2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
6 * Synthesize TLB refill handlers at runtime.
8 * Copyright (C) 2004, 2005, 2006, 2008 Thiemo Seufer
9 * Copyright (C) 2005, 2007, 2008, 2009 Maciej W. Rozycki
10 * Copyright (C) 2006 Ralf Baechle (ralf@linux-mips.org)
11 * Copyright (C) 2008, 2009 Cavium Networks, Inc.
12 * Copyright (C) 2011 MIPS Technologies, Inc.
14 * ... and the days got worse and worse and now you see
15 * I've gone completly out of my mind.
17 * They're coming to take me a away haha
18 * they're coming to take me a away hoho hihi haha
19 * to the funny farm where code is beautiful all the time ...
21 * (Condolences to Napoleon XIV)
24 #include <linux/bug.h>
25 #include <linux/kernel.h>
26 #include <linux/types.h>
27 #include <linux/smp.h>
28 #include <linux/string.h>
29 #include <linux/cache.h>
31 #include <asm/cacheflush.h>
32 #include <asm/cpu-type.h>
33 #include <asm/pgtable.h>
36 #include <asm/setup.h>
38 static int __cpuinitdata mips_xpa_disabled;
40 static int __init xpa_disable(char *s)
42 mips_xpa_disabled = 1;
47 __setup("noxpa", xpa_disable);
50 * TLB load/store/modify handlers.
52 * Only the fastpath gets synthesized at runtime, the slowpath for
53 * do_page_fault remains normal asm.
55 extern void tlb_do_page_fault_0(void);
56 extern void tlb_do_page_fault_1(void);
58 struct work_registers {
67 } ____cacheline_aligned_in_smp;
69 static struct tlb_reg_save handler_reg_save[NR_CPUS];
71 static inline int r45k_bvahwbug(void)
73 /* XXX: We should probe for the presence of this bug, but we don't. */
77 static inline int r4k_250MHZhwbug(void)
79 /* XXX: We should probe for the presence of this bug, but we don't. */
83 static inline int __maybe_unused bcm1250_m3_war(void)
85 return BCM1250_M3_WAR;
88 static inline int __maybe_unused r10000_llsc_war(void)
90 return R10000_LLSC_WAR;
93 static int use_bbit_insns(void)
95 switch (current_cpu_type()) {
96 case CPU_CAVIUM_OCTEON:
97 case CPU_CAVIUM_OCTEON_PLUS:
98 case CPU_CAVIUM_OCTEON2:
99 case CPU_CAVIUM_OCTEON3:
106 static int use_lwx_insns(void)
108 switch (current_cpu_type()) {
109 case CPU_CAVIUM_OCTEON2:
110 case CPU_CAVIUM_OCTEON3:
116 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
117 CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
118 static bool scratchpad_available(void)
122 static int scratchpad_offset(int i)
125 * CVMSEG starts at address -32768 and extends for
126 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
128 i += 1; /* Kernel use starts at the top and works down. */
129 return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
132 static bool scratchpad_available(void)
136 static int scratchpad_offset(int i)
139 /* Really unreachable, but evidently some GCC want this. */
144 * Found by experiment: At least some revisions of the 4kc throw under
145 * some circumstances a machine check exception, triggered by invalid
146 * values in the index register. Delaying the tlbp instruction until
147 * after the next branch, plus adding an additional nop in front of
148 * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
149 * why; it's not an issue caused by the core RTL.
152 static int m4kc_tlbp_war(void)
154 return (current_cpu_data.processor_id & 0xffff00) ==
155 (PRID_COMP_MIPS | PRID_IMP_4KC);
158 /* Handle labels (which must be positive integers). */
160 label_second_part = 1,
165 label_split = label_tlbw_hazard_0 + 8,
166 label_tlbl_goaround1,
167 label_tlbl_goaround2,
171 label_smp_pgtable_change,
172 label_r3000_write_probe_fail,
173 label_large_segbits_fault,
174 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
175 label_tlb_huge_update,
179 UASM_L_LA(_second_part)
182 UASM_L_LA(_vmalloc_done)
183 /* _tlbw_hazard_x is handled differently. */
185 UASM_L_LA(_tlbl_goaround1)
186 UASM_L_LA(_tlbl_goaround2)
187 UASM_L_LA(_nopage_tlbl)
188 UASM_L_LA(_nopage_tlbs)
189 UASM_L_LA(_nopage_tlbm)
190 UASM_L_LA(_smp_pgtable_change)
191 UASM_L_LA(_r3000_write_probe_fail)
192 UASM_L_LA(_large_segbits_fault)
193 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
194 UASM_L_LA(_tlb_huge_update)
197 static int hazard_instance;
199 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
203 uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
210 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
214 uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
222 * pgtable bits are assigned dynamically depending on processor feature
223 * and statically based on kernel configuration. This spits out the actual
224 * values the kernel is using. Required to make sense from disassembled
225 * TLB exception handlers.
227 static void output_pgtable_bits_defines(void)
229 #define pr_define(fmt, ...) \
230 pr_debug("#define " fmt, ##__VA_ARGS__)
232 pr_debug("#include <asm/asm.h>\n");
233 pr_debug("#include <asm/regdef.h>\n");
236 pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
237 pr_define("_PAGE_READ_SHIFT %d\n", _PAGE_READ_SHIFT);
238 pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
239 pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
240 pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
241 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
242 pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
243 pr_define("_PAGE_SPLITTING_SHIFT %d\n", _PAGE_SPLITTING_SHIFT);
245 #ifdef CONFIG_CPU_MIPSR2
247 #ifdef _PAGE_NO_EXEC_SHIFT
248 pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
249 pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
253 pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
254 pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
255 pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
256 pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
260 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
264 pr_debug("LEAF(%s)\n", symbol);
266 pr_debug("\t.set push\n");
267 pr_debug("\t.set noreorder\n");
269 for (i = 0; i < count; i++)
270 pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
272 pr_debug("\t.set\tpop\n");
274 pr_debug("\tEND(%s)\n", symbol);
277 /* The only general purpose registers allowed in TLB handlers. */
281 /* Some CP0 registers */
282 #define C0_INDEX 0, 0
283 #define C0_ENTRYLO0 2, 0
284 #define C0_TCBIND 2, 2
285 #define C0_ENTRYLO1 3, 0
286 #define C0_CONTEXT 4, 0
287 #define C0_PAGEMASK 5, 0
288 #define C0_BADVADDR 8, 0
289 #define C0_ENTRYHI 10, 0
291 #define C0_XCONTEXT 20, 0
294 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
296 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
299 /* The worst case length of the handler is around 18 instructions for
300 * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
301 * Maximum space available is 32 instructions for R3000 and 64
302 * instructions for R4000.
