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/init.h>
30 #include <linux/cache.h>
32 #include <asm/cacheflush.h>
33 #include <asm/pgtable.h>
36 #include <asm/setup.h>
39 * TLB load/store/modify handlers.
41 * Only the fastpath gets synthesized at runtime, the slowpath for
42 * do_page_fault remains normal asm.
44 extern void tlb_do_page_fault_0(void);
45 extern void tlb_do_page_fault_1(void);
47 struct work_registers {
56 } ____cacheline_aligned_in_smp;
58 static struct tlb_reg_save handler_reg_save[NR_CPUS];
60 static inline int r45k_bvahwbug(void)
62 /* XXX: We should probe for the presence of this bug, but we don't. */
66 static inline int r4k_250MHZhwbug(void)
68 /* XXX: We should probe for the presence of this bug, but we don't. */
72 static inline int __maybe_unused bcm1250_m3_war(void)
74 return BCM1250_M3_WAR;
77 static inline int __maybe_unused r10000_llsc_war(void)
79 return R10000_LLSC_WAR;
82 static int use_bbit_insns(void)
84 switch (current_cpu_type()) {
85 case CPU_CAVIUM_OCTEON:
86 case CPU_CAVIUM_OCTEON_PLUS:
87 case CPU_CAVIUM_OCTEON2:
94 static int use_lwx_insns(void)
96 switch (current_cpu_type()) {
97 case CPU_CAVIUM_OCTEON2:
103 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
104 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
105 static bool scratchpad_available(void)
109 static int scratchpad_offset(int i)
112 * CVMSEG starts at address -32768 and extends for
113 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
115 i += 1; /* Kernel use starts at the top and works down. */
116 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
119 static bool scratchpad_available(void)
123 static int scratchpad_offset(int i)
126 /* Really unreachable, but evidently some GCC want this. */
131 * Found by experiment: At least some revisions of the 4kc throw under
132 * some circumstances a machine check exception, triggered by invalid
133 * values in the index register. Delaying the tlbp instruction until
134 * after the next branch, plus adding an additional nop in front of
135 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
136 * why; it's not an issue caused by the core RTL.
139 static int __cpuinit m4kc_tlbp_war(void)
141 return (current_cpu_data.processor_id & 0xffff00) ==
142 (PRID_COMP_MIPS | PRID_IMP_4KC);
145 /* Handle labels (which must be positive integers). */
147 label_second_part = 1,
152 label_split = label_tlbw_hazard_0 + 8,
153 label_tlbl_goaround1,
154 label_tlbl_goaround2,
158 label_smp_pgtable_change,
159 label_r3000_write_probe_fail,
160 label_large_segbits_fault,
161 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
162 label_tlb_huge_update,
166 UASM_L_LA(_second_part)
169 UASM_L_LA(_vmalloc_done)
170 /* _tlbw_hazard_x is handled differently. */
172 UASM_L_LA(_tlbl_goaround1)
173 UASM_L_LA(_tlbl_goaround2)
174 UASM_L_LA(_nopage_tlbl)
175 UASM_L_LA(_nopage_tlbs)
176 UASM_L_LA(_nopage_tlbm)
177 UASM_L_LA(_smp_pgtable_change)
178 UASM_L_LA(_r3000_write_probe_fail)
179 UASM_L_LA(_large_segbits_fault)
180 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
181 UASM_L_LA(_tlb_huge_update)
184 static int __cpuinitdata hazard_instance;
186 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
190 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
197 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
201 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
209 * pgtable bits are assigned dynamically depending on processor feature
210 * and statically based on kernel configuration. This spits out the actual
211 * values the kernel is using. Required to make sense from disassembled
212 * TLB exception handlers.
214 static void output_pgtable_bits_defines(void)
216 #define pr_define(fmt, ...) \
217 pr_debug("#define " fmt, ##__VA_ARGS__)
219 pr_debug("#include <asm/asm.h>\n");
220 pr_debug("#include <asm/regdef.h>\n");
223 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
224 pr_define("_PAGE_READ_SHIFT %d\n", _PAGE_READ_SHIFT);
225 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
226 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
227 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
228 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
229 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
230 pr_define("_PAGE_SPLITTING_SHIFT %d\n", _PAGE_SPLITTING_SHIFT);
233 #ifdef _PAGE_NO_EXEC_SHIFT
234 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
236 #ifdef _PAGE_NO_READ_SHIFT
237 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
240 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
241 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
242 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
243 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
247 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
251 pr_debug("LEAF(%s)\n", symbol);
253 pr_debug("\t.set push\n");
254 pr_debug("\t.set noreorder\n");
256 for (i = 0; i < count; i++)
257 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
259 pr_debug("\t.set\tpop\n");
261 pr_debug("\tEND(%s)\n", symbol);
264 /* The only general purpose registers allowed in TLB handlers. */
268 /* Some CP0 registers */
269 #define C0_INDEX 0, 0
270 #define C0_ENTRYLO0 2, 0
271 #define C0_TCBIND 2, 2
272 #define C0_ENTRYLO1 3, 0
273 #define C0_CONTEXT 4, 0
274 #define C0_PAGEMASK 5, 0
275 #define C0_BADVADDR 8, 0
276 #define C0_ENTRYHI 10, 0
278 #define C0_XCONTEXT 20, 0
281 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
283 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
286 /* The worst case length of the handler is around 18 instructions for
287 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
288 * Maximum space available is 32 instructions for R3000 and 64
289 * instructions for R4000.
291 * We deliberately chose a buffer size of 128, so we won't scribble
292 * over anything important on overflow before we panic.
