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 __cpuinit uasm_bgezl_hazard(u32 **p,
187 struct uasm_reloc **r,
192 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
199 static void __cpuinit uasm_bgezl_label(struct uasm_label **l,
205 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
213 * pgtable bits are assigned dynamically depending on processor feature
214 * and statically based on kernel configuration. This spits out the actual
215 * values the kernel is using. Required to make sense from disassembled
216 * TLB exception handlers.
218 static void output_pgtable_bits_defines(void)
220 #define pr_define(fmt, ...) \
221 pr_debug("#define " fmt, ##__VA_ARGS__)
223 pr_debug("#include <asm/asm.h>\n");
224 pr_debug("#include <asm/regdef.h>\n");
227 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
228 pr_define("_PAGE_READ_SHIFT %d\n", _PAGE_READ_SHIFT);
229 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
230 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
231 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
232 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
233 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
234 pr_define("_PAGE_SPLITTING_SHIFT %d\n", _PAGE_SPLITTING_SHIFT);
237 #ifdef _PAGE_NO_EXEC_SHIFT
238 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
240 #ifdef _PAGE_NO_READ_SHIFT
241 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
244 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
245 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
246 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
247 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
251 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
255 pr_debug("LEAF(%s)\n", symbol);
257 pr_debug("\t.set push\n");
258 pr_debug("\t.set noreorder\n");
260 for (i = 0; i < count; i++)
261 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
263 pr_debug("\t.set\tpop\n");
265 pr_debug("\tEND(%s)\n", symbol);
268 /* The only general purpose registers allowed in TLB handlers. */
272 /* Some CP0 registers */
273 #define C0_INDEX 0, 0
274 #define C0_ENTRYLO0 2, 0
275 #define C0_TCBIND 2, 2
276 #define C0_ENTRYLO1 3, 0
277 #define C0_CONTEXT 4, 0
278 #define C0_PAGEMASK 5, 0
279 #define C0_BADVADDR 8, 0
280 #define C0_ENTRYHI 10, 0
282 #define C0_XCONTEXT 20, 0
285 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
287 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
290 /* The worst case length of the handler is around 18 instructions for
291 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
292 * Maximum space available is 32 instructions for R3000 and 64
293 * instructions for R4000.
295 * We deliberately chose a buffer size of 128, so we won't scribble
296 * over anything important on overflow before we panic.
298 static u32 tlb_handler[128] __cpuinitdata;
300 /* simply assume worst case size for labels and relocs */
301 static struct uasm_label labels[128] __cpuinitdata;
302 static struct uasm_reloc relocs[128] __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 */
633 panic("No TLB refill handler yet (CPU type: %d)",
634 current_cpu_data.cputype);
639 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
643 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
645 #ifdef CONFIG_64BIT_PHYS_ADDR
646 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
648 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
653 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
655 static __cpuinit void build_restore_pagemask(u32 **p,
656 struct uasm_reloc **r,
661 if (restore_scratch) {
662 /* Reset default page size */
663 if (PM_DEFAULT_MASK >> 16) {
664 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
665 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
666 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
667 uasm_il_b(p, r, lid);
668 } else if (PM_DEFAULT_MASK) {
669 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
670 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
671 uasm_il_b(p, r, lid);
673 uasm_i_mtc0(p, 0, C0_PAGEMASK);
674 uasm_il_b(p, r, lid);
677 UASM_i_MFC0(p, 1, 31, scratch_reg);
679 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
681 /* Reset default page size */
682 if (PM_DEFAULT_MASK >> 16) {
683 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
684 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
685 uasm_il_b(p, r, lid);
686 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
687 } else if (PM_DEFAULT_MASK) {
688 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
689 uasm_il_b(p, r, lid);
690 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
692 uasm_il_b(p, r, lid);
693 uasm_i_mtc0(p, 0, C0_PAGEMASK);
698 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
699 struct uasm_label **l,
700 struct uasm_reloc **r,
702 enum tlb_write_entry wmode,
705 /* Set huge page tlb entry size */
706 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
707 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
708 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
710 build_tlb_write_entry(p, l, r, wmode);
712 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
716 * Check if Huge PTE is present, if so then jump to LABEL.
718 static void __cpuinit
719 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
720 unsigned int pmd, int lid)
722 UASM_i_LW(p, tmp, 0, pmd);
723 if (use_bbit_insns()) {
724 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
726 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
727 uasm_il_bnez(p, r, tmp, lid);
731 static __cpuinit void build_huge_update_entries(u32 **p,
738 * A huge PTE describes an area the size of the
739 * configured huge page size. This is twice the
740 * of the large TLB entry size we intend to use.
741 * A TLB entry half the size of the configured
742 * huge page size is configured into entrylo0
743 * and entrylo1 to cover the contiguous huge PTE
746 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
748 /* We can clobber tmp. It isn't used after this.*/
750 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
752 build_convert_pte_to_entrylo(p, pte);
753 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
754 /* convert to entrylo1 */
756 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
758 UASM_i_ADDU(p, pte, pte, tmp);
760 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
763 static __cpuinit void build_huge_handler_tail(u32 **p,
764 struct uasm_reloc **r,
765 struct uasm_label **l,
770 UASM_i_SC(p, pte, 0, ptr);
771 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
772 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
774 UASM_i_SW(p, pte, 0, ptr);
776 build_huge_update_entries(p, pte, ptr);
777 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
779 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
783 * TMP and PTR are scratch.
