6); /* convert to entrylo */
464 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* load it */
465 /* convert to entrylo1 */
467 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
469 UASM_i_ADDU(p, pte, pte, tmp);
471 uasm_i_mtc0(p, pte, C0_ENTRYLO1); /* load it */
474 static __cpuinit void build_huge_handler_tail(u32 **p,
475 struct uasm_reloc **r,
476 struct uasm_label **l,
481 UASM_i_SC(p, pte, 0, ptr);
482 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
483 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
485 UASM_i_SW(p, pte, 0, ptr);
487 build_huge_update_entries(p, pte, ptr);
488 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed);
490 #endif /* CONFIG_HUGETLB_PAGE */
494 * TMP and PTR are scratch.
495 * TMP will be clobbered, PTR will hold the pmd entry.
497 static void __cpuinit
498 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
499 unsigned int tmp, unsigned int ptr)
501 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
502 long pgdc = (long)pgd_current;
505 * The vmalloc handling is not in the hotpath.
507 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
508 uasm_il_bltz(p, r, tmp, label_vmalloc);
509 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
511 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
513 * &pgd << 11 stored in CONTEXT [23..63].
515 UASM_i_MFC0(p, ptr, C0_CONTEXT);
516 uasm_i_dins(p, ptr, 0, 0, 23); /* Clear lower 23 bits of context. */
517 uasm_i_ori(p, ptr, ptr, 0x540); /* 1 0 1 0 1 << 6 xkphys cached */
518 uasm_i_drotr(p, ptr, ptr, 11);
519 #elif defined(CONFIG_SMP)
520 # ifdef CONFIG_MIPS_MT_SMTC
522 * SMTC uses TCBind value as "CPU" index
524 uasm_i_mfc0(p, ptr, C0_TCBIND);
525 uasm_i_dsrl(p, ptr, ptr, 19);
528 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
531 uasm_i_dmfc0(p, ptr, C0_CONTEXT);
532 uasm_i_dsrl(p, ptr, ptr, 23);
534 UASM_i_LA_mostly(p, tmp, pgdc);
535 uasm_i_daddu(p, ptr, ptr, tmp);
536 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
537 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
539 UASM_i_LA_mostly(p, ptr, pgdc);
540 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
543 uasm_l_vmalloc_done(l, *p);
545 if (PGDIR_SHIFT - 3 < 32) /* get pgd offset in bytes */
546 uasm_i_dsrl(p, tmp, tmp, PGDIR_SHIFT-3);
548 uasm_i_dsrl32(p, tmp, tmp, PGDIR_SHIFT - 3 - 32);
550 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
551 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
552 #ifndef __PAGETABLE_PMD_FOLDED
553 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
554 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
555 uasm_i_dsrl(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
556 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
557 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
562 * BVADDR is the faulting address, PTR is scratch.
563 * PTR will hold the pgd for vmalloc.
565 static void __cpuinit
566 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
567 unsigned int bvaddr, unsigned int ptr)
569 long swpd = (long)swapper_pg_dir;
571 uasm_l_vmalloc(l, *p);
573 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
574 uasm_il_b(p, r, label_vmalloc_done);
575 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
577 UASM_i_LA_mostly(p, ptr, swpd);
578 uasm_il_b(p, r, label_vmalloc_done);
579 if (uasm_in_compat_space_p(swpd))
580 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
582 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
586 #else /* !CONFIG_64BIT */
589 * TMP and PTR are scratch.
590 * TMP will be clobbered, PTR will hold the pgd entry.
592 static void __cpuinit __maybe_unused
593 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
595 long pgdc = (long)pgd_current;
597 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
599 #ifdef CONFIG_MIPS_MT_SMTC
601 * SMTC uses TCBind value as "CPU" index
603 uasm_i_mfc0(p, ptr, C0_TCBIND);
604 UASM_i_LA_mostly(p, tmp, pgdc);
605 uasm_i_srl(p, ptr, ptr, 19);
608 * smp_processor_id() << 3 is stored in CONTEXT.