294 static u32 tlb_handler[128] __cpuinitdata;
296 /* simply assume worst case size for labels and relocs */
297 static struct uasm_label labels[128] __cpuinitdata;
298 static struct uasm_reloc relocs[128] __cpuinitdata;
301 static int check_for_high_segbits __cpuinitdata;
304 static int check_for_high_segbits __cpuinitdata;
306 static unsigned int kscratch_used_mask __cpuinitdata;
308 static int __cpuinit allocate_kscratch(void)
311 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
318 r--; /* make it zero based */
320 kscratch_used_mask |= (1 << r);
325 static int scratch_reg __cpuinitdata;
326 static int pgd_reg __cpuinitdata;
327 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
329 static struct work_registers __cpuinit build_get_work_registers(u32 **p)
331 struct work_registers r;
333 int smp_processor_id_reg;
334 int smp_processor_id_sel;
335 int smp_processor_id_shift;
337 if (scratch_reg > 0) {
338 /* Save in CPU local C0_KScratch? */
339 UASM_i_MTC0(p, 1, 31, scratch_reg);
346 if (num_possible_cpus() > 1) {
347 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
348 smp_processor_id_shift = 51;
349 smp_processor_id_reg = 20; /* XContext */
350 smp_processor_id_sel = 0;
353 smp_processor_id_shift = 25;
354 smp_processor_id_reg = 4; /* Context */
355 smp_processor_id_sel = 0;
358 smp_processor_id_shift = 26;
359 smp_processor_id_reg = 4; /* Context */
360 smp_processor_id_sel = 0;
363 /* Get smp_processor_id */
364 UASM_i_MFC0(p, K0, smp_processor_id_reg, smp_processor_id_sel);
365 UASM_i_SRL_SAFE(p, K0, K0, smp_processor_id_shift);
367 /* handler_reg_save index in K0 */
368 UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
370 UASM_i_LA(p, K1, (long)&handler_reg_save);
371 UASM_i_ADDU(p, K0, K0, K1);
373 UASM_i_LA(p, K0, (long)&handler_reg_save);
375 /* K0 now points to save area, save $1 and $2 */
376 UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
377 UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
385 static void __cpuinit build_restore_work_registers(u32 **p)
387 if (scratch_reg > 0) {
388 UASM_i_MFC0(p, 1, 31, scratch_reg);
391 /* K0 already points to save area, restore $1 and $2 */
392 UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
393 UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
396 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
399 * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
400 * we cannot do r3000 under these circumstances.
402 * Declare pgd_current here instead of including mmu_context.h to avoid type
403 * conflicts for tlbmiss_handler_setup_pgd
405 extern unsigned long pgd_current[];
408 * The R3000 TLB handler is simple.
410 static void __cpuinit build_r3000_tlb_refill_handler(void)
412 long pgdc = (long)pgd_current;
415 memset(tlb_handler, 0, sizeof(tlb_handler));
418 uasm_i_mfc0(&p, K0, C0_BADVADDR);
419 uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
420 uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
421 uasm_i_srl(&p, K0, K0, 22); /* load delay */
422 uasm_i_sll(&p, K0, K0, 2);
423 uasm_i_addu(&p, K1, K1, K0);
424 uasm_i_mfc0(&p, K0, C0_CONTEXT);
425 uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
426 uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
427 uasm_i_addu(&p, K1, K1, K0);
428 uasm_i_lw(&p, K0, 0, K1);
429 uasm_i_nop(&p); /* load delay */
430 uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
431 uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
432 uasm_i_tlbwr(&p); /* cp0 delay */
434 uasm_i_rfe(&p); /* branch delay */
436 if (p > tlb_handler + 32)
437 panic("TLB refill handler space exceeded");
439 pr_debug("Wrote TLB refill handler (%u instructions).\n",
440 (unsigned int)(p - tlb_handler));
442 memcpy((void *)ebase, tlb_handler, 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] __cpuinitdata;
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 __cpuinit __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 __cpuinit 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) {
516 * The architecture spec says an ehb is required here,
517 * but a number of cores do not have the hazard and
518 * using an ehb causes an expensive pipeline stall.
520 switch (current_cpu_type()) {
533 switch (current_cpu_type()) {
541 * This branch uses up a mtc0 hazard nop slot and saves
542 * two nops after the tlbw instruction.
544 uasm_bgezl_hazard(p, r, hazard_instance);
546 uasm_bgezl_label(l, p, hazard_instance);
560 uasm_i_nop(p); /* QED specifies 2 nops hazard */
561 uasm_i_nop(p); /* QED specifies 2 nops hazard */
608 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
609 * use of the JTLB for instructions should not occur for 4
610 * cpu cycles and use for data translations should not occur
650 panic("No TLB refill handler yet (CPU type: %d)",
651 current_cpu_data.cputype);
656 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
660 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
662 #ifdef CONFIG_64BIT_PHYS_ADDR
663 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
665 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
670 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
672 static __cpuinit void build_restore_pagemask(u32 **p,
673 struct uasm_reloc **r,
678 if (restore_scratch) {
679 /* Reset default page size */
680 if (PM_DEFAULT_MASK >> 16) {
681 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
682 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
683 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
684 uasm_il_b(p, r, lid);
685 } else if (PM_DEFAULT_MASK) {
686 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
687 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
688 uasm_il_b(p, r, lid);
690 uasm_i_mtc0(p, 0, C0_PAGEMASK);
691 uasm_il_b(p, r, lid);
694 UASM_i_MFC0(p, 1, 31, scratch_reg);
696 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
698 /* Reset default page size */
699 if (PM_DEFAULT_MASK >> 16) {
700 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
701 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
702 uasm_il_b(p, r, lid);
703 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
704 } else if (PM_DEFAULT_MASK) {
705 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
706 uasm_il_b(p, r, lid);
707 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
709 uasm_il_b(p, r, lid);
710 uasm_i_mtc0(p, 0, C0_PAGEMASK);
715 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
716 struct uasm_label **l,
717 struct uasm_reloc **r,
719 enum tlb_write_entry wmode,
722 /* Set huge page tlb entry size */
723 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
724 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
725 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
727 build_tlb_write_entry(p, l, r, wmode);
729 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
733 * Check if Huge PTE is present, if so then jump to LABEL.
735 static void __cpuinit
736 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
737 unsigned int pmd, int lid)
739 UASM_i_LW(p, tmp, 0, pmd);
740 if (use_bbit_insns()) {
741 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
743 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
744 uasm_il_bnez(p, r, tmp, lid);
748 static __cpuinit void build_huge_update_entries(u32 **p,
755 * A huge PTE describes an area the size of the
756 * configured huge page size. This is twice the
757 * of the large TLB entry size we intend to use.