784 * TMP will be clobbered, PTR will hold the pmd entry.
786 static void __cpuinit
787 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
788 unsigned int tmp, unsigned int ptr)
790 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
791 long pgdc = (long)pgd_current;
794 * The vmalloc handling is not in the hotpath.
796 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
798 if (check_for_high_segbits) {
800 * The kernel currently implicitely assumes that the
801 * MIPS SEGBITS parameter for the processor is
802 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
803 * allocate virtual addresses outside the maximum
804 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
805 * that doesn't prevent user code from accessing the
806 * higher xuseg addresses. Here, we make sure that
807 * everything but the lower xuseg addresses goes down
808 * the module_alloc/vmalloc path.
810 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
811 uasm_il_bnez(p, r, ptr, label_vmalloc);
813 uasm_il_bltz(p, r, tmp, label_vmalloc);
815 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
817 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
819 /* pgd is in pgd_reg */
820 UASM_i_MFC0(p, ptr, 31, pgd_reg);
823 * &pgd << 11 stored in CONTEXT [23..63].
825 UASM_i_MFC0(p, ptr, C0_CONTEXT);
827 /* Clear lower 23 bits of context. */
828 uasm_i_dins(p, ptr, 0, 0, 23);
830 /* 1 0 1 0 1 << 6 xkphys cached */
831 uasm_i_ori(p, ptr, ptr, 0x540);
832 uasm_i_drotr(p, ptr, ptr, 11);
834 #elif defined(CONFIG_SMP)
835 # ifdef CONFIG_MIPS_MT_SMTC
837 * SMTC uses TCBind value as "CPU" index
839 uasm_i_mfc0(p, ptr, C0_TCBIND);
840 uasm_i_dsrl_safe(p, ptr, ptr, 19);
843 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
846 uasm_i_dmfc0(p, ptr, C0_CONTEXT);
847 uasm_i_dsrl_safe(p, ptr, ptr, 23);
849 UASM_i_LA_mostly(p, tmp, pgdc);
850 uasm_i_daddu(p, ptr, ptr, tmp);
851 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
852 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
854 UASM_i_LA_mostly(p, ptr, pgdc);
855 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
858 uasm_l_vmalloc_done(l, *p);
860 /* get pgd offset in bytes */
861 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
863 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
864 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
865 #ifndef __PAGETABLE_PMD_FOLDED
866 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
867 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
868 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
869 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
870 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
875 * BVADDR is the faulting address, PTR is scratch.
876 * PTR will hold the pgd for vmalloc.
878 static void __cpuinit
879 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
880 unsigned int bvaddr, unsigned int ptr,
881 enum vmalloc64_mode mode)
883 long swpd = (long)swapper_pg_dir;
884 int single_insn_swpd;
885 int did_vmalloc_branch = 0;
887 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
889 uasm_l_vmalloc(l, *p);
891 if (mode != not_refill && check_for_high_segbits) {
892 if (single_insn_swpd) {
893 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
894 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
895 did_vmalloc_branch = 1;
898 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
901 if (!did_vmalloc_branch) {
902 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
903 uasm_il_b(p, r, label_vmalloc_done);
904 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
906 UASM_i_LA_mostly(p, ptr, swpd);
907 uasm_il_b(p, r, label_vmalloc_done);
908 if (uasm_in_compat_space_p(swpd))
909 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
911 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
914 if (mode != not_refill && check_for_high_segbits) {
915 uasm_l_large_segbits_fault(l, *p);
917 * We get here if we are an xsseg address, or if we are
918 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
920 * Ignoring xsseg (assume disabled so would generate
921 * (address errors?), the only remaining possibility
922 * is the upper xuseg addresses. On processors with
923 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
924 * addresses would have taken an address error. We try
925 * to mimic that here by taking a load/istream page
928 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
931 if (mode == refill_scratch) {
933 UASM_i_MFC0(p, 1, 31, scratch_reg);
935 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
942 #else /* !CONFIG_64BIT */
945 * TMP and PTR are scratch.
946 * TMP will be clobbered, PTR will hold the pgd entry.
948 static void __cpuinit __maybe_unused
949 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
951 long pgdc = (long)pgd_current;
953 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
955 #ifdef CONFIG_MIPS_MT_SMTC
957 * SMTC uses TCBind value as "CPU" index
959 uasm_i_mfc0(p, ptr, C0_TCBIND);
960 UASM_i_LA_mostly(p, tmp, pgdc);
961 uasm_i_srl(p, ptr, ptr, 19);
964 * smp_processor_id() << 3 is stored in CONTEXT.