610 uasm_i_mfc0(p, ptr, C0_CONTEXT);
611 UASM_i_LA_mostly(p, tmp, pgdc);
612 uasm_i_srl(p, ptr, ptr, 23);
614 uasm_i_addu(p, ptr, tmp, ptr);
616 UASM_i_LA_mostly(p, ptr, pgdc);
618 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
619 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
620 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
621 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
622 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
625 #endif /* !CONFIG_64BIT */
627 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
629 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
630 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
632 switch (current_cpu_type()) {
649 UASM_i_SRL(p, ctx, ctx, shift);
650 uasm_i_andi(p, ctx, ctx, mask);
653 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
656 * Bug workaround for the Nevada. It seems as if under certain
657 * circumstances the move from cp0_context might produce a
658 * bogus result when the mfc0 instruction and its consumer are
659 * in a different cacheline or a load instruction, probably any
660 * memory reference, is between them.
662 switch (current_cpu_type()) {
664 UASM_i_LW(p, ptr, 0, ptr);
665 GET_CONTEXT(p, tmp); /* get context reg */
669 GET_CONTEXT(p, tmp); /* get context reg */
670 UASM_i_LW(p, ptr, 0, ptr);
674 build_adjust_context(p, tmp);
675 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
678 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
682 * 64bit address support (36bit on a 32bit CPU) in a 32bit
683 * Kernel is a special case. Only a few CPUs use it.
685 #ifdef CONFIG_64BIT_PHYS_ADDR
686 if (cpu_has_64bits) {
687 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
688 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
689 uasm_i_dsrl(p, tmp, tmp, 6); /* convert to entrylo0 */
690 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
691 uasm_i_dsrl(p, ptep, ptep, 6); /* convert to entrylo1 */
692 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
694 int pte_off_even = sizeof(pte_t) / 2;
695 int pte_off_odd = pte_off_even + sizeof(pte_t);
697 /* The pte entries are pre-shifted */
698 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
699 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
700 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
701 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
704 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
705 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
707 build_tlb_probe_entry(p);
708 UASM_i_SRL(p, tmp, tmp, 6); /* convert to entrylo0 */
709 if (r4k_250MHZhwbug())
710 uasm_i_mtc0(p, 0, C0_ENTRYLO0);
711 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
712 UASM_i_SRL(p, ptep, ptep, 6); /* convert to entrylo1 */
714 uasm_i_mfc0(p, tmp, C0_INDEX);
715 if (r4k_250MHZhwbug())
716 uasm_i_mtc0(p, 0, C0_ENTRYLO1);
717 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
722 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
723 * because EXL == 0. If we wrap, we can also use the 32 instruction
724 * slots before the XTLB refill exception handler which belong to the
725 * unused TLB refill exception.
727 #define MIPS64_REFILL_INSNS 32
729 static void __cpuinit build_r4000_tlb_refill_handler(void)
731 u32 *p = tlb_handler;
732 struct uasm_label *l = labels;
733 struct uasm_reloc *r = relocs;
735 unsigned int final_len;
737 memset(tlb_handler, 0, sizeof(tlb_handler));
738 memset(labels, 0, sizeof(labels));
739 memset(relocs, 0, sizeof(relocs));
740 memset(final_handler, 0, sizeof(final_handler));
743 * create the plain linear handler
745 if (bcm1250_m3_war()) {
746 UASM_i_MFC0(&p, K0, C0_BADVADDR);
747 UASM_i_MFC0(&p, K1, C0_ENTRYHI);
748 uasm_i_xor(&p, K0, K0, K1);
749 UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
750 uasm_il_bnez(&p, &r, K0, label_leave);
751 /* No need for uasm_i_nop */
755 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
757 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
760 #ifdef CONFIG_HUGETLB_PAGE
761 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
764 build_get_ptep(&p, K0, K1);
765 build_update_entries(&p, K0, K1);
766 build_tlb_write_entry(&p, &l, &r, tlb_random);
768 uasm_i_eret(&p); /* return from trap */
770 #ifdef CONFIG_HUGETLB_PAGE
771 uasm_l_tlb_huge_update(&l, p);
772 UASM_i_LW(&p, K0, 0, K1);
773 build_huge_update_entries(&p, K0, K1);
774 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random);
778 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1);
782 * Overflow check: For the 64bit handler, we need at least one
783 * free instruction slot for the wrap-around branch. In worst
784 * case, if the intended insertion point is a delay slot, we
785 * need three, with the second nop'ed and the third being
788 /* Loongson2 ebase is different than r4k, we have more space */
789 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
790 if ((p - tlb_handler) > 64)
791 panic("TLB refill handler space exceeded");
793 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
794 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
795 && uasm_insn_has_bdelay(relocs,
796 tlb_handler + MIPS64_REFILL_INSNS - 3)))
797 panic("TLB refill handler space exceeded");
801 * Now fold the handler in the TLB refill handler space.