304 * We deliberately chose a buffer size of 128, so we won't scribble
305 * over anything important on overflow before we panic.
307 static u32 tlb_handler[128];
309 /* simply assume worst case size for labels and relocs */
310 static struct uasm_label labels[128];
311 static struct uasm_reloc relocs[128];
313 static int check_for_high_segbits;
315 static unsigned int kscratch_used_mask;
317 static inline int __maybe_unused c0_kscratch(void)
319 switch (current_cpu_type()) {
328 static int allocate_kscratch(void)
331 unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
338 r--; /* make it zero based */
340 kscratch_used_mask |= (1 << r);
345 static int scratch_reg;
347 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
349 static struct work_registers build_get_work_registers(u32 **p)
351 struct work_registers r;
353 if (scratch_reg >= 0) {
354 /* Save in CPU local C0_KScratch? */
355 UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
362 if (num_possible_cpus() > 1) {
363 /* Get smp_processor_id */
364 UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
365 UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
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 build_restore_work_registers(u32 **p)
387 if (scratch_reg >= 0) {
388 UASM_i_MFC0(p, 1, c0_kscratch(), 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 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);
443 local_flush_icache_range(ebase, ebase + 0x80);
445 dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
447 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
450 * The R4000 TLB handler is much more complicated. We have two
451 * consecutive handler areas with 32 instructions space each.
452 * Since they aren't used at the same time, we can overflow in the
453 * other one.To keep things simple, we first assume linear space,
454 * then we relocate it to the final handler layout as needed.
456 static u32 final_handler[64];
461 * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
462 * 2. A timing hazard exists for the TLBP instruction.
464 * stalling_instruction
467 * The JTLB is being read for the TLBP throughout the stall generated by the
468 * previous instruction. This is not really correct as the stalling instruction
469 * can modify the address used to access the JTLB. The failure symptom is that
470 * the TLBP instruction will use an address created for the stalling instruction
471 * and not the address held in C0_ENHI and thus report the wrong results.
473 * The software work-around is to not allow the instruction preceding the TLBP
474 * to stall - make it an NOP or some other instruction guaranteed not to stall.
476 * Errata 2 will not be fixed. This errata is also on the R5000.
478 * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
480 static void __maybe_unused build_tlb_probe_entry(u32 **p)
482 switch (current_cpu_type()) {
483 /* Found by experiment: R4600 v2.0/R4700 needs this, too. */
499 * Write random or indexed TLB entry, and care about the hazards from
500 * the preceding mtc0 and for the following eret.
502 enum tlb_write_entry { tlb_random, tlb_indexed };
504 static void build_tlb_write_entry(u32 **p, struct uasm_label **l,
505 struct uasm_reloc **r,
506 enum tlb_write_entry wmode)
508 void(*tlbw)(u32 **) = NULL;
511 case tlb_random: tlbw = uasm_i_tlbwr; break;
512 case tlb_indexed: tlbw = uasm_i_tlbwi; break;
515 if (cpu_has_mips_r2_exec_hazard) {
517 * The architecture spec says an ehb is required here,
518 * but a number of cores do not have the hazard and
519 * using an ehb causes an expensive pipeline stall.
521 switch (current_cpu_type()) {
528 case CPU_QEMU_GENERIC:
539 switch (current_cpu_type()) {
547 * This branch uses up a mtc0 hazard nop slot and saves
548 * two nops after the tlbw instruction.
550 uasm_bgezl_hazard(p, r, hazard_instance);
552 uasm_bgezl_label(l, p, hazard_instance);
566 uasm_i_nop(p); /* QED specifies 2 nops hazard */
567 uasm_i_nop(p); /* QED specifies 2 nops hazard */
640 panic("No TLB refill handler yet (CPU type: %d)",
646 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
650 UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
652 #ifdef CONFIG_PHYS_ADDR_T_64BIT
653 uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
655 UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
660 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
662 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
663 unsigned int tmp, enum label_id lid,
666 if (restore_scratch) {
667 /* Reset default page size */
668 if (PM_DEFAULT_MASK >> 16) {
669 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
670 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
671 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
672 uasm_il_b(p, r, lid);
673 } else if (PM_DEFAULT_MASK) {
674 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
675 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
676 uasm_il_b(p, r, lid);
678 uasm_i_mtc0(p, 0, C0_PAGEMASK);
679 uasm_il_b(p, r, lid);
681 if (scratch_reg >= 0)
682 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
684 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
686 /* Reset default page size */
687 if (PM_DEFAULT_MASK >> 16) {
688 uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
689 uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
690 uasm_il_b(p, r, lid);
691 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
692 } else if (PM_DEFAULT_MASK) {
693 uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
694 uasm_il_b(p, r, lid);
695 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
697 uasm_il_b(p, r, lid);
698 uasm_i_mtc0(p, 0, C0_PAGEMASK);
703 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
704 struct uasm_reloc **r,
706 enum tlb_write_entry wmode,
709 /* Set huge page tlb entry size */
710 uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
711 uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
712 uasm_i_mtc0(p, tmp, C0_PAGEMASK);
714 build_tlb_write_entry(p, l, r, wmode);
716 build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
720 * Check if Huge PTE is present, if so then jump to LABEL.
723 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
724 unsigned int pmd, int lid)
726 UASM_i_LW(p, tmp, 0, pmd);
727 if (use_bbit_insns()) {
728 uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
730 uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
731 uasm_il_bnez(p, r, tmp, lid);
735 static void build_huge_update_entries(u32 **p, unsigned int pte,
741 * A huge PTE describes an area the size of the
742 * configured huge page size. This is twice the
743 * of the large TLB entry size we intend to use.
744 * A TLB entry half the size of the configured
745 * huge page size is configured into entrylo0
746 * and entrylo1 to cover the contiguous huge PTE
749 small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
751 /* We can clobber tmp. It isn't used after this.*/
753 uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
755 build_convert_pte_to_entrylo(p, pte);
756 UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
757 /* convert to entrylo1 */
759 UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
761 UASM_i_ADDU(p, pte, pte, tmp);
763 UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
766 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
767 struct uasm_label **l,
772 UASM_i_SC(p, pte, 0, ptr);
773 uasm_il_beqz(p, r, pte, label_tlb_huge_update);
774 UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
776 UASM_i_SW(p, pte, 0, ptr);
778 build_huge_update_entries(p, pte, ptr);
779 build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
781 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
785 * TMP and PTR are scratch.
786 * TMP will be clobbered, PTR will hold the pmd entry.
789 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
790 unsigned int tmp, unsigned int ptr)
792 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
793 long pgdc = (long)pgd_current;
796 * The vmalloc handling is not in the hotpath.