966 uasm_i_mfc0(p, ptr, C0_CONTEXT);
967 UASM_i_LA_mostly(p, tmp, pgdc);
968 uasm_i_srl(p, ptr, ptr, 23);
970 uasm_i_addu(p, ptr, tmp, ptr);
972 UASM_i_LA_mostly(p, ptr, pgdc);
974 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
975 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
976 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
977 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
978 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
981 #endif /* !CONFIG_64BIT */
983 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
985 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
986 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
988 switch (current_cpu_type()) {
1005 UASM_i_SRL(p, ctx, ctx, shift);
1006 uasm_i_andi(p, ctx, ctx, mask);
1009 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1012 * Bug workaround for the Nevada. It seems as if under certain
1013 * circumstances the move from cp0_context might produce a
1014 * bogus result when the mfc0 instruction and its consumer are
1015 * in a different cacheline or a load instruction, probably any
1016 * memory reference, is between them.
1018 switch (current_cpu_type()) {
1020 UASM_i_LW(p, ptr, 0, ptr);
1021 GET_CONTEXT(p, tmp); /* get context reg */
1025 GET_CONTEXT(p, tmp); /* get context reg */
1026 UASM_i_LW(p, ptr, 0, ptr);
1030 build_adjust_context(p, tmp);
1031 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1034 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
1038 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1039 * Kernel is a special case. Only a few CPUs use it.
1041 #ifdef CONFIG_64BIT_PHYS_ADDR
1042 if (cpu_has_64bits) {
1043 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1044 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1046 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1047 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1048 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1050 uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1051 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1052 uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1054 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1056 int pte_off_even = sizeof(pte_t) / 2;
1057 int pte_off_odd = pte_off_even + sizeof(pte_t);
1059 /* The pte entries are pre-shifted */
1060 uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1061 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1062 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1063 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1066 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1067 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1068 if (r45k_bvahwbug())
1069 build_tlb_probe_entry(p);
1071 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1072 if (r4k_250MHZhwbug())
1073 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1074 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1075 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1077 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1078 if (r4k_250MHZhwbug())
1079 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1080 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1081 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1082 if (r45k_bvahwbug())
1083 uasm_i_mfc0(p, tmp, C0_INDEX);
1085 if (r4k_250MHZhwbug())
1086 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1087 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1091 struct mips_huge_tlb_info {
1093 int restore_scratch;
1094 bool need_reload_pte;
1097 static struct mips_huge_tlb_info __cpuinit
1098 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1099 struct uasm_reloc **r, unsigned int tmp,
1100 unsigned int ptr, int c0_scratch)
1102 struct mips_huge_tlb_info rv;
1103 unsigned int even, odd;
1104 int vmalloc_branch_delay_filled = 0;
1105 const int scratch = 1; /* Our extra working register */
1107 rv.huge_pte = scratch;
1108 rv.restore_scratch = 0;
1109 rv.need_reload_pte = false;
1111 if (check_for_high_segbits) {
1112 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1115 UASM_i_MFC0(p, ptr, 31, pgd_reg);
1117 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1119 if (c0_scratch >= 0)
1120 UASM_i_MTC0(p, scratch, 31, c0_scratch);
1122 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1124 uasm_i_dsrl_safe(p, scratch, tmp,
1125 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1126 uasm_il_bnez(p, r, scratch, label_vmalloc);
1128 if (pgd_reg == -1) {
1129 vmalloc_branch_delay_filled = 1;
1130 /* Clear lower 23 bits of context. */
1131 uasm_i_dins(p, ptr, 0, 0, 23);
1135 UASM_i_MFC0(p, ptr, 31, pgd_reg);
1137 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1139 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1141 if (c0_scratch >= 0)
1142 UASM_i_MTC0(p, scratch, 31, c0_scratch);
1144 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1147 /* Clear lower 23 bits of context. */
1148 uasm_i_dins(p, ptr, 0, 0, 23);
1150 uasm_il_bltz(p, r, tmp, label_vmalloc);
1153 if (pgd_reg == -1) {
1154 vmalloc_branch_delay_filled = 1;
1155 /* 1 0 1 0 1 << 6 xkphys cached */
1156 uasm_i_ori(p, ptr, ptr, 0x540);
1157 uasm_i_drotr(p, ptr, ptr, 11);
1160 #ifdef __PAGETABLE_PMD_FOLDED
1161 #define LOC_PTEP scratch
1163 #define LOC_PTEP ptr
1166 if (!vmalloc_branch_delay_filled)
1167 /* get pgd offset in bytes */
1168 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1170 uasm_l_vmalloc_done(l, *p);
1174 * fall-through case = badvaddr *pgd_current
1175 * vmalloc case = badvaddr swapper_pg_dir
1178 if (vmalloc_branch_delay_filled)
1179 /* get pgd offset in bytes */
1180 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1182 #ifdef __PAGETABLE_PMD_FOLDED
1183 GET_CONTEXT(p, tmp); /* get context reg */
1185 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1187 if (use_lwx_insns()) {
1188 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1190 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1191 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1194 #ifndef __PAGETABLE_PMD_FOLDED
1195 /* get pmd offset in bytes */
1196 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1197 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1198 GET_CONTEXT(p, tmp); /* get context reg */
1200 if (use_lwx_insns()) {
1201 UASM_i_LWX(p, scratch, scratch, ptr);
1203 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1204 UASM_i_LW(p, scratch, 0, ptr);
1207 /* Adjust the context during the load latency. */
1208 build_adjust_context(p, tmp);
1210 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1211 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1213 * The in the LWX case we don't want to do the load in the
1214 * delay slot. It cannot issue in the same cycle and may be
1215 * speculative and unneeded.
