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>
34 static inline int r45k_bvahwbug(void)
36 /* XXX: We should probe for the presence of this bug, but we don't. */
40 static inline int r4k_250MHZhwbug(void)
42 /* XXX: We should probe for the presence of this bug, but we don't. */
46 static inline int __maybe_unused bcm1250_m3_war(void)
48 return BCM1250_M3_WAR;
51 static inline int __maybe_unused r10000_llsc_war(void)
53 return R10000_LLSC_WAR;
57 * Found by experiment: At least some revisions of the 4kc throw under
58 * some circumstances a machine check exception, triggered by invalid
59 * values in the index register. Delaying the tlbp instruction until
60 * after the next branch, plus adding an additional nop in front of
61 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
62 * why; it's not an issue caused by the core RTL.
65 static int __cpuinit m4kc_tlbp_war(void)
67 return (current_cpu_data.processor_id & 0xffff00) ==
68 (PRID_COMP_MIPS | PRID_IMP_4KC);
71 /* Handle labels (which must be positive integers). */
73 label_second_part = 1,
82 label_smp_pgtable_change,
83 label_r3000_write_probe_fail,
84 #ifdef CONFIG_HUGETLB_PAGE
85 label_tlb_huge_update,
89 UASM_L_LA(_second_part)
92 UASM_L_LA(_vmalloc_done)
93 UASM_L_LA(_tlbw_hazard)
95 UASM_L_LA(_nopage_tlbl)
96 UASM_L_LA(_nopage_tlbs)
97 UASM_L_LA(_nopage_tlbm)
98 UASM_L_LA(_smp_pgtable_change)
99 UASM_L_LA(_r3000_write_probe_fail)
100 #ifdef CONFIG_HUGETLB_PAGE
101 UASM_L_LA(_tlb_huge_update)
105 * For debug purposes.
107 static inline void dump_handler(const u32 *handler, int count)
111 pr_debug("\t.set push\n");
112 pr_debug("\t.set noreorder\n");
114 for (i = 0; i < count; i++)
115 pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
117 pr_debug("\t.set pop\n");
120 /* The only general purpose registers allowed in TLB handlers. */
124 /* Some CP0 registers */
125 #define C0_INDEX 0, 0
126 #define C0_ENTRYLO0 2, 0
127 #define C0_TCBIND 2, 2
128 #define C0_ENTRYLO1 3, 0
129 #define C0_CONTEXT 4, 0
130 #define C0_PAGEMASK 5, 0
131 #define C0_BADVADDR 8, 0
132 #define C0_ENTRYHI 10, 0
134 #define C0_XCONTEXT 20, 0
137 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
139 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
142 /* The worst case length of the handler is around 18 instructions for
143 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
144 * Maximum space available is 32 instructions for R3000 and 64
145 * instructions for R4000.
147 * We deliberately chose a buffer size of 128, so we won't scribble
148 * over anything important on overflow before we panic.
150 static u32 tlb_handler[128] __cpuinitdata;
152 /* simply assume worst case size for labels and relocs */
153 static struct uasm_label labels[128] __cpuinitdata;
154 static struct uasm_reloc relocs[128] __cpuinitdata;
156 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
158 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
159 * we cannot do r3000 under these circumstances.
163 * The R3000 TLB handler is simple.
165 static void __cpuinit build_r3000_tlb_refill_handler(void)
167 long pgdc = (long)pgd_current;
170 memset(tlb_handler, 0, sizeof(tlb_handler));
173 uasm_i_mfc0(&p, K0, C0_BADVADDR);
174 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
175 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
176 uasm_i_srl(&p, K0, K0, 22); /* load delay */
177 uasm_i_sll(&p, K0, K0, 2);
178 uasm_i_addu(&p, K1, K1, K0);
179 uasm_i_mfc0(&p, K0, C0_CONTEXT);
180 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
181 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
182 uasm_i_addu(&p, K1, K1, K0);
183 uasm_i_lw(&p, K0, 0, K1);
184 uasm_i_nop(&p); /* load delay */
185 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
186 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
187 uasm_i_tlbwr(&p); /* cp0 delay */
189 uasm_i_rfe(&p); /* branch delay */
191 if (p > tlb_handler + 32)
192 panic("TLB refill handler space exceeded");
194 pr_debug("Wrote TLB refill handler (%u instructions).\n",
195 (unsigned int)(p - tlb_handler));
197 memcpy((void *)ebase, tlb_handler, 0x80);
199 dump_handler((u32 *)ebase, 32);
201 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
204 * The R4000 TLB handler is much more complicated. We have two
205 * consecutive handler areas with 32 instructions space each.