798 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
800 if (check_for_high_segbits) {
802 * The kernel currently implicitely assumes that the
803 * MIPS SEGBITS parameter for the processor is
804 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
805 * allocate virtual addresses outside the maximum
806 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
807 * that doesn't prevent user code from accessing the
808 * higher xuseg addresses. Here, we make sure that
809 * everything but the lower xuseg addresses goes down
810 * the module_alloc/vmalloc path.
812 uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
813 uasm_il_bnez(p, r, ptr, label_vmalloc);
815 uasm_il_bltz(p, r, tmp, label_vmalloc);
817 /* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
820 /* pgd is in pgd_reg */
821 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
823 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
825 * &pgd << 11 stored in CONTEXT [23..63].
827 UASM_i_MFC0(p, ptr, C0_CONTEXT);
829 /* Clear lower 23 bits of context. */
830 uasm_i_dins(p, ptr, 0, 0, 23);
832 /* 1 0 1 0 1 << 6 xkphys cached */
833 uasm_i_ori(p, ptr, ptr, 0x540);
834 uasm_i_drotr(p, ptr, ptr, 11);
835 #elif defined(CONFIG_SMP)
836 UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
837 uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
838 UASM_i_LA_mostly(p, tmp, pgdc);
839 uasm_i_daddu(p, ptr, ptr, tmp);
840 uasm_i_dmfc0(p, tmp, C0_BADVADDR);
841 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
843 UASM_i_LA_mostly(p, ptr, pgdc);
844 uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
848 uasm_l_vmalloc_done(l, *p);
850 /* get pgd offset in bytes */
851 uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
853 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
854 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
855 #ifndef __PAGETABLE_PMD_FOLDED
856 uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
857 uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
858 uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
859 uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
860 uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
865 * BVADDR is the faulting address, PTR is scratch.
866 * PTR will hold the pgd for vmalloc.
869 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
870 unsigned int bvaddr, unsigned int ptr,
871 enum vmalloc64_mode mode)
873 long swpd = (long)swapper_pg_dir;
874 int single_insn_swpd;
875 int did_vmalloc_branch = 0;
877 single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
879 uasm_l_vmalloc(l, *p);
881 if (mode != not_refill && check_for_high_segbits) {
882 if (single_insn_swpd) {
883 uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
884 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
885 did_vmalloc_branch = 1;
888 uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
891 if (!did_vmalloc_branch) {
892 if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
893 uasm_il_b(p, r, label_vmalloc_done);
894 uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
896 UASM_i_LA_mostly(p, ptr, swpd);
897 uasm_il_b(p, r, label_vmalloc_done);
898 if (uasm_in_compat_space_p(swpd))
899 uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
901 uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
904 if (mode != not_refill && check_for_high_segbits) {
905 uasm_l_large_segbits_fault(l, *p);
907 * We get here if we are an xsseg address, or if we are
908 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
910 * Ignoring xsseg (assume disabled so would generate
911 * (address errors?), the only remaining possibility
912 * is the upper xuseg addresses. On processors with
913 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
914 * addresses would have taken an address error. We try
915 * to mimic that here by taking a load/istream page
918 UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
921 if (mode == refill_scratch) {
922 if (scratch_reg >= 0)
923 UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
925 UASM_i_LW(p, 1, scratchpad_offset(0), 0);
932 #else /* !CONFIG_64BIT */
935 * TMP and PTR are scratch.
936 * TMP will be clobbered, PTR will hold the pgd entry.
938 static void __maybe_unused
939 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
942 /* pgd is in pgd_reg */
943 uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
944 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
946 long pgdc = (long)pgd_current;
948 /* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
950 uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
951 UASM_i_LA_mostly(p, tmp, pgdc);
952 uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
953 uasm_i_addu(p, ptr, tmp, ptr);
955 UASM_i_LA_mostly(p, ptr, pgdc);
957 uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
958 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
960 uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
961 uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
962 uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
965 #endif /* !CONFIG_64BIT */
967 static void build_adjust_context(u32 **p, unsigned int ctx)
969 unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
970 unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
972 switch (current_cpu_type()) {
989 UASM_i_SRL(p, ctx, ctx, shift);
990 uasm_i_andi(p, ctx, ctx, mask);
993 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
996 * Bug workaround for the Nevada. It seems as if under certain
997 * circumstances the move from cp0_context might produce a
998 * bogus result when the mfc0 instruction and its consumer are
999 * in a different cacheline or a load instruction, probably any
1000 * memory reference, is between them.
1002 switch (current_cpu_type()) {
1004 UASM_i_LW(p, ptr, 0, ptr);
1005 GET_CONTEXT(p, tmp); /* get context reg */
1009 GET_CONTEXT(p, tmp); /* get context reg */
1010 UASM_i_LW(p, ptr, 0, ptr);
1014 build_adjust_context(p, tmp);
1015 UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1018 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1021 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1022 * Kernel is a special case. Only a few CPUs use it.