1217 if (use_lwx_insns())
1219 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1222 /* build_update_entries */
1223 if (use_lwx_insns()) {
1226 UASM_i_LWX(p, even, scratch, tmp);
1227 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1228 UASM_i_LWX(p, odd, scratch, tmp);
1230 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1233 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1234 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1237 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1238 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1239 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1241 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1242 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1243 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1245 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1247 if (c0_scratch >= 0) {
1248 UASM_i_MFC0(p, scratch, 31, c0_scratch);
1249 build_tlb_write_entry(p, l, r, tlb_random);
1250 uasm_l_leave(l, *p);
1251 rv.restore_scratch = 1;
1252 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1253 build_tlb_write_entry(p, l, r, tlb_random);
1254 uasm_l_leave(l, *p);
1255 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1257 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1258 build_tlb_write_entry(p, l, r, tlb_random);
1259 uasm_l_leave(l, *p);
1260 rv.restore_scratch = 1;
1263 uasm_i_eret(p); /* return from trap */
1269 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1270 * because EXL == 0. If we wrap, we can also use the 32 instruction
1271 * slots before the XTLB refill exception handler which belong to the
1272 * unused TLB refill exception.
1274 #define MIPS64_REFILL_INSNS 32
1276 static void __cpuinit build_r4000_tlb_refill_handler(void)
1278 u32 *p = tlb_handler;
1279 struct uasm_label *l = labels;
1280 struct uasm_reloc *r = relocs;
1282 unsigned int final_len;
1283 struct mips_huge_tlb_info htlb_info __maybe_unused;
1284 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1286 memset(tlb_handler, 0, sizeof(tlb_handler));
1287 memset(labels, 0, sizeof(labels));
1288 memset(relocs, 0, sizeof(relocs));
1289 memset(final_handler, 0, sizeof(final_handler));
1291 if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1292 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1294 vmalloc_mode = refill_scratch;
1296 htlb_info.huge_pte = K0;
1297 htlb_info.restore_scratch = 0;
1298 htlb_info.need_reload_pte = true;
1299 vmalloc_mode = refill_noscratch;
1301 * create the plain linear handler
1303 if (bcm1250_m3_war()) {
1304 unsigned int segbits = 44;
1306 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1307 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1308 uasm_i_xor(&p, K0, K0, K1);
1309 uasm_i_dsrl_safe(&p, K1, K0, 62);
1310 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1311 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1312 uasm_i_or(&p, K0, K0, K1);
1313 uasm_il_bnez(&p, &r, K0, label_leave);
1314 /* No need for uasm_i_nop */
1318 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1320 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1323 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1324 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1327 build_get_ptep(&p, K0, K1);
1328 build_update_entries(&p, K0, K1);
1329 build_tlb_write_entry(&p, &l, &r, tlb_random);
1330 uasm_l_leave(&l, p);
1331 uasm_i_eret(&p); /* return from trap */
1333 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1334 uasm_l_tlb_huge_update(&l, p);
1335 if (htlb_info.need_reload_pte)
1336 UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1337 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1338 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1339 htlb_info.restore_scratch);
1343 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1347 * Overflow check: For the 64bit handler, we need at least one
1348 * free instruction slot for the wrap-around branch. In worst
1349 * case, if the intended insertion point is a delay slot, we
1350 * need three, with the second nop'ed and the third being
1353 /* Loongson2 ebase is different than r4k, we have more space */
1354 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1355 if ((p - tlb_handler) > 64)
1356 panic("TLB refill handler space exceeded");
1358 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1359 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1360 && uasm_insn_has_bdelay(relocs,
1361 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1362 panic("TLB refill handler space exceeded");
1366 * Now fold the handler in the TLB refill handler space.
1368 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1370 /* Simplest case, just copy the handler. */
1371 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1372 final_len = p - tlb_handler;
1373 #else /* CONFIG_64BIT */
1374 f = final_handler + MIPS64_REFILL_INSNS;
1375 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1376 /* Just copy the handler. */
1377 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1378 final_len = p - tlb_handler;
1380 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1381 const enum label_id ls = label_tlb_huge_update;
1383 const enum label_id ls = label_vmalloc;
1389 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1391 BUG_ON(i == ARRAY_SIZE(labels));
1392 split = labels[i].addr;
1395 * See if we have overflown one way or the other.
1397 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1398 split < p - MIPS64_REFILL_INSNS)
1403 * Split two instructions before the end. One
1404 * for the branch and one for the instruction
1405 * in the delay slot.
1407 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1410 * If the branch would fall in a delay slot,
1411 * we must back up an additional instruction
1412 * so that it is no longer in a delay slot.
1414 if (uasm_insn_has_bdelay(relocs, split - 1))
1417 /* Copy first part of the handler. */
1418 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1419 f += split - tlb_handler;
1422 /* Insert branch. */
1423 uasm_l_split(&l, final_handler);
1424 uasm_il_b(&f, &r, label_split);
1425 if (uasm_insn_has_bdelay(relocs, split))
1428 uasm_copy_handler(relocs, labels,
1429 split, split + 1, f);
1430 uasm_move_labels(labels, f, f + 1, -1);
1436 /* Copy the rest of the handler. */
1437 uasm_copy_handler(relocs, labels, split, p, final_handler);
1438 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1441 #endif /* CONFIG_64BIT */
1443 uasm_resolve_relocs(relocs, labels);
1444 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1447 memcpy((void *)ebase, final_handler, 0x100);
1449 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1453 * 128 instructions for the fastpath handler is generous and should
1454 * never be exceeded.