206 * Since they aren't used at the same time, we can overflow in the
207 * other one.To keep things simple, we first assume linear space,
208 * then we relocate it to the final handler layout as needed.
210 static u32 final_handler[64] __cpuinitdata;
215 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
216 * 2. A timing hazard exists for the TLBP instruction.
218 * stalling_instruction
221 * The JTLB is being read for the TLBP throughout the stall generated by the
222 * previous instruction. This is not really correct as the stalling instruction
223 * can modify the address used to access the JTLB. The failure symptom is that
224 * the TLBP instruction will use an address created for the stalling instruction
225 * and not the address held in C0_ENHI and thus report the wrong results.
227 * The software work-around is to not allow the instruction preceding the TLBP
228 * to stall - make it an NOP or some other instruction guaranteed not to stall.
230 * Errata 2 will not be fixed. This errata is also on the R5000.
232 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
234 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
236 switch (current_cpu_type()) {
237 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
254 * Write random or indexed TLB entry, and care about the hazards from
255 * the preceeding mtc0 and for the following eret.
257 enum tlb_write_entry { tlb_random, tlb_indexed };
259 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
260 struct uasm_reloc **r,
261 enum tlb_write_entry wmode)
263 void(*tlbw)(u32 **) = NULL;
266 case tlb_random: tlbw = uasm_i_tlbwr; break;
267 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
270 if (cpu_has_mips_r2) {
271 if (cpu_has_mips_r2_exec_hazard)
277 switch (current_cpu_type()) {
285 * This branch uses up a mtc0 hazard nop slot and saves
286 * two nops after the tlbw instruction.
288 uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
290 uasm_l_tlbw_hazard(l, *p);
336 uasm_i_nop(p); /* QED specifies 2 nops hazard */
338 * This branch uses up a mtc0 hazard nop slot and saves
339 * a nop after the tlbw instruction.
341 uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
343 uasm_l_tlbw_hazard(l, *p);
356 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
357 * use of the JTLB for instructions should not occur for 4
358 * cpu cycles and use for data translations should not occur
393 panic("No TLB refill handler yet (CPU type: %d)",
394 current_cpu_data.cputype);
399 #ifdef CONFIG_HUGETLB_PAGE
400 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
401 struct uasm_label **l,
402 struct uasm_reloc **r,
404 enum tlb_write_entry wmode)
406 /* Set huge page tlb entry size */
407 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
408 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
409 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
411 build_tlb_write_entry(p, l, r, wmode);
413 /* Reset default page size */
414 if (PM_DEFAULT_MASK >> 16) {
415 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
416 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
417 uasm_il_b(p, r, label_leave);
418 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
419 } else if (PM_DEFAULT_MASK) {
420 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
421 uasm_il_b(p, r, label_leave);
422 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
424 uasm_il_b(p, r, label_leave);
425 uasm_i_mtc0(p, 0, C0_PAGEMASK);
430 * Check if Huge PTE is present, if so then jump to LABEL.
432 static void __cpuinit
433 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
434 unsigned int pmd, int lid)
436 UASM_i_LW(p, tmp, 0, pmd);
437 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
438 uasm_il_bnez(p, r, tmp, lid);
441 static __cpuinit void build_huge_update_entries(u32 **p,
448 * A huge PTE describes an area the size of the
449 * configured huge page size. This is twice the
450 * of the large TLB entry size we intend to use.
451 * A TLB entry half the size of the configured
452 * huge page size is configured into entrylo0
453 * and entrylo1 to cover the contiguous huge PTE
456 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
458 /* We can clobber tmp. It isn't used after this.*/
460 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
462 UASM_i_SRL(p, pte, pte, 6); /* convert to entrylo */
463 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* load it */
464 /* convert to entrylo1 */
466 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
468 UASM_i_ADDU(p, pte, pte, tmp);
470 uasm_i_mtc0(p, pte, C0_ENTRYLO1); /* load it */
473 static __cpuinit void build_huge_handler_tail(u32 **p,
474 struct uasm_reloc **r,
475 struct uasm_label **l,
480 UASM_i_SC(p, pte, 0, ptr);
481 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
482 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
484 UASM_i_SW(p, pte, 0, ptr);
486 build_huge_update_entries(p, pte, ptr);
487 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed);
489 #endif /* CONFIG_HUGETLB_PAGE */
493 * TMP and PTR are scratch.