1024 #ifdef CONFIG_PHYS_ADDR_T_64BIT
1025 if (cpu_has_64bits) {
1026 uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1027 uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1029 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1030 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1031 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1033 uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1034 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1035 uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1037 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1039 int pte_off_even = sizeof(pte_t) / 2;
1040 int pte_off_odd = pte_off_even + sizeof(pte_t);
1042 const int scratch = 1; /* Our extra working register */
1044 uasm_i_addu(p, scratch, 0, ptep);
1046 uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1047 uasm_i_lw(p, ptep, pte_off_odd, ptep); /* odd pte */
1048 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1049 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1050 UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1051 UASM_i_MTC0(p, ptep, C0_ENTRYLO1);
1053 uasm_i_lw(p, tmp, 0, scratch);
1054 uasm_i_lw(p, ptep, sizeof(pte_t), scratch);
1055 uasm_i_lui(p, scratch, 0xff);
1056 uasm_i_ori(p, scratch, scratch, 0xffff);
1057 uasm_i_and(p, tmp, scratch, tmp);
1058 uasm_i_and(p, ptep, scratch, ptep);
1059 uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1060 uasm_i_mthc0(p, ptep, C0_ENTRYLO1);
1064 UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1065 UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1066 if (r45k_bvahwbug())
1067 build_tlb_probe_entry(p);
1069 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1070 if (r4k_250MHZhwbug())
1071 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1072 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1073 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1075 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1076 if (r4k_250MHZhwbug())
1077 UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1078 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1079 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1080 if (r45k_bvahwbug())
1081 uasm_i_mfc0(p, tmp, C0_INDEX);
1083 if (r4k_250MHZhwbug())
1084 UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1085 UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1089 struct mips_huge_tlb_info {
1091 int restore_scratch;
1092 bool need_reload_pte;
1095 static struct mips_huge_tlb_info
1096 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1097 struct uasm_reloc **r, unsigned int tmp,
1098 unsigned int ptr, int c0_scratch_reg)
1100 struct mips_huge_tlb_info rv;
1101 unsigned int even, odd;
1102 int vmalloc_branch_delay_filled = 0;
1103 const int scratch = 1; /* Our extra working register */
1105 rv.huge_pte = scratch;
1106 rv.restore_scratch = 0;
1107 rv.need_reload_pte = false;
1109 if (check_for_high_segbits) {
1110 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1113 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1115 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1117 if (c0_scratch_reg >= 0)
1118 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1120 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1122 uasm_i_dsrl_safe(p, scratch, tmp,
1123 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1124 uasm_il_bnez(p, r, scratch, label_vmalloc);
1126 if (pgd_reg == -1) {
1127 vmalloc_branch_delay_filled = 1;
1128 /* Clear lower 23 bits of context. */
1129 uasm_i_dins(p, ptr, 0, 0, 23);
1133 UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1135 UASM_i_MFC0(p, ptr, C0_CONTEXT);
1137 UASM_i_MFC0(p, tmp, C0_BADVADDR);
1139 if (c0_scratch_reg >= 0)
1140 UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1142 UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1145 /* Clear lower 23 bits of context. */
1146 uasm_i_dins(p, ptr, 0, 0, 23);
1148 uasm_il_bltz(p, r, tmp, label_vmalloc);
1151 if (pgd_reg == -1) {
1152 vmalloc_branch_delay_filled = 1;
1153 /* 1 0 1 0 1 << 6 xkphys cached */
1154 uasm_i_ori(p, ptr, ptr, 0x540);
1155 uasm_i_drotr(p, ptr, ptr, 11);
1158 #ifdef __PAGETABLE_PMD_FOLDED
1159 #define LOC_PTEP scratch
1161 #define LOC_PTEP ptr
1164 if (!vmalloc_branch_delay_filled)
1165 /* get pgd offset in bytes */
1166 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1168 uasm_l_vmalloc_done(l, *p);
1172 * fall-through case = badvaddr *pgd_current
1173 * vmalloc case = badvaddr swapper_pg_dir
1176 if (vmalloc_branch_delay_filled)
1177 /* get pgd offset in bytes */
1178 uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1180 #ifdef __PAGETABLE_PMD_FOLDED
1181 GET_CONTEXT(p, tmp); /* get context reg */
1183 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1185 if (use_lwx_insns()) {
1186 UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1188 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1189 uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1192 #ifndef __PAGETABLE_PMD_FOLDED
1193 /* get pmd offset in bytes */
1194 uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1195 uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1196 GET_CONTEXT(p, tmp); /* get context reg */
1198 if (use_lwx_insns()) {
1199 UASM_i_LWX(p, scratch, scratch, ptr);
1201 uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1202 UASM_i_LW(p, scratch, 0, ptr);
1205 /* Adjust the context during the load latency. */
1206 build_adjust_context(p, tmp);
1208 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1209 uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1211 * The in the LWX case we don't want to do the load in the
1212 * delay slot. It cannot issue in the same cycle and may be
1213 * speculative and unneeded.
1215 if (use_lwx_insns())
1217 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1220 /* build_update_entries */
1221 if (use_lwx_insns()) {
1224 UASM_i_LWX(p, even, scratch, tmp);
1225 UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1226 UASM_i_LWX(p, odd, scratch, tmp);
1228 UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1231 UASM_i_LW(p, even, 0, ptr); /* get even pte */
1232 UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1235 uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1236 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1237 uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1239 uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1240 UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1241 uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1243 UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1245 if (c0_scratch_reg >= 0) {
1246 UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1247 build_tlb_write_entry(p, l, r, tlb_random);
1248 uasm_l_leave(l, *p);
1249 rv.restore_scratch = 1;
1250 } else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13) {
1251 build_tlb_write_entry(p, l, r, tlb_random);
1252 uasm_l_leave(l, *p);
1253 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1255 UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1256 build_tlb_write_entry(p, l, r, tlb_random);
1257 uasm_l_leave(l, *p);
1258 rv.restore_scratch = 1;
1261 uasm_i_eret(p); /* return from trap */
1267 * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1268 * because EXL == 0. If we wrap, we can also use the 32 instruction
1269 * slots before the XTLB refill exception handler which belong to the
1270 * unused TLB refill exception.
1272 #define MIPS64_REFILL_INSNS 32
1274 static void build_r4000_tlb_refill_handler(void)
1276 u32 *p = tlb_handler;
1277 struct uasm_label *l = labels;
1278 struct uasm_reloc *r = relocs;
1280 unsigned int final_len;
1281 struct mips_huge_tlb_info htlb_info __maybe_unused;
1282 enum vmalloc64_mode vmalloc_mode __maybe_unused;
1284 memset(tlb_handler, 0, sizeof(tlb_handler));
1285 memset(labels, 0, sizeof(labels));
1286 memset(relocs, 0, sizeof(relocs));
1287 memset(final_handler, 0, sizeof(final_handler));
1289 if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1290 htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1292 vmalloc_mode = refill_scratch;
1294 htlb_info.huge_pte = K0;
1295 htlb_info.restore_scratch = 0;
1296 htlb_info.need_reload_pte = true;
1297 vmalloc_mode = refill_noscratch;
1299 * create the plain linear handler
1301 if (bcm1250_m3_war()) {
1302 unsigned int segbits = 44;
1304 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1305 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1306 uasm_i_xor(&p, K0, K0, K1);
1307 uasm_i_dsrl_safe(&p, K1, K0, 62);
1308 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1309 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1310 uasm_i_or(&p, K0, K0, K1);
1311 uasm_il_bnez(&p, &r, K0, label_leave);
1312 /* No need for uasm_i_nop */
1316 build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1318 build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1321 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1322 build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1325 build_get_ptep(&p, K0, K1);
1326 build_update_entries(&p, K0, K1);
1327 build_tlb_write_entry(&p, &l, &r, tlb_random);
1328 uasm_l_leave(&l, p);
1329 uasm_i_eret(&p); /* return from trap */
1331 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1332 uasm_l_tlb_huge_update(&l, p);
1333 if (htlb_info.need_reload_pte)
1334 UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1335 build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1336 build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1337 htlb_info.restore_scratch);
1341 build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1345 * Overflow check: For the 64bit handler, we need at least one
1346 * free instruction slot for the wrap-around branch. In worst
1347 * case, if the intended insertion point is a delay slot, we
1348 * need three, with the second nop'ed and the third being
1351 switch (boot_cpu_type()) {
1353 if (sizeof(long) == 4) {
1355 /* Loongson2 ebase is different than r4k, we have more space */
1356 if ((p - tlb_handler) > 64)
1357 panic("TLB refill handler space exceeded");
1359 * Now fold the handler in the TLB refill handler space.