758 * A TLB entry half the size of the configured
759 * huge page size is configured into entrylo0
760 * and entrylo1 to cover the contiguous huge PTE
763 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
765 /* We can clobber tmp. It isn't used after this.*/
767 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
769 build_convert_pte_to_entrylo(p, pte);
770 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
771 /* convert to entrylo1 */
773 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
775 UASM_i_ADDU(p, pte, pte, tmp);
777 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
780 static __cpuinit void build_huge_handler_tail(u32 **p,
781 struct uasm_reloc **r,
782 struct uasm_label **l,
787 UASM_i_SC(p, pte, 0, ptr);
788 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
789 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
791 UASM_i_SW(p, pte, 0, ptr);
793 build_huge_update_entries(p, pte, ptr);
794 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
796 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
800 * TMP and PTR are scratch.
801 * TMP will be clobbered, PTR will hold the pmd entry.
803 static void __cpuinit
804 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
805 unsigned int tmp, unsigned int ptr)
807 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
808 long pgdc = (long)pgd_current;
811 * The vmalloc handling is not in the hotpath.
813 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
815 if (check_for_high_segbits) {
817 * The kernel currently implicitely assumes that the
818 * MIPS SEGBITS parameter for the processor is
819 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
820 * allocate virtual addresses outside the maximum
821 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
822 * that doesn't prevent user code from accessing the
823 * higher xuseg addresses. Here, we make sure that
824 * everything but the lower xuseg addresses goes down
825 * the module_alloc/vmalloc path.
827 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
828 uasm_il_bnez(p, r, ptr, label_vmalloc);
830 uasm_il_bltz(p, r, tmp, label_vmalloc);
832 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
834 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
836 /* pgd is in pgd_reg */
837 UASM_i_MFC0(p, ptr, 31, pgd_reg);
840 * &pgd << 11 stored in CONTEXT [23..63].
842 UASM_i_MFC0(p, ptr, C0_CONTEXT);
844 /* Clear lower 23 bits of context. */
845 uasm_i_dins(p, ptr, 0, 0, 23);
847 /* 1 0 1 0 1 << 6 xkphys cached */
848 uasm_i_ori(p, ptr, ptr, 0x540);
849 uasm_i_drotr(p, ptr, ptr, 11);
851 #elif defined(CONFIG_SMP)
852 # ifdef CONFIG_MIPS_MT_SMTC
854 * SMTC uses TCBind value as "CPU" index
856 uasm_i_mfc0(p, ptr, C0_TCBIND);
857 uasm_i_dsrl_safe(p, ptr, ptr, 19);
860 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
863 uasm_i_dmfc0(p, ptr, C0_CONTEXT);
864 uasm_i_dsrl_safe(p, ptr, ptr, 23);
866 UASM_i_LA_mostly(p, tmp, pgdc);
867 uasm_i_daddu(p, ptr, ptr, tmp);
868 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
869 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
871 UASM_i_LA_mostly(p, ptr, pgdc);
872 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
875 uasm_l_vmalloc_done(l, *p);
877 /* get pgd offset in bytes */
878 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
880 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
881 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
882 #ifndef __PAGETABLE_PMD_FOLDED
883 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
884 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
885 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
886 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
887 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
892 * BVADDR is the faulting address, PTR is scratch.
893 * PTR will hold the pgd for vmalloc.
895 static void __cpuinit
896 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
897 unsigned int bvaddr, unsigned int ptr,
898 enum vmalloc64_mode mode)
900 long swpd = (long)swapper_pg_dir;
901 int single_insn_swpd;
902 int did_vmalloc_branch = 0;
904 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
906 uasm_l_vmalloc(l, *p);
908 if (mode != not_refill && check_for_high_segbits) {
909 if (single_insn_swpd) {
910 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
911 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
912 did_vmalloc_branch = 1;
915 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
918 if (!did_vmalloc_branch) {
919 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
920 uasm_il_b(p, r, label_vmalloc_done);
921 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
923 UASM_i_LA_mostly(p, ptr, swpd);
924 uasm_il_b(p, r, label_vmalloc_done);
925 if (uasm_in_compat_space_p(swpd))
926 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
928 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
931 if (mode != not_refill && check_for_high_segbits) {
932 uasm_l_large_segbits_fault(l, *p);
934 * We get here if we are an xsseg address, or if we are
935 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
937 * Ignoring xsseg (assume disabled so would generate
938 * (address errors?), the only remaining possibility
939 * is the upper xuseg addresses. On processors with
940 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
941 * addresses would have taken an address error. We try
942 * to mimic that here by taking a load/istream page
945 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
948 if (mode == refill_scratch) {
950 UASM_i_MFC0(p, 1, 31, scratch_reg);
952 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
959 #else /* !CONFIG_64BIT */
962 * TMP and PTR are scratch.
963 * TMP will be clobbered, PTR will hold the pgd entry.
965 static void __cpuinit __maybe_unused
966 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
968 long pgdc = (long)pgd_current;
970 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
972 #ifdef CONFIG_MIPS_MT_SMTC
974 * SMTC uses TCBind value as "CPU" index
976 uasm_i_mfc0(p, ptr, C0_TCBIND);
977 UASM_i_LA_mostly(p, tmp, pgdc);
978 uasm_i_srl(p, ptr, ptr, 19);
981 * smp_processor_id() << 3 is stored in CONTEXT.