494 * TMP will be clobbered, PTR will hold the pmd entry.
496 static void __cpuinit
497 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
498 unsigned int tmp, unsigned int ptr)
500 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
501 long pgdc = (long)pgd_current;
504 * The vmalloc handling is not in the hotpath.
506 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
507 uasm_il_bltz(p, r, tmp, label_vmalloc);
508 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
510 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
512 * &pgd << 11 stored in CONTEXT [23..63].
514 UASM_i_MFC0(p, ptr, C0_CONTEXT);
515 uasm_i_dins(p, ptr, 0, 0, 23); /* Clear lower 23 bits of context. */
516 uasm_i_ori(p, ptr, ptr, 0x540); /* 1 0 1 0 1 << 6 xkphys cached */
517 uasm_i_drotr(p, ptr, ptr, 11);
518 #elif defined(CONFIG_SMP)
519 # ifdef CONFIG_MIPS_MT_SMTC
521 * SMTC uses TCBind value as "CPU" index
523 uasm_i_mfc0(p, ptr, C0_TCBIND);
524 uasm_i_dsrl(p, ptr, ptr, 19);
527 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
530 uasm_i_dmfc0(p, ptr, C0_CONTEXT);
531 uasm_i_dsrl(p, ptr, ptr, 23);
533 UASM_i_LA_mostly(p, tmp, pgdc);
534 uasm_i_daddu(p, ptr, ptr, tmp);
535 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
536 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
538 UASM_i_LA_mostly(p, ptr, pgdc);
539 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
542 uasm_l_vmalloc_done(l, *p);
544 if (PGDIR_SHIFT - 3 < 32) /* get pgd offset in bytes */
545 uasm_i_dsrl(p, tmp, tmp, PGDIR_SHIFT-3);
547 uasm_i_dsrl32(p, tmp, tmp, PGDIR_SHIFT - 3 - 32);
549 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
550 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
551 #ifndef __PAGETABLE_PMD_FOLDED
552 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
553 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
554 uasm_i_dsrl(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
555 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
556 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
561 * BVADDR is the faulting address, PTR is scratch.
562 * PTR will hold the pgd for vmalloc.
564 static void __cpuinit
565 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
566 unsigned int bvaddr, unsigned int ptr)
568 long swpd = (long)swapper_pg_dir;
570 uasm_l_vmalloc(l, *p);
572 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
573 uasm_il_b(p, r, label_vmalloc_done);
574 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
576 UASM_i_LA_mostly(p, ptr, swpd);
577 uasm_il_b(p, r, label_vmalloc_done);
578 if (uasm_in_compat_space_p(swpd))
579 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
581 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
585 #else /* !CONFIG_64BIT */
588 * TMP and PTR are scratch.
589 * TMP will be clobbered, PTR will hold the pgd entry.
591 static void __cpuinit __maybe_unused
592 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
594 long pgdc = (long)pgd_current;
596 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
598 #ifdef CONFIG_MIPS_MT_SMTC
600 * SMTC uses TCBind value as "CPU" index
602 uasm_i_mfc0(p, ptr, C0_TCBIND);
603 UASM_i_LA_mostly(p, tmp, pgdc);
604 uasm_i_srl(p, ptr, ptr, 19);
607 * smp_processor_id() << 3 is stored in CONTEXT.
609 uasm_i_mfc0(p, ptr, C0_CONTEXT);
610 UASM_i_LA_mostly(p, tmp, pgdc);
611 uasm_i_srl(p, ptr, ptr, 23);
613 uasm_i_addu(p, ptr, tmp, ptr);
615 UASM_i_LA_mostly(p, ptr, pgdc);
617 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
618 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
619 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
620 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
621 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
624 #endif /* !CONFIG_64BIT */
626 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
628 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
629 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
631 switch (current_cpu_type()) {
648 UASM_i_SRL(p, ctx, ctx, shift);
649 uasm_i_andi(p, ctx, ctx, mask);
652 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
655 * Bug workaround for the Nevada. It seems as if under certain
656 * circumstances the move from cp0_context might produce a
657 * bogus result when the mfc0 instruction and its consumer are
658 * in a different cacheline or a load instruction, probably any
659 * memory reference, is between them.