1362 /* Simplest case, just copy the handler. */
1363 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1364 final_len = p - tlb_handler;
1367 if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1368 || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1369 && uasm_insn_has_bdelay(relocs,
1370 tlb_handler + MIPS64_REFILL_INSNS - 3)))
1371 panic("TLB refill handler space exceeded");
1373 * Now fold the handler in the TLB refill handler space.
1375 f = final_handler + MIPS64_REFILL_INSNS;
1376 if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1377 /* Just copy the handler. */
1378 uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1379 final_len = p - tlb_handler;
1381 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1382 const enum label_id ls = label_tlb_huge_update;
1384 const enum label_id ls = label_vmalloc;
1390 for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1392 BUG_ON(i == ARRAY_SIZE(labels));
1393 split = labels[i].addr;
1396 * See if we have overflown one way or the other.
1398 if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1399 split < p - MIPS64_REFILL_INSNS)
1404 * Split two instructions before the end. One
1405 * for the branch and one for the instruction
1406 * in the delay slot.
1408 split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1411 * If the branch would fall in a delay slot,
1412 * we must back up an additional instruction
1413 * so that it is no longer in a delay slot.
1415 if (uasm_insn_has_bdelay(relocs, split - 1))
1418 /* Copy first part of the handler. */
1419 uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1420 f += split - tlb_handler;
1423 /* Insert branch. */
1424 uasm_l_split(&l, final_handler);
1425 uasm_il_b(&f, &r, label_split);
1426 if (uasm_insn_has_bdelay(relocs, split))
1429 uasm_copy_handler(relocs, labels,
1430 split, split + 1, f);
1431 uasm_move_labels(labels, f, f + 1, -1);
1437 /* Copy the rest of the handler. */
1438 uasm_copy_handler(relocs, labels, split, p, final_handler);
1439 final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1446 uasm_resolve_relocs(relocs, labels);
1447 pr_debug("Wrote TLB refill handler (%u instructions).\n",
1450 memcpy((void *)ebase, final_handler, 0x100);
1451 local_flush_icache_range(ebase, ebase + 0x100);
1453 dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1456 extern u32 handle_tlbl[], handle_tlbl_end[];
1457 extern u32 handle_tlbs[], handle_tlbs_end[];
1458 extern u32 handle_tlbm[], handle_tlbm_end[];
1459 extern u32 tlbmiss_handler_setup_pgd_start[], tlbmiss_handler_setup_pgd[];
1460 extern u32 tlbmiss_handler_setup_pgd_end[];
1462 static void build_setup_pgd(void)
1465 const int __maybe_unused a1 = 5;
1466 const int __maybe_unused a2 = 6;
1467 u32 *p = tlbmiss_handler_setup_pgd_start;
1468 const int tlbmiss_handler_setup_pgd_size =
1469 tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1470 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1471 long pgdc = (long)pgd_current;
1474 memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1475 sizeof(tlbmiss_handler_setup_pgd[0]));
1476 memset(labels, 0, sizeof(labels));
1477 memset(relocs, 0, sizeof(relocs));
1478 pgd_reg = allocate_kscratch();
1479 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1480 if (pgd_reg == -1) {
1481 struct uasm_label *l = labels;
1482 struct uasm_reloc *r = relocs;
1484 /* PGD << 11 in c0_Context */
1486 * If it is a ckseg0 address, convert to a physical
1487 * address. Shifting right by 29 and adding 4 will
1488 * result in zero for these addresses.
1491 UASM_i_SRA(&p, a1, a0, 29);
1492 UASM_i_ADDIU(&p, a1, a1, 4);
1493 uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1495 uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1496 uasm_l_tlbl_goaround1(&l, p);
1497 UASM_i_SLL(&p, a0, a0, 11);
1499 UASM_i_MTC0(&p, a0, C0_CONTEXT);
1501 /* PGD in c0_KScratch */
1503 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1507 /* Save PGD to pgd_current[smp_processor_id()] */
1508 UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1509 UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1510 UASM_i_LA_mostly(&p, a2, pgdc);
1511 UASM_i_ADDU(&p, a2, a2, a1);
1512 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1514 UASM_i_LA_mostly(&p, a2, pgdc);
1515 UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1519 /* if pgd_reg is allocated, save PGD also to scratch register */
1521 UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1525 if (p >= tlbmiss_handler_setup_pgd_end)
1526 panic("tlbmiss_handler_setup_pgd space exceeded");
1528 uasm_resolve_relocs(relocs, labels);
1529 pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1530 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1532 dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1533 tlbmiss_handler_setup_pgd_size);
1537 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1540 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1542 uasm_i_lld(p, pte, 0, ptr);
1545 UASM_i_LL(p, pte, 0, ptr);
1547 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1549 uasm_i_ld(p, pte, 0, ptr);
1552 UASM_i_LW(p, pte, 0, ptr);
1557 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1560 #ifdef CONFIG_PHYS_ADDR_T_64BIT
1561 unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1563 if (!cpu_has_64bits) {
1564 const int scratch = 1; /* Our extra working register */
1566 uasm_i_lui(p, scratch, (mode >> 16));
1567 uasm_i_or(p, pte, pte, scratch);
1570 uasm_i_ori(p, pte, pte, mode);
1572 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1574 uasm_i_scd(p, pte, 0, ptr);
1577 UASM_i_SC(p, pte, 0, ptr);
1579 if (r10000_llsc_war())
1580 uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1582 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1584 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1585 if (!cpu_has_64bits) {
1586 /* no uasm_i_nop needed */
1587 uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1588 uasm_i_ori(p, pte, pte, hwmode);
1589 uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1590 uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1591 /* no uasm_i_nop needed */
1592 uasm_i_lw(p, pte, 0, ptr);
1599 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1601 uasm_i_sd(p, pte, 0, ptr);
1604 UASM_i_SW(p, pte, 0, ptr);
1606 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1607 if (!cpu_has_64bits) {
1608 uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1609 uasm_i_ori(p, pte, pte, hwmode);
1610 uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1611 uasm_i_lw(p, pte, 0, ptr);
1618 * Check if PTE is present, if not then jump to LABEL. PTR points to
1619 * the page table where this PTE is located, PTE will be re-loaded
1620 * with it's original value.