1456 #define FASTPATH_SIZE 128
1458 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1459 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1460 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1461 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1462 u32 tlbmiss_handler_setup_pgd_array[16] __cacheline_aligned;
1464 static void __cpuinit build_r4000_setup_pgd(void)
1468 u32 *p = tlbmiss_handler_setup_pgd_array;
1469 struct uasm_label *l = labels;
1470 struct uasm_reloc *r = relocs;
1472 memset(tlbmiss_handler_setup_pgd_array, 0, sizeof(tlbmiss_handler_setup_pgd_array));
1473 memset(labels, 0, sizeof(labels));
1474 memset(relocs, 0, sizeof(relocs));
1476 pgd_reg = allocate_kscratch();
1478 if (pgd_reg == -1) {
1479 /* PGD << 11 in c0_Context */
1481 * If it is a ckseg0 address, convert to a physical
1482 * address. Shifting right by 29 and adding 4 will
1483 * result in zero for these addresses.
1486 UASM_i_SRA(&p, a1, a0, 29);
1487 UASM_i_ADDIU(&p, a1, a1, 4);
1488 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1490 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1491 uasm_l_tlbl_goaround1(&l, p);
1492 UASM_i_SLL(&p, a0, a0, 11);
1494 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1496 /* PGD in c0_KScratch */
1498 UASM_i_MTC0(&p, a0, 31, pgd_reg);
1500 if (p - tlbmiss_handler_setup_pgd_array > ARRAY_SIZE(tlbmiss_handler_setup_pgd_array))
1501 panic("tlbmiss_handler_setup_pgd_array space exceeded");
1502 uasm_resolve_relocs(relocs, labels);
1503 pr_debug("Wrote tlbmiss_handler_setup_pgd_array (%u instructions).\n",
1504 (unsigned int)(p - tlbmiss_handler_setup_pgd_array));
1506 dump_handler("tlbmiss_handler",
1507 tlbmiss_handler_setup_pgd_array,
1508 ARRAY_SIZE(tlbmiss_handler_setup_pgd_array));
1512 static void __cpuinit
1513 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1516 # ifdef CONFIG_64BIT_PHYS_ADDR
1518 uasm_i_lld(p, pte, 0, ptr);
1521 UASM_i_LL(p, pte, 0, ptr);
1523 # ifdef CONFIG_64BIT_PHYS_ADDR
1525 uasm_i_ld(p, pte, 0, ptr);
1528 UASM_i_LW(p, pte, 0, ptr);
1532 static void __cpuinit
1533 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1536 #ifdef CONFIG_64BIT_PHYS_ADDR
1537 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1540 uasm_i_ori(p, pte, pte, mode);
1542 # ifdef CONFIG_64BIT_PHYS_ADDR
1544 uasm_i_scd(p, pte, 0, ptr);
1547 UASM_i_SC(p, pte, 0, ptr);
1549 if (r10000_llsc_war())
1550 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1552 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1554 # ifdef CONFIG_64BIT_PHYS_ADDR
1555 if (!cpu_has_64bits) {
1556 /* no uasm_i_nop needed */
1557 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1558 uasm_i_ori(p, pte, pte, hwmode);
1559 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1560 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1561 /* no uasm_i_nop needed */
1562 uasm_i_lw(p, pte, 0, ptr);
1569 # ifdef CONFIG_64BIT_PHYS_ADDR
1571 uasm_i_sd(p, pte, 0, ptr);
1574 UASM_i_SW(p, pte, 0, ptr);
1576 # ifdef CONFIG_64BIT_PHYS_ADDR
1577 if (!cpu_has_64bits) {
1578 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1579 uasm_i_ori(p, pte, pte, hwmode);
1580 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1581 uasm_i_lw(p, pte, 0, ptr);
1588 * Check if PTE is present, if not then jump to LABEL. PTR points to
1589 * the page table where this PTE is located, PTE will be re-loaded
1590 * with it's original value.
1592 static void __cpuinit
1593 build_pte_present(u32 **p, struct uasm_reloc **r,
1594 int pte, int ptr, int scratch, enum label_id lid)
1596 int t = scratch >= 0 ? scratch : pte;
1599 if (use_bbit_insns()) {
1600 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1603 uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1604 uasm_il_beqz(p, r, t, lid);
1606 /* You lose the SMP race :-(*/
1607 iPTE_LW(p, pte, ptr);
1610 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1611 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1612 uasm_il_bnez(p, r, t, lid);
1614 /* You lose the SMP race :-(*/
1615 iPTE_LW(p, pte, ptr);
1619 /* Make PTE valid, store result in PTR. */
1620 static void __cpuinit
1621 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1624 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1626 iPTE_SW(p, r, pte, ptr, mode);
1630 * Check if PTE can be written to, if not branch to LABEL. Regardless
1631 * restore PTE with value from PTR when done.