803 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
805 /* Simplest case, just copy the handler. */
806 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
807 final_len = p - tlb_handler;
808 #else /* CONFIG_64BIT */
809 f = final_handler + MIPS64_REFILL_INSNS;
810 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
811 /* Just copy the handler. */
812 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
813 final_len = p - tlb_handler;
815 #if defined(CONFIG_HUGETLB_PAGE)
816 const enum label_id ls = label_tlb_huge_update;
818 const enum label_id ls = label_vmalloc;
824 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
826 BUG_ON(i == ARRAY_SIZE(labels));
827 split = labels[i].addr;
830 * See if we have overflown one way or the other.
832 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
833 split < p - MIPS64_REFILL_INSNS)
838 * Split two instructions before the end. One
839 * for the branch and one for the instruction
842 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
845 * If the branch would fall in a delay slot,
846 * we must back up an additional instruction
847 * so that it is no longer in a delay slot.
849 if (uasm_insn_has_bdelay(relocs, split - 1))
852 /* Copy first part of the handler. */
853 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
854 f += split - tlb_handler;
858 uasm_l_split(&l, final_handler);
859 uasm_il_b(&f, &r, label_split);
860 if (uasm_insn_has_bdelay(relocs, split))
863 uasm_copy_handler(relocs, labels,
864 split, split + 1, f);
865 uasm_move_labels(labels, f, f + 1, -1);
871 /* Copy the rest of the handler. */
872 uasm_copy_handler(relocs, labels, split, p, final_handler);
873 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
876 #endif /* CONFIG_64BIT */
878 uasm_resolve_relocs(relocs, labels);
879 pr_debug("Wrote TLB refill handler (%u instructions).\n",
882 memcpy((void *)ebase, final_handler, 0x100);
884 dump_handler((u32 *)ebase, 64);
888 * TLB load/store/modify handlers.
890 * Only the fastpath gets synthesized at runtime, the slowpath for
891 * do_page_fault remains normal asm.
893 extern void tlb_do_page_fault_0(void);
894 extern void tlb_do_page_fault_1(void);
897 * 128 instructions for the fastpath handler is generous and should
900 #define FASTPATH_SIZE 128
902 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
903 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
904 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
906 static void __cpuinit
907 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
910 # ifdef CONFIG_64BIT_PHYS_ADDR
912 uasm_i_lld(p, pte, 0, ptr);
915 UASM_i_LL(p, pte, 0, ptr);
917 # ifdef CONFIG_64BIT_PHYS_ADDR
919 uasm_i_ld(p, pte, 0, ptr);
922 UASM_i_LW(p, pte, 0, ptr);
926 static void __cpuinit
927 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
930 #ifdef CONFIG_64BIT_PHYS_ADDR
931 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
934 uasm_i_ori(p, pte, pte, mode);
936 # ifdef CONFIG_64BIT_PHYS_ADDR
938 uasm_i_scd(p, pte, 0, ptr);
941 UASM_i_SC(p, pte, 0, ptr);
943 if (r10000_llsc_war())
944 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
946 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
948 # ifdef CONFIG_64BIT_PHYS_ADDR
949 if (!cpu_has_64bits) {
950 /* no uasm_i_nop needed */
951 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
952 uasm_i_ori(p, pte, pte, hwmode);
953 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
954 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
955 /* no uasm_i_nop needed */
956 uasm_i_lw(p, pte, 0, ptr);
963 # ifdef CONFIG_64BIT_PHYS_ADDR
965 uasm_i_sd(p, pte, 0, ptr);
968 UASM_i_SW(p, pte, 0, ptr);
970 # ifdef CONFIG_64BIT_PHYS_ADDR
971 if (!cpu_has_64bits) {
972 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
973 uasm_i_ori(p, pte, pte, hwmode);
974 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
975 uasm_i_lw(p, pte, 0, ptr);
982 * Check if PTE is present, if not then jump to LABEL. PTR points to
983 * the page table where this PTE is located, PTE will be re-loaded
984 * with it's original value.