1623 build_pte_present(u32 **p, struct uasm_reloc **r,
1624 int pte, int ptr, int scratch, enum label_id lid)
1626 int t = scratch >= 0 ? scratch : pte;
1629 if (use_bbit_insns()) {
1630 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1633 uasm_i_srl(p, t, pte, _PAGE_PRESENT_SHIFT);
1634 uasm_i_andi(p, t, t, 1);
1635 uasm_il_beqz(p, r, t, lid);
1637 /* You lose the SMP race :-(*/
1638 iPTE_LW(p, pte, ptr);
1641 uasm_i_srl(p, t, pte, _PAGE_PRESENT_SHIFT);
1642 uasm_i_andi(p, t, t, 3);
1643 uasm_i_xori(p, t, t, 3);
1644 uasm_il_bnez(p, r, t, lid);
1646 /* You lose the SMP race :-(*/
1647 iPTE_LW(p, pte, ptr);
1651 /* Make PTE valid, store result in PTR. */
1653 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1656 unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1658 iPTE_SW(p, r, pte, ptr, mode);
1662 * Check if PTE can be written to, if not branch to LABEL. Regardless
1663 * restore PTE with value from PTR when done.
1666 build_pte_writable(u32 **p, struct uasm_reloc **r,
1667 unsigned int pte, unsigned int ptr, int scratch,
1670 int t = scratch >= 0 ? scratch : pte;
1672 uasm_i_srl(p, t, pte, _PAGE_PRESENT_SHIFT);
1673 uasm_i_andi(p, t, t, 5);
1674 uasm_i_xori(p, t, t, 5);
1675 uasm_il_bnez(p, r, t, lid);
1677 /* You lose the SMP race :-(*/
1678 iPTE_LW(p, pte, ptr);
1683 /* Make PTE writable, update software status bits as well, then store
1687 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1690 unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1693 iPTE_SW(p, r, pte, ptr, mode);
1697 * Check if PTE can be modified, if not branch to LABEL. Regardless
1698 * restore PTE with value from PTR when done.
1701 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1702 unsigned int pte, unsigned int ptr, int scratch,
1705 if (use_bbit_insns()) {
1706 uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1709 int t = scratch >= 0 ? scratch : pte;
1710 uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1711 uasm_i_andi(p, t, t, 1);
1712 uasm_il_beqz(p, r, t, lid);
1714 /* You lose the SMP race :-(*/
1715 iPTE_LW(p, pte, ptr);
1719 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1723 * R3000 style TLB load/store/modify handlers.
1727 * This places the pte into ENTRYLO0 and writes it with tlbwi.
1731 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1733 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1734 uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1737 uasm_i_rfe(p); /* branch delay */
1741 * This places the pte into ENTRYLO0 and writes it with tlbwi
1742 * or tlbwr as appropriate. This is because the index register
1743 * may have the probe fail bit set as a result of a trap on a
1744 * kseg2 access, i.e. without refill. Then it returns.
1747 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1748 struct uasm_reloc **r, unsigned int pte,
1751 uasm_i_mfc0(p, tmp, C0_INDEX);
1752 uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1753 uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1754 uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1755 uasm_i_tlbwi(p); /* cp0 delay */
1757 uasm_i_rfe(p); /* branch delay */
1758 uasm_l_r3000_write_probe_fail(l, *p);
1759 uasm_i_tlbwr(p); /* cp0 delay */
1761 uasm_i_rfe(p); /* branch delay */
1765 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1768 long pgdc = (long)pgd_current;
1770 uasm_i_mfc0(p, pte, C0_BADVADDR);
1771 uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1772 uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1773 uasm_i_srl(p, pte, pte, 22); /* load delay */
1774 uasm_i_sll(p, pte, pte, 2);
1775 uasm_i_addu(p, ptr, ptr, pte);
1776 uasm_i_mfc0(p, pte, C0_CONTEXT);
1777 uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1778 uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1779 uasm_i_addu(p, ptr, ptr, pte);
1780 uasm_i_lw(p, pte, 0, ptr);
1781 uasm_i_tlbp(p); /* load delay */
1784 static void build_r3000_tlb_load_handler(void)
1786 u32 *p = handle_tlbl;
1787 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1788 struct uasm_label *l = labels;
1789 struct uasm_reloc *r = relocs;
1791 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1792 memset(labels, 0, sizeof(labels));
1793 memset(relocs, 0, sizeof(relocs));
1795 build_r3000_tlbchange_handler_head(&p, K0, K1);
1796 build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1797 uasm_i_nop(&p); /* load delay */
1798 build_make_valid(&p, &r, K0, K1);
1799 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1801 uasm_l_nopage_tlbl(&l, p);
1802 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1805 if (p >= handle_tlbl_end)
1806 panic("TLB load handler fastpath space exceeded");
1808 uasm_resolve_relocs(relocs, labels);
1809 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1810 (unsigned int)(p - handle_tlbl));
1812 dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1815 static void build_r3000_tlb_store_handler(void)
1817 u32 *p = handle_tlbs;
1818 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1819 struct uasm_label *l = labels;
1820 struct uasm_reloc *r = relocs;
1822 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1823 memset(labels, 0, sizeof(labels));
1824 memset(relocs, 0, sizeof(relocs));
1826 build_r3000_tlbchange_handler_head(&p, K0, K1);
1827 build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1828 uasm_i_nop(&p); /* load delay */
1829 build_make_write(&p, &r, K0, K1);
1830 build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1832 uasm_l_nopage_tlbs(&l, p);
1833 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1836 if (p >= handle_tlbs_end)
1837 panic("TLB store handler fastpath space exceeded");
1839 uasm_resolve_relocs(relocs, labels);
1840 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1841 (unsigned int)(p - handle_tlbs));
1843 dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1846 static void build_r3000_tlb_modify_handler(void)
1848 u32 *p = handle_tlbm;
1849 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1850 struct uasm_label *l = labels;
1851 struct uasm_reloc *r = relocs;
1853 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1854 memset(labels, 0, sizeof(labels));
1855 memset(relocs, 0, sizeof(relocs));
1857 build_r3000_tlbchange_handler_head(&p, K0, K1);
1858 build_pte_modifiable(&p, &r, K0, K1, -1, label_nopage_tlbm);
1859 uasm_i_nop(&p); /* load delay */
1860 build_make_write(&p, &r, K0, K1);
1861 build_r3000_pte_reload_tlbwi(&p, K0, K1);
1863 uasm_l_nopage_tlbm(&l, p);
1864 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1867 if (p >= handle_tlbm_end)
1868 panic("TLB modify handler fastpath space exceeded");
1870 uasm_resolve_relocs(relocs, labels);
1871 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1872 (unsigned int)(p - handle_tlbm));
1874 dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1876 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1879 * R4000 style TLB load/store/modify handlers.