661 switch (current_cpu_type()) {
663 UASM_i_LW(p, ptr, 0, ptr);
664 GET_CONTEXT(p, tmp); /* get context reg */
668 GET_CONTEXT(p, tmp); /* get context reg */
669 UASM_i_LW(p, ptr, 0, ptr);
673 build_adjust_context(p, tmp);
674 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
677 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
681 * 64bit address support (36bit on a 32bit CPU) in a 32bit
682 * Kernel is a special case. Only a few CPUs use it.
684 #ifdef CONFIG_64BIT_PHYS_ADDR
685 if (cpu_has_64bits) {
686 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
687 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
688 uasm_i_dsrl(p, tmp, tmp, 6); /* convert to entrylo0 */
689 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
690 uasm_i_dsrl(p, ptep, ptep, 6); /* convert to entrylo1 */
691 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
693 int pte_off_even = sizeof(pte_t) / 2;
694 int pte_off_odd = pte_off_even + sizeof(pte_t);
696 /* The pte entries are pre-shifted */
697 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
698 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
699 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
700 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
703 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
704 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
706 build_tlb_probe_entry(p);
707 UASM_i_SRL(p, tmp, tmp, 6); /* convert to entrylo0 */
708 if (r4k_250MHZhwbug())
709 uasm_i_mtc0(p, 0, C0_ENTRYLO0);
710 uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
711 UASM_i_SRL(p, ptep, ptep, 6); /* convert to entrylo1 */
713 uasm_i_mfc0(p, tmp, C0_INDEX);
714 if (r4k_250MHZhwbug())
715 uasm_i_mtc0(p, 0, C0_ENTRYLO1);
716 uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
721 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
722 * because EXL == 0. If we wrap, we can also use the 32 instruction
723 * slots before the XTLB refill exception handler which belong to the
724 * unused TLB refill exception.
726 #define MIPS64_REFILL_INSNS 32
728 static void __cpuinit build_r4000_tlb_refill_handler(void)
730 u32 *p = tlb_handler;
731 struct uasm_label *l = labels;
732 struct uasm_reloc *r = relocs;
734 unsigned int final_len;
736 memset(tlb_handler, 0, sizeof(tlb_handler));
737 memset(labels, 0, sizeof(labels));
738 memset(relocs, 0, sizeof(relocs));
739 memset(final_handler, 0, sizeof(final_handler));
742 * create the plain linear handler
744 if (bcm1250_m3_war()) {
745 UASM_i_MFC0(&p, K0, C0_BADVADDR);
746 UASM_i_MFC0(&p, K1, C0_ENTRYHI);
747 uasm_i_xor(&p, K0, K0, K1);
748 UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
749 uasm_il_bnez(&p, &r, K0, label_leave);
750 /* No need for uasm_i_nop */
754 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
756 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
759 #ifdef CONFIG_HUGETLB_PAGE
760 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
763 build_get_ptep(&p, K0, K1);
764 build_update_entries(&p, K0, K1);
765 build_tlb_write_entry(&p, &l, &r, tlb_random);
767 uasm_i_eret(&p); /* return from trap */
769 #ifdef CONFIG_HUGETLB_PAGE
770 uasm_l_tlb_huge_update(&l, p);
771 UASM_i_LW(&p, K0, 0, K1);
772 build_huge_update_entries(&p, K0, K1);
773 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random);
777 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1);
781 * Overflow check: For the 64bit handler, we need at least one
782 * free instruction slot for the wrap-around branch. In worst
783 * case, if the intended insertion point is a delay slot, we
784 * need three, with the second nop'ed and the third being
787 /* Loongson2 ebase is different than r4k, we have more space */
788 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
789 if ((p - tlb_handler) > 64)
790 panic("TLB refill handler space exceeded");
792 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
793 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
794 && uasm_insn_has_bdelay(relocs,
795 tlb_handler + MIPS64_REFILL_INSNS - 3)))
796 panic("TLB refill handler space exceeded");
800 * Now fold the handler in the TLB refill handler space.