986 static void __cpuinit
987 build_pte_present(u32 **p, struct uasm_reloc **r,
988 unsigned int pte, unsigned int ptr, enum label_id lid)
990 uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
991 uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
992 uasm_il_bnez(p, r, pte, lid);
993 iPTE_LW(p, pte, ptr);
996 /* Make PTE valid, store result in PTR. */
997 static void __cpuinit
998 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1001 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1003 iPTE_SW(p, r, pte, ptr, mode);
1007 * Check if PTE can be written to, if not branch to LABEL. Regardless
1008 * restore PTE with value from PTR when done.
1010 static void __cpuinit
1011 build_pte_writable(u32 **p, struct uasm_reloc **r,
1012 unsigned int pte, unsigned int ptr, enum label_id lid)
1014 uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1015 uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1016 uasm_il_bnez(p, r, pte, lid);
1017 iPTE_LW(p, pte, ptr);
1020 /* Make PTE writable, update software status bits as well, then store
1023 static void __cpuinit
1024 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1027 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1030 iPTE_SW(p, r, pte, ptr, mode);
1034 * Check if PTE can be modified, if not branch to LABEL. Regardless
1035 * restore PTE with value from PTR when done.
1037 static void __cpuinit
1038 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1039 unsigned int pte, unsigned int ptr, enum label_id lid)
1041 uasm_i_andi(p, pte, pte, _PAGE_WRITE);
1042 uasm_il_beqz(p, r, pte, lid);
1043 iPTE_LW(p, pte, ptr);
1046 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1048 * R3000 style TLB load/store/modify handlers.
1052 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1055 static void __cpuinit
1056 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1058 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1059 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1062 uasm_i_rfe(p); /* branch delay */
1066 * This places the pte into ENTRYLO0 and writes it with tlbwi
1067 * or tlbwr as appropriate. This is because the index register
1068 * may have the probe fail bit set as a result of a trap on a
1069 * kseg2 access, i.e. without refill. Then it returns.
1071 static void __cpuinit
1072 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1073 struct uasm_reloc **r, unsigned int pte,
1076 uasm_i_mfc0(p, tmp, C0_INDEX);
1077 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1078 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1079 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1080 uasm_i_tlbwi(p); /* cp0 delay */
1082 uasm_i_rfe(p); /* branch delay */
1083 uasm_l_r3000_write_probe_fail(l, *p);
1084 uasm_i_tlbwr(p); /* cp0 delay */
1086 uasm_i_rfe(p); /* branch delay */
1089 static void __cpuinit
1090 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1093 long pgdc = (long)pgd_current;
1095 uasm_i_mfc0(p, pte, C0_BADVADDR);
1096 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1097 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1098 uasm_i_srl(p, pte, pte, 22); /* load delay */
1099 uasm_i_sll(p, pte, pte, 2);
1100 uasm_i_addu(p, ptr, ptr, pte);
1101 uasm_i_mfc0(p, pte, C0_CONTEXT);
1102 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1103 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1104 uasm_i_addu(p, ptr, ptr, pte);
1105 uasm_i_lw(p, pte, 0, ptr);
1106 uasm_i_tlbp(p); /* load delay */
1109 static void __cpuinit build_r3000_tlb_load_handler(void)
1111 u32 *p = handle_tlbl;
1112 struct uasm_label *l = labels;
1113 struct uasm_reloc *r = relocs;
1115 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1116 memset(labels, 0, sizeof(labels));
1117 memset(relocs, 0, sizeof(relocs));
1119 build_r3000_tlbchange_handler_head(&p, K0, K1);
1120 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1121 uasm_i_nop(&p); /* load delay */
1122 build_make_valid(&p, &r, K0, K1);
1123 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1125 uasm_l_nopage_tlbl(&l, p);
1126 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1129 if ((p - handle_tlbl) > FASTPATH_SIZE)
1130 panic("TLB load handler fastpath space exceeded");
1132 uasm_resolve_relocs(relocs, labels);
1133 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1134 (unsigned int)(p - handle_tlbl));