1633 static void __cpuinit
1634 build_pte_writable(u32 **p, struct uasm_reloc **r,
1635 unsigned int pte, unsigned int ptr, int scratch,
1638 int t = scratch >= 0 ? scratch : pte;
1640 uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1641 uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1642 uasm_il_bnez(p, r, t, lid);
1644 /* You lose the SMP race :-(*/
1645 iPTE_LW(p, pte, ptr);
1650 /* Make PTE writable, update software status bits as well, then store
1653 static void __cpuinit
1654 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1657 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1660 iPTE_SW(p, r, pte, ptr, mode);
1664 * Check if PTE can be modified, if not branch to LABEL. Regardless
1665 * restore PTE with value from PTR when done.
1667 static void __cpuinit
1668 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1669 unsigned int pte, unsigned int ptr, int scratch,
1672 if (use_bbit_insns()) {
1673 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1676 int t = scratch >= 0 ? scratch : pte;
1677 uasm_i_andi(p, t, pte, _PAGE_WRITE);
1678 uasm_il_beqz(p, r, t, lid);
1680 /* You lose the SMP race :-(*/
1681 iPTE_LW(p, pte, ptr);
1685 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1689 * R3000 style TLB load/store/modify handlers.
1693 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1696 static void __cpuinit
1697 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1699 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1700 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1703 uasm_i_rfe(p); /* branch delay */
1707 * This places the pte into ENTRYLO0 and writes it with tlbwi
1708 * or tlbwr as appropriate. This is because the index register
1709 * may have the probe fail bit set as a result of a trap on a
1710 * kseg2 access, i.e. without refill. Then it returns.
1712 static void __cpuinit
1713 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1714 struct uasm_reloc **r, unsigned int pte,
1717 uasm_i_mfc0(p, tmp, C0_INDEX);
1718 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1719 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1720 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1721 uasm_i_tlbwi(p); /* cp0 delay */
1723 uasm_i_rfe(p); /* branch delay */
1724 uasm_l_r3000_write_probe_fail(l, *p);
1725 uasm_i_tlbwr(p); /* cp0 delay */
1727 uasm_i_rfe(p); /* branch delay */
1730 static void __cpuinit
1731 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1734 long pgdc = (long)pgd_current;
1736 uasm_i_mfc0(p, pte, C0_BADVADDR);
1737 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1738 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1739 uasm_i_srl(p, pte, pte, 22); /* load delay */
1740 uasm_i_sll(p, pte, pte, 2);
1741 uasm_i_addu(p, ptr, ptr, pte);
1742 uasm_i_mfc0(p, pte, C0_CONTEXT);
1743 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1744 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1745 uasm_i_addu(p, ptr, ptr, pte);
1746 uasm_i_lw(p, pte, 0, ptr);
1747 uasm_i_tlbp(p); /* load delay */
1750 static void __cpuinit build_r3000_tlb_load_handler(void)
1752 u32 *p = handle_tlbl;
1753 struct uasm_label *l = labels;
1754 struct uasm_reloc *r = relocs;
1756 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1757 memset(labels, 0, sizeof(labels));
1758 memset(relocs, 0, sizeof(relocs));
1760 build_r3000_tlbchange_handler_head(&p, K0, K1);
1761 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1762 uasm_i_nop(&p); /* load delay */
1763 build_make_valid(&p, &r, K0, K1);
1764 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1766 uasm_l_nopage_tlbl(&l, p);
1767 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1770 if ((p - handle_tlbl) > FASTPATH_SIZE)
1771 panic("TLB load handler fastpath space exceeded");
1773 uasm_resolve_relocs(relocs, labels);
1774 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1775 (unsigned int)(p - handle_tlbl));
1777 dump_handler("r3000_tlb_load", handle_tlbl, ARRAY_SIZE(handle_tlbl));
1780 static void __cpuinit build_r3000_tlb_store_handler(void)
1782 u32 *p = handle_tlbs;
1783 struct uasm_label *l = labels;
1784 struct uasm_reloc *r = relocs;
1786 memset(handle_tlbs, 0, sizeof(handle_tlbs));
1787 memset(labels, 0, sizeof(labels));
1788 memset(relocs, 0, sizeof(relocs));
1790 build_r3000_tlbchange_handler_head(&p, K0, K1);
1791 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1792 uasm_i_nop(&p); /* load delay */
1793 build_make_write(&p, &r, K0, K1);
1794 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1796 uasm_l_nopage_tlbs(&l, p);
1797 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1800 if ((p - handle_tlbs) > FASTPATH_SIZE)
1801 panic("TLB store handler fastpath space exceeded");
1803 uasm_resolve_relocs(relocs, labels);
1804 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1805 (unsigned int)(p - handle_tlbs));
1807 dump_handler("r3000_tlb_store", handle_tlbs, ARRAY_SIZE(handle_tlbs));
1810 static void __cpuinit build_r3000_tlb_modify_handler(void)
1812 u32 *p = handle_tlbm;
1813 struct uasm_label *l = labels;
1814 struct uasm_reloc *r = relocs;
1816 memset(handle_tlbm, 0, sizeof(handle_tlbm));
1817 memset(labels, 0, sizeof(labels));
1818 memset(relocs, 0, sizeof(relocs));
1820 build_r3000_tlbchange_handler_head(&p, K0, K1);
1821 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1822 uasm_i_nop(&p); /* load delay */
1823 build_make_write(&p, &r, K0, K1);
1824 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1826 uasm_l_nopage_tlbm(&l, p);
1827 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1830 if ((p - handle_tlbm) > FASTPATH_SIZE)
1831 panic("TLB modify handler fastpath space exceeded");
1833 uasm_resolve_relocs(relocs, labels);
1834 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1835 (unsigned int)(p - handle_tlbm));
1837 dump_handler("r3000_tlb_modify", handle_tlbm, ARRAY_SIZE(handle_tlbm));
1839 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1842 * R4000 style TLB load/store/modify handlers.