802 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
804 /* Simplest case, just copy the handler. */
805 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
806 final_len = p - tlb_handler;
807 #else /* CONFIG_64BIT */
808 f = final_handler + MIPS64_REFILL_INSNS;
809 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
810 /* Just copy the handler. */
811 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
812 final_len = p - tlb_handler;
814 #if defined(CONFIG_HUGETLB_PAGE)
815 const enum label_id ls = label_tlb_huge_update;
817 const enum label_id ls = label_vmalloc;
823 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
825 BUG_ON(i == ARRAY_SIZE(labels));
826 split = labels[i].addr;
829 * See if we have overflown one way or the other.
831 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
832 split < p - MIPS64_REFILL_INSNS)
837 * Split two instructions before the end. One
838 * for the branch and one for the instruction
841 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
844 * If the branch would fall in a delay slot,
845 * we must back up an additional instruction
846 * so that it is no longer in a delay slot.
848 if (uasm_insn_has_bdelay(relocs, split - 1))
851 /* Copy first part of the handler. */
852 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
853 f += split - tlb_handler;
857 uasm_l_split(&l, final_handler);
858 uasm_il_b(&f, &r, label_split);
859 if (uasm_insn_has_bdelay(relocs, split))
862 uasm_copy_handler(relocs, labels,
863 split, split + 1, f);
864 uasm_move_labels(labels, f, f + 1, -1);
870 /* Copy the rest of the handler. */
871 uasm_copy_handler(relocs, labels, split, p, final_handler);
872 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
875 #endif /* CONFIG_64BIT */
877 uasm_resolve_relocs(relocs, labels);
878 pr_debug("Wrote TLB refill handler (%u instructions).\n",
881 memcpy((void *)ebase, final_handler, 0x100);
883 dump_handler((u32 *)ebase, 64);
887 * TLB load/store/modify handlers.
889 * Only the fastpath gets synthesized at runtime, the slowpath for
890 * do_page_fault remains normal asm.
892 extern void tlb_do_page_fault_0(void);
893 extern void tlb_do_page_fault_1(void);
896 * 128 instructions for the fastpath handler is generous and should
899 #define FASTPATH_SIZE 128
901 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
902 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
903 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
905 static void __cpuinit
906 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
909 # ifdef CONFIG_64BIT_PHYS_ADDR
911 uasm_i_lld(p, pte, 0, ptr);
914 UASM_i_LL(p, pte, 0, ptr);
916 # ifdef CONFIG_64BIT_PHYS_ADDR
918 uasm_i_ld(p, pte, 0, ptr);
921 UASM_i_LW(p, pte, 0, ptr);
925 static void __cpuinit
926 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
929 #ifdef CONFIG_64BIT_PHYS_ADDR
930 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
933 uasm_i_ori(p, pte, pte, mode);
935 # ifdef CONFIG_64BIT_PHYS_ADDR
937 uasm_i_scd(p, pte, 0, ptr);
940 UASM_i_SC(p, pte, 0, ptr);
942 if (r10000_llsc_war())
943 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
945 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
947 # ifdef CONFIG_64BIT_PHYS_ADDR
948 if (!cpu_has_64bits) {
949 /* no uasm_i_nop needed */
950 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
951 uasm_i_ori(p, pte, pte, hwmode);
952 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
953 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
954 /* no uasm_i_nop needed */
955 uasm_i_lw(p, pte, 0, ptr);
962 # ifdef CONFIG_64BIT_PHYS_ADDR
964 uasm_i_sd(p, pte, 0, ptr);
967 UASM_i_SW(p, pte, 0, ptr);
969 # ifdef CONFIG_64BIT_PHYS_ADDR
970 if (!cpu_has_64bits) {
971 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
972 uasm_i_ori(p, pte, pte, hwmode);
973 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
974 uasm_i_lw(p, pte, 0, ptr);
981 * Check if PTE is present, if not then jump to LABEL. PTR points to
982 * the page table where this PTE is located, PTE will be re-loaded
983 * with it's original value.