1881 static struct work_registers
1882 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1883 struct uasm_reloc **r)
1885 struct work_registers wr = build_get_work_registers(p);
1888 build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1890 build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1893 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1895 * For huge tlb entries, pmd doesn't contain an address but
1896 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1897 * see if we need to jump to huge tlb processing.
1899 build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1902 UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1903 UASM_i_LW(p, wr.r2, 0, wr.r2);
1904 UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1905 uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1906 UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1909 uasm_l_smp_pgtable_change(l, *p);
1911 iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1912 if (!m4kc_tlbp_war()) {
1913 build_tlb_probe_entry(p);
1915 /* race condition happens, leaving */
1917 uasm_i_mfc0(p, wr.r3, C0_INDEX);
1918 uasm_il_bltz(p, r, wr.r3, label_leave);
1926 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1927 struct uasm_reloc **r, unsigned int tmp,
1930 uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1931 uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1932 build_update_entries(p, tmp, ptr);
1933 build_tlb_write_entry(p, l, r, tlb_indexed);
1934 uasm_l_leave(l, *p);
1935 build_restore_work_registers(p);
1936 uasm_i_eret(p); /* return from trap */
1939 build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1943 static void build_r4000_tlb_load_handler(void)
1945 u32 *p = handle_tlbl;
1946 const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1947 struct uasm_label *l = labels;
1948 struct uasm_reloc *r = relocs;
1949 struct work_registers wr;
1951 memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1952 memset(labels, 0, sizeof(labels));
1953 memset(relocs, 0, sizeof(relocs));
1955 if (bcm1250_m3_war()) {
1956 unsigned int segbits = 44;
1958 uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1959 uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1960 uasm_i_xor(&p, K0, K0, K1);
1961 uasm_i_dsrl_safe(&p, K1, K0, 62);
1962 uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1963 uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1964 uasm_i_or(&p, K0, K0, K1);
1965 uasm_il_bnez(&p, &r, K0, label_leave);
1966 /* No need for uasm_i_nop */
1969 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1970 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1971 if (m4kc_tlbp_war())
1972 build_tlb_probe_entry(&p);
1974 if (cpu_has_rixi && !cpu_has_rixiex) {
1976 * If the page is not _PAGE_VALID, RI or XI could not
1977 * have triggered it. Skip the expensive test..
1979 if (use_bbit_insns()) {
1980 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1981 label_tlbl_goaround1);
1983 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1984 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1990 switch (current_cpu_type()) {
1992 if (cpu_has_mips_r2_exec_hazard) {
1995 case CPU_CAVIUM_OCTEON:
1996 case CPU_CAVIUM_OCTEON_PLUS:
1997 case CPU_CAVIUM_OCTEON2:
2002 /* Examine entrylo 0 or 1 based on ptr. */
2003 if (use_bbit_insns()) {
2004 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2006 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2007 uasm_i_beqz(&p, wr.r3, 8);
2009 /* load it in the delay slot*/
2010 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2011 /* load it if ptr is odd */
2012 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2014 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2015 * XI must have triggered it.
2017 if (use_bbit_insns()) {
2018 uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2020 uasm_l_tlbl_goaround1(&l, p);
2022 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2023 uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2026 uasm_l_tlbl_goaround1(&l, p);
2028 build_make_valid(&p, &r, wr.r1, wr.r2);
2029 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2031 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2033 * This is the entry point when build_r4000_tlbchange_handler_head
2034 * spots a huge page.
2036 uasm_l_tlb_huge_update(&l, p);
2037 iPTE_LW(&p, wr.r1, wr.r2);
2038 build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2039 build_tlb_probe_entry(&p);
2041 if (cpu_has_rixi && !cpu_has_rixiex) {
2043 * If the page is not _PAGE_VALID, RI or XI could not
2044 * have triggered it. Skip the expensive test..
2046 if (use_bbit_insns()) {
2047 uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2048 label_tlbl_goaround2);
2050 uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2051 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2057 switch (current_cpu_type()) {
2059 if (cpu_has_mips_r2_exec_hazard) {
2062 case CPU_CAVIUM_OCTEON:
2063 case CPU_CAVIUM_OCTEON_PLUS:
2064 case CPU_CAVIUM_OCTEON2:
2069 /* Examine entrylo 0 or 1 based on ptr. */
2070 if (use_bbit_insns()) {
2071 uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2073 uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2074 uasm_i_beqz(&p, wr.r3, 8);
2076 /* load it in the delay slot*/
2077 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2078 /* load it if ptr is odd */
2079 UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2081 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2082 * XI must have triggered it.
2084 if (use_bbit_insns()) {
2085 uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2087 uasm_i_andi(&p, wr.r3, wr.r3, 2);
2088 uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2090 if (PM_DEFAULT_MASK == 0)
2093 * We clobbered C0_PAGEMASK, restore it. On the other branch
2094 * it is restored in build_huge_tlb_write_entry.
2096 build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2098 uasm_l_tlbl_goaround2(&l, p);
2100 uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2101 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2104 uasm_l_nopage_tlbl(&l, p);
2105 build_restore_work_registers(&p);
2106 #ifdef CONFIG_CPU_MICROMIPS
2107 if ((unsigned long)tlb_do_page_fault_0 & 1) {
2108 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2109 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2113 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2116 if (p >= handle_tlbl_end)
2117 panic("TLB load handler fastpath space exceeded");
2119 uasm_resolve_relocs(relocs, labels);
2120 pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2121 (unsigned int)(p - handle_tlbl));
2123 dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2126 static void build_r4000_tlb_store_handler(void)
2128 u32 *p = handle_tlbs;
2129 const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2130 struct uasm_label *l = labels;
2131 struct uasm_reloc *r = relocs;
2132 struct work_registers wr;
2134 memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2135 memset(labels, 0, sizeof(labels));
2136 memset(relocs, 0, sizeof(relocs));
2138 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2139 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2140 if (m4kc_tlbp_war())
2141 build_tlb_probe_entry(&p);
2142 build_make_write(&p, &r, wr.r1, wr.r2);
2143 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2145 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2147 * This is the entry point when
2148 * build_r4000_tlbchange_handler_head spots a huge page.