983 uasm_i_mfc0(p, ptr, C0_CONTEXT);
984 UASM_i_LA_mostly(p, tmp, pgdc);
985 uasm_i_srl(p, ptr, ptr, 23);
987 uasm_i_addu(p, ptr, tmp, ptr);
989 UASM_i_LA_mostly(p, ptr, pgdc);
991 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
992 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
994 if (cpu_has_mips_r2) {
995 uasm_i_ext(p, tmp, tmp, PGDIR_SHIFT, (32 - PGDIR_SHIFT));
996 uasm_i_ins(p, ptr, tmp, PGD_T_LOG2, (32 - PGDIR_SHIFT));
1000 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
1001 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
1002 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
1005 #endif /* !CONFIG_64BIT */
1007 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
1009 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
1010 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
1012 switch (current_cpu_type()) {
1029 UASM_i_SRL(p, ctx, ctx, shift);
1030 uasm_i_andi(p, ctx, ctx, mask);
1033 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1035 if (cpu_has_mips_r2) {
1036 /* PTE ptr offset is obtained from BadVAddr */
1037 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1038 UASM_i_LW(p, ptr, 0, ptr);
1039 uasm_i_ext(p, tmp, tmp, PAGE_SHIFT+1, PGDIR_SHIFT-PAGE_SHIFT-1);
1040 uasm_i_ins(p, ptr, tmp, PTE_T_LOG2+1, PGDIR_SHIFT-PAGE_SHIFT-1);
1045 * Bug workaround for the Nevada. It seems as if under certain
1046 * circumstances the move from cp0_context might produce a
1047 * bogus result when the mfc0 instruction and its consumer are
1048 * in a different cacheline or a load instruction, probably any
1049 * memory reference, is between them.
1051 switch (current_cpu_type()) {
1053 UASM_i_LW(p, ptr, 0, ptr);
1054 GET_CONTEXT(p, tmp); /* get context reg */
1058 GET_CONTEXT(p, tmp); /* get context reg */
1059 UASM_i_LW(p, ptr, 0, ptr);
1063 build_adjust_context(p, tmp);
1064 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1067 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
1071 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1072 * Kernel is a special case. Only a few CPUs use it.
1074 #ifdef CONFIG_64BIT_PHYS_ADDR
1075 if (cpu_has_64bits) {
1076 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1077 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1079 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1080 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1081 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1083 uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1084 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1085 uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1087 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1089 int pte_off_even = sizeof(pte_t) / 2;
1090 int pte_off_odd = pte_off_even + sizeof(pte_t);
1092 /* The pte entries are pre-shifted */
1093 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1094 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1095 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1096 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1099 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1100 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1101 if (r45k_bvahwbug())
1102 build_tlb_probe_entry(p);
1104 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1105 if (r4k_250MHZhwbug())
1106 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1107 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1108 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1110 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1111 if (r4k_250MHZhwbug())
1112 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1113 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1114 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1115 if (r45k_bvahwbug())
1116 uasm_i_mfc0(p, tmp, C0_INDEX);
1118 if (r4k_250MHZhwbug())
1119 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1120 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1124 struct mips_huge_tlb_info {
1126 int restore_scratch;
1129 static struct mips_huge_tlb_info __cpuinit
1130 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1131 struct uasm_reloc **r, unsigned int tmp,
1132 unsigned int ptr, int c0_scratch)
1134 struct mips_huge_tlb_info rv;
1135 unsigned int even, odd;
1136 int vmalloc_branch_delay_filled = 0;
1137 const int scratch = 1; /* Our extra working register */
1139 rv.huge_pte = scratch;
1140 rv.restore_scratch = 0;
1142 if (check_for_high_segbits) {
1143 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1146 UASM_i_MFC0(p, ptr, 31, pgd_reg);
1148 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1150 if (c0_scratch >= 0)
1151 UASM_i_MTC0(p, scratch, 31, c0_scratch);
1153 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1155 uasm_i_dsrl_safe(p, scratch, tmp,
1156 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1157 uasm_il_bnez(p, r, scratch, label_vmalloc);
1159 if (pgd_reg == -1) {
1160 vmalloc_branch_delay_filled = 1;
1161 /* Clear lower 23 bits of context. */
1162 uasm_i_dins(p, ptr, 0, 0, 23);
1166 UASM_i_MFC0(p, ptr, 31, pgd_reg);
1168 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1170 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1172 if (c0_scratch >= 0)
1173 UASM_i_MTC0(p, scratch, 31, c0_scratch);
1175 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1178 /* Clear lower 23 bits of context. */
1179 uasm_i_dins(p, ptr, 0, 0, 23);
1181 uasm_il_bltz(p, r, tmp, label_vmalloc);
1184 if (pgd_reg == -1) {
1185 vmalloc_branch_delay_filled = 1;
1186 /* 1 0 1 0 1 << 6 xkphys cached */
1187 uasm_i_ori(p, ptr, ptr, 0x540);
1188 uasm_i_drotr(p, ptr, ptr, 11);
1191 #ifdef __PAGETABLE_PMD_FOLDED
1192 #define LOC_PTEP scratch
1194 #define LOC_PTEP ptr
1197 if (!vmalloc_branch_delay_filled)
1198 /* get pgd offset in bytes */
1199 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1201 uasm_l_vmalloc_done(l, *p);
1205 * fall-through case = badvaddr *pgd_current
1206 * vmalloc case = badvaddr swapper_pg_dir
1209 if (vmalloc_branch_delay_filled)
1210 /* get pgd offset in bytes */
1211 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1213 #ifdef __PAGETABLE_PMD_FOLDED
1214 GET_CONTEXT(p, tmp); /* get context reg */
1216 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1218 if (use_lwx_insns()) {
1219 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1221 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1222 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1225 #ifndef __PAGETABLE_PMD_FOLDED
1226 /* get pmd offset in bytes */
1227 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1228 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1229 GET_CONTEXT(p, tmp); /* get context reg */
1231 if (use_lwx_insns()) {
1232 UASM_i_LWX(p, scratch, scratch, ptr);
1234 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1235 UASM_i_LW(p, scratch, 0, ptr);
1238 /* Adjust the context during the load latency. */
1239 build_adjust_context(p, tmp);
1241 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1242 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1244 * The in the LWX case we don't want to do the load in the
1245 * delay slot. It cannot issue in the same cycle and may be
1246 * speculative and unneeded.