985 static void __cpuinit
986 build_pte_present(u32 **p, struct uasm_reloc **r,
987 unsigned int pte, unsigned int ptr, enum label_id lid)
989 uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
990 uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
991 uasm_il_bnez(p, r, pte, lid);
992 iPTE_LW(p, pte, ptr);
995 /* Make PTE valid, store result in PTR. */
996 static void __cpuinit
997 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1000 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1002 iPTE_SW(p, r, pte, ptr, mode);
1006 * Check if PTE can be written to, if not branch to LABEL. Regardless
1007 * restore PTE with value from PTR when done.
1009 static void __cpuinit
1010 build_pte_writable(u32 **p, struct uasm_reloc **r,
1011 unsigned int pte, unsigned int ptr, enum label_id lid)
1013 uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1014 uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1015 uasm_il_bnez(p, r, pte, lid);
1016 iPTE_LW(p, pte, ptr);
1019 /* Make PTE writable, update software status bits as well, then store
1022 static void __cpuinit
1023 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1026 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1029 iPTE_SW(p, r, pte, ptr, mode);
1033 * Check if PTE can be modified, if not branch to LABEL. Regardless
1034 * restore PTE with value from PTR when done.
1036 static void __cpuinit
1037 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1038 unsigned int pte, unsigned int ptr, enum label_id lid)
1040 uasm_i_andi(p, pte, pte, _PAGE_WRITE);
1041 uasm_il_beqz(p, r, pte, lid);
1042 iPTE_LW(p, pte, ptr);
1045 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1047 * R3000 style TLB load/store/modify handlers.
1051 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1054 static void __cpuinit
1055 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1057 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1058 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1061 uasm_i_rfe(p); /* branch delay */
1065 * This places the pte into ENTRYLO0 and writes it with tlbwi
1066 * or tlbwr as appropriate. This is because the index register
1067 * may have the probe fail bit set as a result of a trap on a
1068 * kseg2 access, i.e. without refill. Then it returns.
1070 static void __cpuinit
1071 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1072 struct uasm_reloc **r, unsigned int pte,
1075 uasm_i_mfc0(p, tmp, C0_INDEX);
1076 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1077 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1078 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1079 uasm_i_tlbwi(p); /* cp0 delay */
1081 uasm_i_rfe(p); /* branch delay */
1082 uasm_l_r3000_write_probe_fail(l, *p);
1083 uasm_i_tlbwr(p); /* cp0 delay */
1085 uasm_i_rfe(p); /* branch delay */
1088 static void __cpuinit
1089 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1092 long pgdc = (long)pgd_current;
1094 uasm_i_mfc0(p, pte, C0_BADVADDR);
1095 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1096 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1097 uasm_i_srl(p, pte, pte, 22); /* load delay */
1098 uasm_i_sll(p, pte, pte, 2);
1099 uasm_i_addu(p, ptr, ptr, pte);
1100 uasm_i_mfc0(p, pte, C0_CONTEXT);
1101 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1102 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1103 uasm_i_addu(p, ptr, ptr, pte);
1104 uasm_i_lw(p, pte, 0, ptr);
1105 uasm_i_tlbp(p); /* load delay */
1108 static void __cpuinit build_r3000_tlb_load_handler(void)
1110 u32 *p = handle_tlbl;
1111 struct uasm_label *l = labels;
1112 struct uasm_reloc *r = relocs;
1114 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1115 memset(labels, 0, sizeof(labels));
1116 memset(relocs, 0, sizeof(relocs));
1118 build_r3000_tlbchange_handler_head(&p, K0, K1);
1119 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1120 uasm_i_nop(&p); /* load delay */
1121 build_make_valid(&p, &r, K0, K1);
1122 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1124 uasm_l_nopage_tlbl(&l, p);
1125 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1128 if ((p - handle_tlbl) > FASTPATH_SIZE)
1129 panic("TLB load handler fastpath space exceeded");
1131 uasm_resolve_relocs(relocs, labels);
1132 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1133 (unsigned int)(p - handle_tlbl));