1844 static struct work_registers __cpuinit
1845 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1846 struct uasm_reloc **r)
1848 struct work_registers wr = build_get_work_registers(p);
1851 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1853 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1856 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1858 * For huge tlb entries, pmd doesn't contain an address but
1859 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1860 * see if we need to jump to huge tlb processing.
1862 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1865 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1866 UASM_i_LW(p, wr.r2, 0, wr.r2);
1867 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1868 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1869 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1872 uasm_l_smp_pgtable_change(l, *p);
1874 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1875 if (!m4kc_tlbp_war())
1876 build_tlb_probe_entry(p);
1880 static void __cpuinit
1881 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1882 struct uasm_reloc **r, unsigned int tmp,
1885 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1886 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1887 build_update_entries(p, tmp, ptr);
1888 build_tlb_write_entry(p, l, r, tlb_indexed);
1889 uasm_l_leave(l, *p);
1890 build_restore_work_registers(p);
1891 uasm_i_eret(p); /* return from trap */
1894 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1898 static void __cpuinit build_r4000_tlb_load_handler(void)
1900 u32 *p = handle_tlbl;
1901 struct uasm_label *l = labels;
1902 struct uasm_reloc *r = relocs;
1903 struct work_registers wr;
1905 memset(handle_tlbl, 0, sizeof(handle_tlbl));
1906 memset(labels, 0, sizeof(labels));
1907 memset(relocs, 0, sizeof(relocs));
1909 if (bcm1250_m3_war()) {
1910 unsigned int segbits = 44;
1912 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1913 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1914 uasm_i_xor(&p, K0, K0, K1);
1915 uasm_i_dsrl_safe(&p, K1, K0, 62);
1916 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1917 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1918 uasm_i_or(&p, K0, K0, K1);
1919 uasm_il_bnez(&p, &r, K0, label_leave);
1920 /* No need for uasm_i_nop */
1923 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1924 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1925 if (m4kc_tlbp_war())
1926 build_tlb_probe_entry(&p);
1930 * If the page is not _PAGE_VALID, RI or XI could not
1931 * have triggered it. Skip the expensive test..
1933 if (use_bbit_insns()) {
1934 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1935 label_tlbl_goaround1);
1937 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1938 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1943 /* Examine entrylo 0 or 1 based on ptr. */
1944 if (use_bbit_insns()) {
1945 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1947 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1948 uasm_i_beqz(&p, wr.r3, 8);
1950 /* load it in the delay slot*/
1951 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1952 /* load it if ptr is odd */
1953 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1955 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1956 * XI must have triggered it.
1958 if (use_bbit_insns()) {
1959 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1961 uasm_l_tlbl_goaround1(&l, p);
1963 uasm_i_andi(&p, wr.r3, wr.r3, 2);
1964 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1967 uasm_l_tlbl_goaround1(&l, p);
1969 build_make_valid(&p, &r, wr.r1, wr.r2);
1970 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1972 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1974 * This is the entry point when build_r4000_tlbchange_handler_head
1975 * spots a huge page.
1977 uasm_l_tlb_huge_update(&l, p);
1978 iPTE_LW(&p, wr.r1, wr.r2);
1979 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1980 build_tlb_probe_entry(&p);
1984 * If the page is not _PAGE_VALID, RI or XI could not
1985 * have triggered it. Skip the expensive test..
1987 if (use_bbit_insns()) {
1988 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1989 label_tlbl_goaround2);
1991 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1992 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1997 /* Examine entrylo 0 or 1 based on ptr. */
1998 if (use_bbit_insns()) {
1999 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2001 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2002 uasm_i_beqz(&p, wr.r3, 8);
2004 /* load it in the delay slot*/
2005 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2006 /* load it if ptr is odd */
2007 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2009 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2010 * XI must have triggered it.
2012 if (use_bbit_insns()) {
2013 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2015 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2016 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2018 if (PM_DEFAULT_MASK == 0)
2021 * We clobbered C0_PAGEMASK, restore it. On the other branch
2022 * it is restored in build_huge_tlb_write_entry.