1248 if (use_lwx_insns())
1250 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1253 /* build_update_entries */
1254 if (use_lwx_insns()) {
1257 UASM_i_LWX(p, even, scratch, tmp);
1258 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1259 UASM_i_LWX(p, odd, scratch, tmp);
1261 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1264 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1265 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1268 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1269 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1270 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1272 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1273 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1274 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1276 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1278 if (c0_scratch >= 0) {
1279 UASM_i_MFC0(p, scratch, 31, c0_scratch);
1280 build_tlb_write_entry(p, l, r, tlb_random);
1281 uasm_l_leave(l, *p);
1282 rv.restore_scratch = 1;
1283 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1284 build_tlb_write_entry(p, l, r, tlb_random);
1285 uasm_l_leave(l, *p);
1286 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1288 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1289 build_tlb_write_entry(p, l, r, tlb_random);
1290 uasm_l_leave(l, *p);
1291 rv.restore_scratch = 1;
1294 uasm_i_eret(p); /* return from trap */
1300 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1301 * because EXL == 0. If we wrap, we can also use the 32 instruction
1302 * slots before the XTLB refill exception handler which belong to the
1303 * unused TLB refill exception.
1305 #define MIPS64_REFILL_INSNS 32
1307 static void __cpuinit build_r4000_tlb_refill_handler(void)
1309 u32 *p = tlb_handler;
1310 struct uasm_label *l = labels;
1311 struct uasm_reloc *r = relocs;
1313 unsigned int final_len;
1314 struct mips_huge_tlb_info htlb_info __maybe_unused;
1315 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1317 memset(tlb_handler, 0, sizeof(tlb_handler));
1318 memset(labels, 0, sizeof(labels));
1319 memset(relocs, 0, sizeof(relocs));
1320 memset(final_handler, 0, sizeof(final_handler));
1322 if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1323 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1325 vmalloc_mode = refill_scratch;
1327 htlb_info.huge_pte = K0;
1328 htlb_info.restore_scratch = 0;
1329 vmalloc_mode = refill_noscratch;
1331 * create the plain linear handler
1333 if (bcm1250_m3_war()) {
1334 unsigned int segbits = 44;
1336 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1337 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1338 uasm_i_xor(&p, K0, K0, K1);
1339 uasm_i_dsrl_safe(&p, K1, K0, 62);
1340 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1341 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1342 uasm_i_or(&p, K0, K0, K1);
1343 uasm_il_bnez(&p, &r, K0, label_leave);
1344 /* No need for uasm_i_nop */
1348 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1350 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1353 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1354 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1357 build_get_ptep(&p, K0, K1);
1358 build_update_entries(&p, K0, K1);
1359 build_tlb_write_entry(&p, &l, &r, tlb_random);
1360 uasm_l_leave(&l, p);
1361 uasm_i_eret(&p); /* return from trap */
1363 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1364 uasm_l_tlb_huge_update(&l, p);
1365 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1366 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1367 htlb_info.restore_scratch);
1371 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1375 * Overflow check: For the 64bit handler, we need at least one
1376 * free instruction slot for the wrap-around branch. In worst
1377 * case, if the intended insertion point is a delay slot, we
1378 * need three, with the second nop'ed and the third being
1381 /* Loongson2 ebase is different than r4k, we have more space */
1382 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1383 if ((p - tlb_handler) > 64)
1384 panic("TLB refill handler space exceeded");
1386 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1387 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1388 && uasm_insn_has_bdelay(relocs,
1389 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1390 panic("TLB refill handler space exceeded");
1394 * Now fold the handler in the TLB refill handler space.
1396 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1398 /* Simplest case, just copy the handler. */
1399 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1400 final_len = p - tlb_handler;
1401 #else /* CONFIG_64BIT */
1402 f = final_handler + MIPS64_REFILL_INSNS;
1403 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1404 /* Just copy the handler. */
1405 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1406 final_len = p - tlb_handler;
1408 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1409 const enum label_id ls = label_tlb_huge_update;
1411 const enum label_id ls = label_vmalloc;
1417 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1419 BUG_ON(i == ARRAY_SIZE(labels));
1420 split = labels[i].addr;
1423 * See if we have overflown one way or the other.
1425 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1426 split < p - MIPS64_REFILL_INSNS)
1431 * Split two instructions before the end. One
1432 * for the branch and one for the instruction
1433 * in the delay slot.
1435 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1438 * If the branch would fall in a delay slot,
1439 * we must back up an additional instruction
1440 * so that it is no longer in a delay slot.
1442 if (uasm_insn_has_bdelay(relocs, split - 1))
1445 /* Copy first part of the handler. */
1446 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1447 f += split - tlb_handler;
1450 /* Insert branch. */
1451 uasm_l_split(&l, final_handler);
1452 uasm_il_b(&f, &r, label_split);
1453 if (uasm_insn_has_bdelay(relocs, split))
1456 uasm_copy_handler(relocs, labels,
1457 split, split + 1, f);
1458 uasm_move_labels(labels, f, f + 1, -1);
1464 /* Copy the rest of the handler. */
1465 uasm_copy_handler(relocs, labels, split, p, final_handler);
1466 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1469 #endif /* CONFIG_64BIT */
1471 uasm_resolve_relocs(relocs, labels);
1472 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1475 memcpy((void *)ebase, final_handler, 0x100);
1477 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1481 * 128 instructions for the fastpath handler is generous and should
1482 * never be exceeded.
1484 #define FASTPATH_SIZE 128
1486 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1487 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1488 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1489 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1490 u32 tlbmiss_handler_setup_pgd[16] __cacheline_aligned;
1492 static void __cpuinit build_r4000_setup_pgd(void)
1496 u32 *p = tlbmiss_handler_setup_pgd;
1497 struct uasm_label *l = labels;
1498 struct uasm_reloc *r = relocs;
1500 memset(tlbmiss_handler_setup_pgd, 0, sizeof(tlbmiss_handler_setup_pgd));
1501 memset(labels, 0, sizeof(labels));
1502 memset(relocs, 0, sizeof(relocs));
1504 pgd_reg = allocate_kscratch();
1506 if (pgd_reg == -1) {
1507 /* PGD << 11 in c0_Context */
1509 * If it is a ckseg0 address, convert to a physical
1510 * address. Shifting right by 29 and adding 4 will
1511 * result in zero for these addresses.