1135 dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1138 static void __cpuinit build_r3000_tlb_store_handler(void)
1140 u32 *p = handle_tlbs;
1141 struct uasm_label *l = labels;
1142 struct uasm_reloc *r = relocs;
1144 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1145 memset(labels, 0, sizeof(labels));
1146 memset(relocs, 0, sizeof(relocs));
1148 build_r3000_tlbchange_handler_head(&p, K0, K1);
1149 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1150 uasm_i_nop(&p); /* load delay */
1151 build_make_write(&p, &r, K0, K1);
1152 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1154 uasm_l_nopage_tlbs(&l, p);
1155 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1158 if ((p - handle_tlbs) > FASTPATH_SIZE)
1159 panic("TLB store handler fastpath space exceeded");
1161 uasm_resolve_relocs(relocs, labels);
1162 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1163 (unsigned int)(p - handle_tlbs));
1165 dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1168 static void __cpuinit build_r3000_tlb_modify_handler(void)
1170 u32 *p = handle_tlbm;
1171 struct uasm_label *l = labels;
1172 struct uasm_reloc *r = relocs;
1174 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1175 memset(labels, 0, sizeof(labels));
1176 memset(relocs, 0, sizeof(relocs));
1178 build_r3000_tlbchange_handler_head(&p, K0, K1);
1179 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1180 uasm_i_nop(&p); /* load delay */
1181 build_make_write(&p, &r, K0, K1);
1182 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1184 uasm_l_nopage_tlbm(&l, p);
1185 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1188 if ((p - handle_tlbm) > FASTPATH_SIZE)
1189 panic("TLB modify handler fastpath space exceeded");
1191 uasm_resolve_relocs(relocs, labels);
1192 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1193 (unsigned int)(p - handle_tlbm));
1195 dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1197 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1200 * R4000 style TLB load/store/modify handlers.
1202 static void __cpuinit
1203 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1204 struct uasm_reloc **r, unsigned int pte,
1208 build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1210 build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1213 #ifdef CONFIG_HUGETLB_PAGE
1215 * For huge tlb entries, pmd doesn't contain an address but
1216 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1217 * see if we need to jump to huge tlb processing.
1219 build_is_huge_pte(p, r, pte, ptr, label_tlb_huge_update);
1222 UASM_i_MFC0(p, pte, C0_BADVADDR);
1223 UASM_i_LW(p, ptr, 0, ptr);
1224 UASM_i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1225 uasm_i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1226 UASM_i_ADDU(p, ptr, ptr, pte);
1229 uasm_l_smp_pgtable_change(l, *p);
1231 iPTE_LW(p, pte, ptr); /* get even pte */
1232 if (!m4kc_tlbp_war())
1233 build_tlb_probe_entry(p);
1236 static void __cpuinit
1237 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1238 struct uasm_reloc **r, unsigned int tmp,
1241 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1242 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1243 build_update_entries(p, tmp, ptr);
1244 build_tlb_write_entry(p, l, r, tlb_indexed);
1245 uasm_l_leave(l, *p);
1246 uasm_i_eret(p); /* return from trap */
1249 build_get_pgd_vmalloc64(p, l, r, tmp, ptr);
1253 static void __cpuinit build_r4000_tlb_load_handler(void)
1255 u32 *p = handle_tlbl;
1256 struct uasm_label *l = labels;
1257 struct uasm_reloc *r = relocs;
1259 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1260 memset(labels, 0, sizeof(labels));
1261 memset(relocs, 0, sizeof(relocs));
1263 if (bcm1250_m3_war()) {
1264 UASM_i_MFC0(&p, K0, C0_BADVADDR);
1265 UASM_i_MFC0(&p, K1, C0_ENTRYHI);
1266 uasm_i_xor(&p, K0, K0, K1);
1267 UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
1268 uasm_il_bnez(&p, &r, K0, label_leave);
1269 /* No need for uasm_i_nop */
1272 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1273 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1274 if (m4kc_tlbp_war())
1275 build_tlb_probe_entry(&p);
1276 build_make_valid(&p, &r, K0, K1);
1277 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1279 #ifdef CONFIG_HUGETLB_PAGE
1281 * This is the entry point when build_r4000_tlbchange_handler_head
1282 * spots a huge page.