1136 dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1139 static void __cpuinit build_r3000_tlb_store_handler(void)
1141 u32 *p = handle_tlbs;
1142 struct uasm_label *l = labels;
1143 struct uasm_reloc *r = relocs;
1145 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1146 memset(labels, 0, sizeof(labels));
1147 memset(relocs, 0, sizeof(relocs));
1149 build_r3000_tlbchange_handler_head(&p, K0, K1);
1150 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1151 uasm_i_nop(&p); /* load delay */
1152 build_make_write(&p, &r, K0, K1);
1153 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1155 uasm_l_nopage_tlbs(&l, p);
1156 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1159 if ((p - handle_tlbs) > FASTPATH_SIZE)
1160 panic("TLB store handler fastpath space exceeded");
1162 uasm_resolve_relocs(relocs, labels);
1163 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1164 (unsigned int)(p - handle_tlbs));
1166 dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1169 static void __cpuinit build_r3000_tlb_modify_handler(void)
1171 u32 *p = handle_tlbm;
1172 struct uasm_label *l = labels;
1173 struct uasm_reloc *r = relocs;
1175 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1176 memset(labels, 0, sizeof(labels));
1177 memset(relocs, 0, sizeof(relocs));
1179 build_r3000_tlbchange_handler_head(&p, K0, K1);
1180 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1181 uasm_i_nop(&p); /* load delay */
1182 build_make_write(&p, &r, K0, K1);
1183 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1185 uasm_l_nopage_tlbm(&l, p);
1186 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1189 if ((p - handle_tlbm) > FASTPATH_SIZE)
1190 panic("TLB modify handler fastpath space exceeded");
1192 uasm_resolve_relocs(relocs, labels);
1193 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1194 (unsigned int)(p - handle_tlbm));
1196 dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1198 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1201 * R4000 style TLB load/store/modify handlers.
1203 static void __cpuinit
1204 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1205 struct uasm_reloc **r, unsigned int pte,
1209 build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1211 build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1214 #ifdef CONFIG_HUGETLB_PAGE
1216 * For huge tlb entries, pmd doesn't contain an address but
1217 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1218 * see if we need to jump to huge tlb processing.
1220 build_is_huge_pte(p, r, pte, ptr, label_tlb_huge_update);
1223 UASM_i_MFC0(p, pte, C0_BADVADDR);
1224 UASM_i_LW(p, ptr, 0, ptr);
1225 UASM_i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1226 uasm_i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1227 UASM_i_ADDU(p, ptr, ptr, pte);
1230 uasm_l_smp_pgtable_change(l, *p);
1232 iPTE_LW(p, pte, ptr); /* get even pte */
1233 if (!m4kc_tlbp_war())
1234 build_tlb_probe_entry(p);
1237 static void __cpuinit
1238 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1239 struct uasm_reloc **r, unsigned int tmp,
1242 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1243 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1244 build_update_entries(p, tmp, ptr);
1245 build_tlb_write_entry(p, l, r, tlb_indexed);
1246 uasm_l_leave(l, *p);
1247 uasm_i_eret(p); /* return from trap */
1250 build_get_pgd_vmalloc64(p, l, r, tmp, ptr);
1254 static void __cpuinit build_r4000_tlb_load_handler(void)
1256 u32 *p = handle_tlbl;
1257 struct uasm_label *l = labels;
1258 struct uasm_reloc *r = relocs;
1260 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1261 memset(labels, 0, sizeof(labels));
1262 memset(relocs, 0, sizeof(relocs));
1264 if (bcm1250_m3_war()) {
1265 UASM_i_MFC0(&p, K0, C0_BADVADDR);
1266 UASM_i_MFC0(&p, K1, C0_ENTRYHI);
1267 uasm_i_xor(&p, K0, K0, K1);
1268 UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
1269 uasm_il_bnez(&p, &r, K0, label_leave);
1270 /* No need for uasm_i_nop */
1273 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1274 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1275 if (m4kc_tlbp_war())
1276 build_tlb_probe_entry(&p);
1277 build_make_valid(&p, &r, K0, K1);
1278 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1280 #ifdef CONFIG_HUGETLB_PAGE
1282 * This is the entry point when build_r4000_tlbchange_handler_head
1283 * spots a huge page.