1514 UASM_i_SRA(&p, a1, a0, 29);
1515 UASM_i_ADDIU(&p, a1, a1, 4);
1516 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1518 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1519 uasm_l_tlbl_goaround1(&l, p);
1520 UASM_i_SLL(&p, a0, a0, 11);
1522 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1524 /* PGD in c0_KScratch */
1526 UASM_i_MTC0(&p, a0, 31, pgd_reg);
1528 if (p - tlbmiss_handler_setup_pgd > ARRAY_SIZE(tlbmiss_handler_setup_pgd))
1529 panic("tlbmiss_handler_setup_pgd space exceeded");
1530 uasm_resolve_relocs(relocs, labels);
1531 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1532 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1534 dump_handler("tlbmiss_handler",
1535 tlbmiss_handler_setup_pgd,
1536 ARRAY_SIZE(tlbmiss_handler_setup_pgd));
1540 static void __cpuinit
1541 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1544 # ifdef CONFIG_64BIT_PHYS_ADDR
1546 uasm_i_lld(p, pte, 0, ptr);
1549 UASM_i_LL(p, pte, 0, ptr);
1551 # ifdef CONFIG_64BIT_PHYS_ADDR
1553 uasm_i_ld(p, pte, 0, ptr);
1556 UASM_i_LW(p, pte, 0, ptr);
1560 static void __cpuinit
1561 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1564 #ifdef CONFIG_64BIT_PHYS_ADDR
1565 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1568 uasm_i_ori(p, pte, pte, mode);
1570 # ifdef CONFIG_64BIT_PHYS_ADDR
1572 uasm_i_scd(p, pte, 0, ptr);
1575 UASM_i_SC(p, pte, 0, ptr);
1577 if (r10000_llsc_war())
1578 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1580 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1582 # ifdef CONFIG_64BIT_PHYS_ADDR
1583 if (!cpu_has_64bits) {
1584 /* no uasm_i_nop needed */
1585 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1586 uasm_i_ori(p, pte, pte, hwmode);
1587 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1588 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1589 /* no uasm_i_nop needed */
1590 uasm_i_lw(p, pte, 0, ptr);
1597 # ifdef CONFIG_64BIT_PHYS_ADDR
1599 uasm_i_sd(p, pte, 0, ptr);
1602 UASM_i_SW(p, pte, 0, ptr);
1604 # ifdef CONFIG_64BIT_PHYS_ADDR
1605 if (!cpu_has_64bits) {
1606 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1607 uasm_i_ori(p, pte, pte, hwmode);
1608 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1609 uasm_i_lw(p, pte, 0, ptr);
1616 * Check if PTE is present, if not then jump to LABEL. PTR points to
1617 * the page table where this PTE is located, PTE will be re-loaded
1618 * with it's original value.
1620 static void __cpuinit
1621 build_pte_present(u32 **p, struct uasm_reloc **r,
1622 int pte, int ptr, int scratch, enum label_id lid)
1624 int t = scratch >= 0 ? scratch : pte;
1627 if (use_bbit_insns()) {
1628 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1631 uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1632 uasm_il_beqz(p, r, t, lid);
1634 /* You lose the SMP race :-(*/
1635 iPTE_LW(p, pte, ptr);
1638 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1639 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1640 uasm_il_bnez(p, r, t, lid);
1642 /* You lose the SMP race :-(*/
1643 iPTE_LW(p, pte, ptr);
1647 /* Make PTE valid, store result in PTR. */
1648 static void __cpuinit
1649 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1652 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1654 iPTE_SW(p, r, pte, ptr, mode);
1658 * Check if PTE can be written to, if not branch to LABEL. Regardless
1659 * restore PTE with value from PTR when done.
1661 static void __cpuinit
1662 build_pte_writable(u32 **p, struct uasm_reloc **r,
1663 unsigned int pte, unsigned int ptr, int scratch,
1666 int t = scratch >= 0 ? scratch : pte;
1668 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1669 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1670 uasm_il_bnez(p, r, t, lid);
1672 /* You lose the SMP race :-(*/
1673 iPTE_LW(p, pte, ptr);
1678 /* Make PTE writable, update software status bits as well, then store
1681 static void __cpuinit
1682 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1685 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1688 iPTE_SW(p, r, pte, ptr, mode);
1692 * Check if PTE can be modified, if not branch to LABEL. Regardless
1693 * restore PTE with value from PTR when done.
1695 static void __cpuinit
1696 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1697 unsigned int pte, unsigned int ptr, int scratch,
1700 if (use_bbit_insns()) {
1701 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1704 int t = scratch >= 0 ? scratch : pte;
1705 uasm_i_andi(p, t, pte, _PAGE_WRITE);
1706 uasm_il_beqz(p, r, t, lid);
1708 /* You lose the SMP race :-(*/
1709 iPTE_LW(p, pte, ptr);
1713 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1717 * R3000 style TLB load/store/modify handlers.
1721 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1724 static void __cpuinit
1725 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1727 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1728 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1731 uasm_i_rfe(p); /* branch delay */
1735 * This places the pte into ENTRYLO0 and writes it with tlbwi
1736 * or tlbwr as appropriate. This is because the index register
1737 * may have the probe fail bit set as a result of a trap on a
1738 * kseg2 access, i.e. without refill. Then it returns.