1284 uasm_l_tlb_huge_update(&l, p);
1285 iPTE_LW(&p, K0, K1);
1286 build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1287 build_tlb_probe_entry(&p);
1288 uasm_i_ori(&p, K0, K0, (_PAGE_ACCESSED | _PAGE_VALID));
1289 build_huge_handler_tail(&p, &r, &l, K0, K1);
1292 uasm_l_nopage_tlbl(&l, p);
1293 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1296 if ((p - handle_tlbl) > FASTPATH_SIZE)
1297 panic("TLB load handler fastpath space exceeded");
1299 uasm_resolve_relocs(relocs, labels);
1300 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1301 (unsigned int)(p - handle_tlbl));
1303 dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1306 static void __cpuinit build_r4000_tlb_store_handler(void)
1308 u32 *p = handle_tlbs;
1309 struct uasm_label *l = labels;
1310 struct uasm_reloc *r = relocs;
1312 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1313 memset(labels, 0, sizeof(labels));
1314 memset(relocs, 0, sizeof(relocs));
1316 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1317 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1318 if (m4kc_tlbp_war())
1319 build_tlb_probe_entry(&p);
1320 build_make_write(&p, &r, K0, K1);
1321 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1323 #ifdef CONFIG_HUGETLB_PAGE
1325 * This is the entry point when
1326 * build_r4000_tlbchange_handler_head spots a huge page.
1328 uasm_l_tlb_huge_update(&l, p);
1329 iPTE_LW(&p, K0, K1);
1330 build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1331 build_tlb_probe_entry(&p);
1332 uasm_i_ori(&p, K0, K0,
1333 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1334 build_huge_handler_tail(&p, &r, &l, K0, K1);
1337 uasm_l_nopage_tlbs(&l, p);
1338 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1341 if ((p - handle_tlbs) > FASTPATH_SIZE)
1342 panic("TLB store handler fastpath space exceeded");
1344 uasm_resolve_relocs(relocs, labels);
1345 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1346 (unsigned int)(p - handle_tlbs));
1348 dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1351 static void __cpuinit build_r4000_tlb_modify_handler(void)
1353 u32 *p = handle_tlbm;
1354 struct uasm_label *l = labels;
1355 struct uasm_reloc *r = relocs;
1357 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1358 memset(labels, 0, sizeof(labels));
1359 memset(relocs, 0, sizeof(relocs));
1361 build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1362 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1363 if (m4kc_tlbp_war())
1364 build_tlb_probe_entry(&p);
1365 /* Present and writable bits set, set accessed and dirty bits. */
1366 build_make_write(&p, &r, K0, K1);
1367 build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1369 #ifdef CONFIG_HUGETLB_PAGE
1371 * This is the entry point when
1372 * build_r4000_tlbchange_handler_head spots a huge page.
1374 uasm_l_tlb_huge_update(&l, p);
1375 iPTE_LW(&p, K0, K1);
1376 build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1377 build_tlb_probe_entry(&p);
1378 uasm_i_ori(&p, K0, K0,
1379 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1380 build_huge_handler_tail(&p, &r, &l, K0, K1);
1383 uasm_l_nopage_tlbm(&l, p);
1384 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1387 if ((p - handle_tlbm) > FASTPATH_SIZE)
1388 panic("TLB modify handler fastpath space exceeded");
1390 uasm_resolve_relocs(relocs, labels);
1391 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1392 (unsigned int)(p - handle_tlbm));
1394 dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1397 void __cpuinit build_tlb_refill_handler(void)
1400 * The refill handler is generated per-CPU, multi-node systems
1401 * may have local storage for it. The other handlers are only
1404 static int run_once = 0;
1406 switch (current_cpu_type()) {
1414 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1415 build_r3000_tlb_refill_handler();
1417 build_r3000_tlb_load_handler();
1418 build_r3000_tlb_store_handler();
1419 build_r3000_tlb_modify_handler();
1423 panic("No R3000 TLB refill handler");
1429 panic("No R6000 TLB refill handler yet");
1433 panic("No R8000 TLB refill handler yet");
1437 build_r4000_tlb_refill_handler();
1439 build_r4000_tlb_load_handler();
1440 build_r4000_tlb_store_handler();
1441 build_r4000_tlb_modify_handler();
1447 void __cpuinit flush_tlb_handlers(void)
1449 local_flush_icache_range((unsigned long)handle_tlbl,
1450 (unsigned long)handle_tlbl + sizeof(handle_tlbl));
1451 local_flush_icache_range((unsigned long)handle_tlbs,
1452 (unsigned long)handle_tlbs + sizeof(handle_tlbs));
1453 local_flush_icache_range((unsigned long)handle_tlbm,
1454 (unsigned long)handle_tlbm + sizeof(handle_tlbm));