1285 uasm_l_tlb_huge_update(&l, p);
1286 iPTE_LW(&p, K0, K1);
1287 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1288 build_tlb_probe_entry(&p);
1289 uasm_i_ori(&p, K0, K0, (_PAGE_ACCESSED | _PAGE_VALID));
1290 build_huge_handler_tail(&p, &r, &l, K0, K1);
1293 uasm_l_nopage_tlbl(&l, p);
1294 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1297 if ((p - handle_tlbl) > FASTPATH_SIZE)
1298 panic("TLB load handler fastpath space exceeded");
1300 uasm_resolve_relocs(relocs, labels);
1301 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1302 (unsigned int)(p - handle_tlbl));
1304 dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1307 static void __cpuinit build_r4000_tlb_store_handler(void)
1309 u32 *p = handle_tlbs;
1310 struct uasm_label *l = labels;
1311 struct uasm_reloc *r = relocs;
1313 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1314 memset(labels, 0, sizeof(labels));
1315 memset(relocs, 0, sizeof(relocs));
1317 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1318 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1319 if (m4kc_tlbp_war())
1320 build_tlb_probe_entry(&p);
1321 build_make_write(&p, &r, K0, K1);
1322 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1324 #ifdef CONFIG_HUGETLB_PAGE
1326 * This is the entry point when
1327 * build_r4000_tlbchange_handler_head spots a huge page.
1329 uasm_l_tlb_huge_update(&l, p);
1330 iPTE_LW(&p, K0, K1);
1331 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1332 build_tlb_probe_entry(&p);
1333 uasm_i_ori(&p, K0, K0,
1334 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1335 build_huge_handler_tail(&p, &r, &l, K0, K1);
1338 uasm_l_nopage_tlbs(&l, p);
1339 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1342 if ((p - handle_tlbs) > FASTPATH_SIZE)
1343 panic("TLB store handler fastpath space exceeded");
1345 uasm_resolve_relocs(relocs, labels);
1346 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1347 (unsigned int)(p - handle_tlbs));
1349 dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1352 static void __cpuinit build_r4000_tlb_modify_handler(void)
1354 u32 *p = handle_tlbm;
1355 struct uasm_label *l = labels;
1356 struct uasm_reloc *r = relocs;
1358 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1359 memset(labels, 0, sizeof(labels));
1360 memset(relocs, 0, sizeof(relocs));
1362 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1363 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1364 if (m4kc_tlbp_war())
1365 build_tlb_probe_entry(&p);
1366 /* Present and writable bits set, set accessed and dirty bits. */
1367 build_make_write(&p, &r, K0, K1);
1368 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1370 #ifdef CONFIG_HUGETLB_PAGE
1372 * This is the entry point when
1373 * build_r4000_tlbchange_handler_head spots a huge page.
1375 uasm_l_tlb_huge_update(&l, p);
1376 iPTE_LW(&p, K0, K1);
1377 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1378 build_tlb_probe_entry(&p);
1379 uasm_i_ori(&p, K0, K0,
1380 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1381 build_huge_handler_tail(&p, &r, &l, K0, K1);
1384 uasm_l_nopage_tlbm(&l, p);
1385 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1388 if ((p - handle_tlbm) > FASTPATH_SIZE)
1389 panic("TLB modify handler fastpath space exceeded");
1391 uasm_resolve_relocs(relocs, labels);
1392 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1393 (unsigned int)(p - handle_tlbm));
1395 dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1398 void __cpuinit build_tlb_refill_handler(void)
1401 * The refill handler is generated per-CPU, multi-node systems
1402 * may have local storage for it. The other handlers are only
1405 static int run_once = 0;
1407 switch (current_cpu_type()) {
1415 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1416 build_r3000_tlb_refill_handler();
1418 build_r3000_tlb_load_handler();
1419 build_r3000_tlb_store_handler();
1420 build_r3000_tlb_modify_handler();
1424 panic("No R3000 TLB refill handler");
1430 panic("No R6000 TLB refill handler yet");
1434 panic("No R8000 TLB refill handler yet");
1438 build_r4000_tlb_refill_handler();
1440 build_r4000_tlb_load_handler();
1441 build_r4000_tlb_store_handler();
1442 build_r4000_tlb_modify_handler();
1448 void __cpuinit flush_tlb_handlers(void)
1450 local_flush_icache_range((unsigned long)handle_tlbl,
1451 (unsigned long)handle_tlbl + sizeof(handle_tlbl));
1452 local_flush_icache_range((unsigned long)handle_tlbs,
1453 (unsigned long)handle_tlbs + sizeof(handle_tlbs));
1454 local_flush_icache_range((unsigned long)handle_tlbm,
1455 (unsigned long)handle_tlbm + sizeof(handle_tlbm));