1740 static void __cpuinit
1741 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1742 struct uasm_reloc **r, unsigned int pte,
1745 uasm_i_mfc0(p, tmp, C0_INDEX);
1746 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1747 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1748 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1749 uasm_i_tlbwi(p); /* cp0 delay */
1751 uasm_i_rfe(p); /* branch delay */
1752 uasm_l_r3000_write_probe_fail(l, *p);
1753 uasm_i_tlbwr(p); /* cp0 delay */
1755 uasm_i_rfe(p); /* branch delay */
1758 static void __cpuinit
1759 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1762 long pgdc = (long)pgd_current;
1764 uasm_i_mfc0(p, pte, C0_BADVADDR);
1765 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1766 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1767 uasm_i_srl(p, pte, pte, 22); /* load delay */
1768 uasm_i_sll(p, pte, pte, 2);
1769 uasm_i_addu(p, ptr, ptr, pte);
1770 uasm_i_mfc0(p, pte, C0_CONTEXT);
1771 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1772 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1773 uasm_i_addu(p, ptr, ptr, pte);
1774 uasm_i_lw(p, pte, 0, ptr);
1775 uasm_i_tlbp(p); /* load delay */
1778 static void __cpuinit build_r3000_tlb_load_handler(void)
1780 u32 *p = handle_tlbl;
1781 struct uasm_label *l = labels;
1782 struct uasm_reloc *r = relocs;
1784 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1785 memset(labels, 0, sizeof(labels));
1786 memset(relocs, 0, sizeof(relocs));
1788 build_r3000_tlbchange_handler_head(&p, K0, K1);
1789 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1790 uasm_i_nop(&p); /* load delay */
1791 build_make_valid(&p, &r, K0, K1);
1792 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1794 uasm_l_nopage_tlbl(&l, p);
1795 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1798 if ((p - handle_tlbl) > FASTPATH_SIZE)
1799 panic("TLB load handler fastpath space exceeded");
1801 uasm_resolve_relocs(relocs, labels);
1802 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1803 (unsigned int)(p - handle_tlbl));
1805 dump_handler("r3000_tlb_load", handle_tlbl, ARRAY_SIZE(handle_tlbl));
1808 static void __cpuinit build_r3000_tlb_store_handler(void)
1810 u32 *p = handle_tlbs;
1811 struct uasm_label *l = labels;
1812 struct uasm_reloc *r = relocs;
1814 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1815 memset(labels, 0, sizeof(labels));
1816 memset(relocs, 0, sizeof(relocs));
1818 build_r3000_tlbchange_handler_head(&p, K0, K1);
1819 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1820 uasm_i_nop(&p); /* load delay */
1821 build_make_write(&p, &r, K0, K1);
1822 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1824 uasm_l_nopage_tlbs(&l, p);
1825 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1828 if ((p - handle_tlbs) > FASTPATH_SIZE)
1829 panic("TLB store handler fastpath space exceeded");
1831 uasm_resolve_relocs(relocs, labels);
1832 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1833 (unsigned int)(p - handle_tlbs));
1835 dump_handler("r3000_tlb_store", handle_tlbs, ARRAY_SIZE(handle_tlbs));
1838 static void __cpuinit build_r3000_tlb_modify_handler(void)
1840 u32 *p = handle_tlbm;
1841 struct uasm_label *l = labels;
1842 struct uasm_reloc *r = relocs;
1844 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1845 memset(labels, 0, sizeof(labels));
1846 memset(relocs, 0, sizeof(relocs));
1848 build_r3000_tlbchange_handler_head(&p, K0, K1);
1849 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1850 uasm_i_nop(&p); /* load delay */
1851 build_make_write(&p, &r, K0, K1);
1852 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1854 uasm_l_nopage_tlbm(&l, p);
1855 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1858 if ((p - handle_tlbm) > FASTPATH_SIZE)
1859 panic("TLB modify handler fastpath space exceeded");
1861 uasm_resolve_relocs(relocs, labels);
1862 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1863 (unsigned int)(p - handle_tlbm));
1865 dump_handler("r3000_tlb_modify", handle_tlbm, ARRAY_SIZE(handle_tlbm));
1867 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1870 * R4000 style TLB load/store/modify handlers.
1872 static struct work_registers __cpuinit
1873 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1874 struct uasm_reloc **r)
1876 struct work_registers wr = build_get_work_registers(p);
1879 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1881 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1884 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1886 * For huge tlb entries, pmd doesn't contain an address but
1887 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1888 * see if we need to jump to huge tlb processing.
1890 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1893 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1894 UASM_i_LW(p, wr.r2, 0, wr.r2);
1895 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1896 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1897 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1900 uasm_l_smp_pgtable_change(l, *p);
1902 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1903 if (!m4kc_tlbp_war())
1904 build_tlb_probe_entry(p);
1908 static void __cpuinit
1909 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1910 struct uasm_reloc **r, unsigned int tmp,
1913 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1914 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1915 build_update_entries(p, tmp, ptr);
1916 build_tlb_write_entry(p, l, r, tlb_indexed);
1917 uasm_l_leave(l, *p);
1918 build_restore_work_registers(p);
1919 uasm_i_eret(p); /* return from trap */
1922 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1926 static void __cpuinit build_r4000_tlb_load_handler(void)
1928 u32 *p = handle_tlbl;
1929 struct uasm_label *l = labels;
1930 struct uasm_reloc *r = relocs;
1931 struct work_registers wr;
1933 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1934 memset(labels, 0, sizeof(labels));
1935 memset(relocs, 0, sizeof(relocs));
1937 if (bcm1250_m3_war()) {
1938 unsigned int segbits = 44;
1940 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1941 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1942 uasm_i_xor(&p, K0, K0, K1);
1943 uasm_i_dsrl_safe(&p, K1, K0, 62);
1944 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1945 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1946 uasm_i_or(&p, K0, K0, K1);
1947 uasm_il_bnez(&p, &r, K0, label_leave);
1948 /* No need for uasm_i_nop */
1951 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1952 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1953 if (m4kc_tlbp_war())
1954 build_tlb_probe_entry(&p);
1958 * If the page is not _PAGE_VALID, RI or XI could not
1959 * have triggered it. Skip the expensive test..
1961 if (use_bbit_insns()) {
1962 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1963 label_tlbl_goaround1);
1965 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1966 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1971 /* Examine entrylo 0 or 1 based on ptr. */
1972 if (use_bbit_insns()) {
1973 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1975 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1976 uasm_i_beqz(&p, wr.r3, 8);
1978 /* load it in the delay slot*/
1979 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1980 /* load it if ptr is odd */
1981 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1983 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1984 * XI must have triggered it.
1986 if (use_bbit_insns()) {
1987 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1989 uasm_l_tlbl_goaround1(&l, p);
1991 uasm_i_andi(&p, wr.r3, wr.r3, 2);
1992 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1995 uasm_l_tlbl_goaround1(&l, p);
1997 build_make_valid(&p, &r, wr.r1, wr.r2);
1998 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2000 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2002 * This is the entry point when build_r4000_tlbchange_handler_head
2003 * spots a huge page.