2024 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2026 uasm_l_tlbl_goaround2(&l, p);
2028 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2029 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2032 uasm_l_nopage_tlbl(&l, p);
2033 build_restore_work_registers(&p);
2034 #ifdef CONFIG_CPU_MICROMIPS
2035 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2036 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2037 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2041 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2044 if ((p - handle_tlbl) > FASTPATH_SIZE)
2045 panic("TLB load handler fastpath space exceeded");
2047 uasm_resolve_relocs(relocs, labels);
2048 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2049 (unsigned int)(p - handle_tlbl));
2051 dump_handler("r4000_tlb_load", handle_tlbl, ARRAY_SIZE(handle_tlbl));
2054 static void __cpuinit build_r4000_tlb_store_handler(void)
2056 u32 *p = handle_tlbs;
2057 struct uasm_label *l = labels;
2058 struct uasm_reloc *r = relocs;
2059 struct work_registers wr;
2061 memset(handle_tlbs, 0, sizeof(handle_tlbs));
2062 memset(labels, 0, sizeof(labels));
2063 memset(relocs, 0, sizeof(relocs));
2065 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2066 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2067 if (m4kc_tlbp_war())
2068 build_tlb_probe_entry(&p);
2069 build_make_write(&p, &r, wr.r1, wr.r2);
2070 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2072 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2074 * This is the entry point when
2075 * build_r4000_tlbchange_handler_head spots a huge page.
2077 uasm_l_tlb_huge_update(&l, p);
2078 iPTE_LW(&p, wr.r1, wr.r2);
2079 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2080 build_tlb_probe_entry(&p);
2081 uasm_i_ori(&p, wr.r1, wr.r1,
2082 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2083 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2086 uasm_l_nopage_tlbs(&l, p);
2087 build_restore_work_registers(&p);
2088 #ifdef CONFIG_CPU_MICROMIPS
2089 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2090 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2091 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2095 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2098 if ((p - handle_tlbs) > FASTPATH_SIZE)
2099 panic("TLB store handler fastpath space exceeded");
2101 uasm_resolve_relocs(relocs, labels);
2102 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2103 (unsigned int)(p - handle_tlbs));
2105 dump_handler("r4000_tlb_store", handle_tlbs, ARRAY_SIZE(handle_tlbs));
2108 static void __cpuinit build_r4000_tlb_modify_handler(void)
2110 u32 *p = handle_tlbm;
2111 struct uasm_label *l = labels;
2112 struct uasm_reloc *r = relocs;
2113 struct work_registers wr;
2115 memset(handle_tlbm, 0, sizeof(handle_tlbm));
2116 memset(labels, 0, sizeof(labels));
2117 memset(relocs, 0, sizeof(relocs));
2119 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2120 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2121 if (m4kc_tlbp_war())
2122 build_tlb_probe_entry(&p);
2123 /* Present and writable bits set, set accessed and dirty bits. */
2124 build_make_write(&p, &r, wr.r1, wr.r2);
2125 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2127 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2129 * This is the entry point when
2130 * build_r4000_tlbchange_handler_head spots a huge page.
2132 uasm_l_tlb_huge_update(&l, p);
2133 iPTE_LW(&p, wr.r1, wr.r2);
2134 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2135 build_tlb_probe_entry(&p);
2136 uasm_i_ori(&p, wr.r1, wr.r1,
2137 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2138 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2141 uasm_l_nopage_tlbm(&l, p);
2142 build_restore_work_registers(&p);
2143 #ifdef CONFIG_CPU_MICROMIPS
2144 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2145 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2146 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2150 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2153 if ((p - handle_tlbm) > FASTPATH_SIZE)
2154 panic("TLB modify handler fastpath space exceeded");
2156 uasm_resolve_relocs(relocs, labels);
2157 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2158 (unsigned int)(p - handle_tlbm));
2160 dump_handler("r4000_tlb_modify", handle_tlbm, ARRAY_SIZE(handle_tlbm));
2163 void __cpuinit build_tlb_refill_handler(void)
2166 * The refill handler is generated per-CPU, multi-node systems
2167 * may have local storage for it. The other handlers are only
2170 static int run_once = 0;
2172 output_pgtable_bits_defines();
2175 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2178 switch (current_cpu_type()) {
2186 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2187 if (cpu_has_local_ebase)
2188 build_r3000_tlb_refill_handler();
2190 if (!cpu_has_local_ebase)
2191 build_r3000_tlb_refill_handler();
2192 build_r3000_tlb_load_handler();
2193 build_r3000_tlb_store_handler();
2194 build_r3000_tlb_modify_handler();
2198 panic("No R3000 TLB refill handler");
2204 panic("No R6000 TLB refill handler yet");
2208 panic("No R8000 TLB refill handler yet");
2213 scratch_reg = allocate_kscratch();
2214 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2215 build_r4000_setup_pgd();
2217 build_r4000_tlb_load_handler();
2218 build_r4000_tlb_store_handler();
2219 build_r4000_tlb_modify_handler();
2220 if (!cpu_has_local_ebase)
2221 build_r4000_tlb_refill_handler();
2224 if (cpu_has_local_ebase)
2225 build_r4000_tlb_refill_handler();
2229 void __cpuinit flush_tlb_handlers(void)
2231 local_flush_icache_range((unsigned long)handle_tlbl,
2232 (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2233 local_flush_icache_range((unsigned long)handle_tlbs,
2234 (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2235 local_flush_icache_range((unsigned long)handle_tlbm,
2236 (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2237 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2238 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd_array,
2239 (unsigned long)tlbmiss_handler_setup_pgd_array + sizeof(handle_tlbm));