2150 uasm_l_tlb_huge_update(&l, p);
2151 iPTE_LW(&p, wr.r1, wr.r2);
2152 build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2153 build_tlb_probe_entry(&p);
2154 uasm_i_ori(&p, wr.r1, wr.r1,
2155 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2156 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2159 uasm_l_nopage_tlbs(&l, p);
2160 build_restore_work_registers(&p);
2161 #ifdef CONFIG_CPU_MICROMIPS
2162 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2163 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2164 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2168 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2171 if (p >= handle_tlbs_end)
2172 panic("TLB store handler fastpath space exceeded");
2174 uasm_resolve_relocs(relocs, labels);
2175 pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2176 (unsigned int)(p - handle_tlbs));
2178 dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2181 static void build_r4000_tlb_modify_handler(void)
2183 u32 *p = handle_tlbm;
2184 const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2185 struct uasm_label *l = labels;
2186 struct uasm_reloc *r = relocs;
2187 struct work_registers wr;
2189 memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2190 memset(labels, 0, sizeof(labels));
2191 memset(relocs, 0, sizeof(relocs));
2193 wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2194 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2195 if (m4kc_tlbp_war())
2196 build_tlb_probe_entry(&p);
2197 /* Present and writable bits set, set accessed and dirty bits. */
2198 build_make_write(&p, &r, wr.r1, wr.r2);
2199 build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2201 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2203 * This is the entry point when
2204 * build_r4000_tlbchange_handler_head spots a huge page.
2206 uasm_l_tlb_huge_update(&l, p);
2207 iPTE_LW(&p, wr.r1, wr.r2);
2208 build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2209 build_tlb_probe_entry(&p);
2210 uasm_i_ori(&p, wr.r1, wr.r1,
2211 _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2212 build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2215 uasm_l_nopage_tlbm(&l, p);
2216 build_restore_work_registers(&p);
2217 #ifdef CONFIG_CPU_MICROMIPS
2218 if ((unsigned long)tlb_do_page_fault_1 & 1) {
2219 uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2220 uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2224 uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2227 if (p >= handle_tlbm_end)
2228 panic("TLB modify handler fastpath space exceeded");
2230 uasm_resolve_relocs(relocs, labels);
2231 pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2232 (unsigned int)(p - handle_tlbm));
2234 dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2237 static void flush_tlb_handlers(void)
2239 local_flush_icache_range((unsigned long)handle_tlbl,
2240 (unsigned long)handle_tlbl_end);
2241 local_flush_icache_range((unsigned long)handle_tlbs,
2242 (unsigned long)handle_tlbs_end);
2243 local_flush_icache_range((unsigned long)handle_tlbm,
2244 (unsigned long)handle_tlbm_end);
2245 local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2246 (unsigned long)tlbmiss_handler_setup_pgd_end);
2249 static void print_htw_config(void)
2251 unsigned long config;
2253 const int field = 2 * sizeof(unsigned long);
2255 config = read_c0_pwfield();
2256 pr_debug("PWField (0x%0*lx): GDI: 0x%02lx UDI: 0x%02lx MDI: 0x%02lx PTI: 0x%02lx PTEI: 0x%02lx\n",
2258 (config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2259 (config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2260 (config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2261 (config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2262 (config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2264 config = read_c0_pwsize();
2265 pr_debug("PWSize (0x%0*lx): GDW: 0x%02lx UDW: 0x%02lx MDW: 0x%02lx PTW: 0x%02lx PTEW: 0x%02lx\n",
2267 (config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2268 (config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2269 (config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2270 (config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2271 (config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2273 pwctl = read_c0_pwctl();
2274 pr_debug("PWCtl (0x%x): PWEn: 0x%x DPH: 0x%x HugePg: 0x%x Psn: 0x%x\n",
2276 (pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2277 (pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2278 (pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2279 (pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2282 static void config_htw_params(void)
2284 unsigned long pwfield, pwsize, ptei;
2285 unsigned int config;
2288 * We are using 2-level page tables, so we only need to
2289 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2290 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2291 * write values less than 0xc in these fields because the entire
2292 * write will be dropped. As a result of which, we must preserve
2293 * the original reset values and overwrite only what we really want.
2296 pwfield = read_c0_pwfield();
2297 /* re-initialize the GDI field */
2298 pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2299 pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2300 /* re-initialize the PTI field including the even/odd bit */
2301 pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2302 pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2303 /* Set the PTEI right shift */
2304 ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2306 write_c0_pwfield(pwfield);
2307 /* Check whether the PTEI value is supported */
2308 back_to_back_c0_hazard();
2309 pwfield = read_c0_pwfield();
2310 if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2312 pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2315 * Drop option to avoid HTW being enabled via another path
2318 current_cpu_data.options &= ~MIPS_CPU_HTW;
2322 pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2323 pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2325 /* If XPA has been enabled, PTEs are 64-bit in size. */
2326 if (read_c0_pagegrain() & PG_ELPA)
2329 write_c0_pwsize(pwsize);
2331 /* Make sure everything is set before we enable the HTW */
2332 back_to_back_c0_hazard();
2334 /* Enable HTW and disable the rest of the pwctl fields */
2335 config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2336 write_c0_pwctl(config);
2337 pr_info("Hardware Page Table Walker enabled\n");
2342 static void config_xpa_params(void)
2345 unsigned int pagegrain;
2347 if (mips_xpa_disabled) {
2348 pr_info("Extended Physical Addressing (XPA) disabled\n");
2352 pagegrain = read_c0_pagegrain();
2353 write_c0_pagegrain(pagegrain | PG_ELPA);
2354 back_to_back_c0_hazard();
2355 pagegrain = read_c0_pagegrain();
2357 if (pagegrain & PG_ELPA)
2358 pr_info("Extended Physical Addressing (XPA) enabled\n");
2360 panic("Extended Physical Addressing (XPA) disabled");
2364 void build_tlb_refill_handler(void)
2367 * The refill handler is generated per-CPU, multi-node systems
2368 * may have local storage for it. The other handlers are only
2371 static int run_once = 0;
2373 output_pgtable_bits_defines();
2376 check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2379 switch (current_cpu_type()) {
2387 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2388 if (cpu_has_local_ebase)
2389 build_r3000_tlb_refill_handler();
2391 if (!cpu_has_local_ebase)
2392 build_r3000_tlb_refill_handler();
2394 build_r3000_tlb_load_handler();
2395 build_r3000_tlb_store_handler();
2396 build_r3000_tlb_modify_handler();
2397 flush_tlb_handlers();
2401 panic("No R3000 TLB refill handler");
2407 panic("No R6000 TLB refill handler yet");
2411 panic("No R8000 TLB refill handler yet");
2416 scratch_reg = allocate_kscratch();
2418 build_r4000_tlb_load_handler();
2419 build_r4000_tlb_store_handler();
2420 build_r4000_tlb_modify_handler();
2421 if (!cpu_has_local_ebase)
2422 build_r4000_tlb_refill_handler();
2423 flush_tlb_handlers();
2426 if (cpu_has_local_ebase)
2427 build_r4000_tlb_refill_handler();
2429 config_xpa_params();
2431 config_htw_params();