2005 uasm_l_tlb_huge_update(&l, p);
2006 iPTE_LW(&p, wr.r1, wr.r2);
2007 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2008 build_tlb_probe_entry(&p);
2012 * If the page is not _PAGE_VALID, RI or XI could not
2013 * have triggered it. Skip the expensive test..
2015 if (use_bbit_insns()) {
2016 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2017 label_tlbl_goaround2);
2019 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2020 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2025 /* Examine entrylo 0 or 1 based on ptr. */
2026 if (use_bbit_insns()) {
2027 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2029 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2030 uasm_i_beqz(&p, wr.r3, 8);
2032 /* load it in the delay slot*/
2033 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2034 /* load it if ptr is odd */
2035 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2037 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2038 * XI must have triggered it.
2040 if (use_bbit_insns()) {
2041 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2043 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2044 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2046 if (PM_DEFAULT_MASK == 0)
2049 * We clobbered C0_PAGEMASK, restore it. On the other branch
2050 * it is restored in build_huge_tlb_write_entry.
2052 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2054 uasm_l_tlbl_goaround2(&l, p);
2056 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2057 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2060 uasm_l_nopage_tlbl(&l, p);
2061 build_restore_work_registers(&p);
2062 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2065 if ((p - handle_tlbl) > FASTPATH_SIZE)
2066 panic("TLB load handler fastpath space exceeded");
2068 uasm_resolve_relocs(relocs, labels);
2069 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2070 (unsigned int)(p - handle_tlbl));
2072 dump_handler("r4000_tlb_load", handle_tlbl, ARRAY_SIZE(handle_tlbl));
2075 static void __cpuinit build_r4000_tlb_store_handler(void)
2077 u32 *p = handle_tlbs;
2078 struct uasm_label *l = labels;
2079 struct uasm_reloc *r = relocs;
2080 struct work_registers wr;
2082 memset(handle_tlbs, 0, sizeof(handle_tlbs));
2083 memset(labels, 0, sizeof(labels));
2084 memset(relocs, 0, sizeof(relocs));
2086 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2087 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2088 if (m4kc_tlbp_war())
2089 build_tlb_probe_entry(&p);
2090 build_make_write(&p, &r, wr.r1, wr.r2);
2091 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2093 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2095 * This is the entry point when
2096 * build_r4000_tlbchange_handler_head spots a huge page.
2098 uasm_l_tlb_huge_update(&l, p);
2099 iPTE_LW(&p, wr.r1, wr.r2);
2100 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2101 build_tlb_probe_entry(&p);
2102 uasm_i_ori(&p, wr.r1, wr.r1,
2103 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2104 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2107 uasm_l_nopage_tlbs(&l, p);
2108 build_restore_work_registers(&p);
2109 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2112 if ((p - handle_tlbs) > FASTPATH_SIZE)
2113 panic("TLB store handler fastpath space exceeded");
2115 uasm_resolve_relocs(relocs, labels);
2116 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2117 (unsigned int)(p - handle_tlbs));
2119 dump_handler("r4000_tlb_store", handle_tlbs, ARRAY_SIZE(handle_tlbs));
2122 static void __cpuinit build_r4000_tlb_modify_handler(void)
2124 u32 *p = handle_tlbm;
2125 struct uasm_label *l = labels;
2126 struct uasm_reloc *r = relocs;
2127 struct work_registers wr;
2129 memset(handle_tlbm, 0, sizeof(handle_tlbm));
2130 memset(labels, 0, sizeof(labels));
2131 memset(relocs, 0, sizeof(relocs));
2133 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2134 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2135 if (m4kc_tlbp_war())
2136 build_tlb_probe_entry(&p);
2137 /* Present and writable bits set, set accessed and dirty bits. */
2138 build_make_write(&p, &r, wr.r1, wr.r2);
2139 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2141 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2143 * This is the entry point when
2144 * build_r4000_tlbchange_handler_head spots a huge page.
2146 uasm_l_tlb_huge_update(&l, p);
2147 iPTE_LW(&p, wr.r1, wr.r2);
2148 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2149 build_tlb_probe_entry(&p);
2150 uasm_i_ori(&p, wr.r1, wr.r1,
2151 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2152 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2155 uasm_l_nopage_tlbm(&l, p);
2156 build_restore_work_registers(&p);
2157 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2160 if ((p - handle_tlbm) > FASTPATH_SIZE)
2161 panic("TLB modify handler fastpath space exceeded");
2163 uasm_resolve_relocs(relocs, labels);
2164 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2165 (unsigned int)(p - handle_tlbm));
2167 dump_handler("r4000_tlb_modify", handle_tlbm, ARRAY_SIZE(handle_tlbm));
2170 void __cpuinit build_tlb_refill_handler(void)
2173 * The refill handler is generated per-CPU, multi-node systems
2174 * may have local storage for it. The other handlers are only
2177 static int run_once = 0;
2179 output_pgtable_bits_defines();
2182 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2185 switch (current_cpu_type()) {
2193 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2194 build_r3000_tlb_refill_handler();
2196 build_r3000_tlb_load_handler();
2197 build_r3000_tlb_store_handler();
2198 build_r3000_tlb_modify_handler();
2202 panic("No R3000 TLB refill handler");
2208 panic("No R6000 TLB refill handler yet");
2212 panic("No R8000 TLB refill handler yet");
2217 scratch_reg = allocate_kscratch();
2218 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2219 build_r4000_setup_pgd();
2221 build_r4000_tlb_load_handler();
2222 build_r4000_tlb_store_handler();
2223 build_r4000_tlb_modify_handler();
2226 build_r4000_tlb_refill_handler();
2230 void __cpuinit flush_tlb_handlers(void)
2232 local_flush_icache_range((unsigned long)handle_tlbl,
2233 (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2234 local_flush_icache_range((unsigned long)handle_tlbs,
2235 (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2236 local_flush_icache_range((unsigned long)handle_tlbm,
2237 (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2238 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2239 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2240 (unsigned long)tlbmiss_handler_setup_pgd + sizeof(handle_tlbm));