xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision 3805e6a1)
1 /*
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
4  * for more details.
5  *
6  * Synthesize TLB refill handlers at runtime.
7  *
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.
13  *
14  * ... and the days got worse and worse and now you see
15  * I've gone completely out of my mind.
16  *
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 ...
20  *
21  * (Condolences to Napoleon XIV)
22  */
23 
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>
30 
31 #include <asm/cacheflush.h>
32 #include <asm/cpu-type.h>
33 #include <asm/pgtable.h>
34 #include <asm/war.h>
35 #include <asm/uasm.h>
36 #include <asm/setup.h>
37 
38 static int mips_xpa_disabled;
39 
40 static int __init xpa_disable(char *s)
41 {
42 	mips_xpa_disabled = 1;
43 
44 	return 1;
45 }
46 
47 __setup("noxpa", xpa_disable);
48 
49 /*
50  * TLB load/store/modify handlers.
51  *
52  * Only the fastpath gets synthesized at runtime, the slowpath for
53  * do_page_fault remains normal asm.
54  */
55 extern void tlb_do_page_fault_0(void);
56 extern void tlb_do_page_fault_1(void);
57 
58 struct work_registers {
59 	int r1;
60 	int r2;
61 	int r3;
62 };
63 
64 struct tlb_reg_save {
65 	unsigned long a;
66 	unsigned long b;
67 } ____cacheline_aligned_in_smp;
68 
69 static struct tlb_reg_save handler_reg_save[NR_CPUS];
70 
71 static inline int r45k_bvahwbug(void)
72 {
73 	/* XXX: We should probe for the presence of this bug, but we don't. */
74 	return 0;
75 }
76 
77 static inline int r4k_250MHZhwbug(void)
78 {
79 	/* XXX: We should probe for the presence of this bug, but we don't. */
80 	return 0;
81 }
82 
83 static inline int __maybe_unused bcm1250_m3_war(void)
84 {
85 	return BCM1250_M3_WAR;
86 }
87 
88 static inline int __maybe_unused r10000_llsc_war(void)
89 {
90 	return R10000_LLSC_WAR;
91 }
92 
93 static int use_bbit_insns(void)
94 {
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:
100 		return 1;
101 	default:
102 		return 0;
103 	}
104 }
105 
106 static int use_lwx_insns(void)
107 {
108 	switch (current_cpu_type()) {
109 	case CPU_CAVIUM_OCTEON2:
110 	case CPU_CAVIUM_OCTEON3:
111 		return 1;
112 	default:
113 		return 0;
114 	}
115 }
116 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
117     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
118 static bool scratchpad_available(void)
119 {
120 	return true;
121 }
122 static int scratchpad_offset(int i)
123 {
124 	/*
125 	 * CVMSEG starts at address -32768 and extends for
126 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
127 	 */
128 	i += 1; /* Kernel use starts at the top and works down. */
129 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
130 }
131 #else
132 static bool scratchpad_available(void)
133 {
134 	return false;
135 }
136 static int scratchpad_offset(int i)
137 {
138 	BUG();
139 	/* Really unreachable, but evidently some GCC want this. */
140 	return 0;
141 }
142 #endif
143 /*
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.
150  *
151  */
152 static int m4kc_tlbp_war(void)
153 {
154 	return (current_cpu_data.processor_id & 0xffff00) ==
155 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
156 }
157 
158 /* Handle labels (which must be positive integers). */
159 enum label_id {
160 	label_second_part = 1,
161 	label_leave,
162 	label_vmalloc,
163 	label_vmalloc_done,
164 	label_tlbw_hazard_0,
165 	label_split = label_tlbw_hazard_0 + 8,
166 	label_tlbl_goaround1,
167 	label_tlbl_goaround2,
168 	label_nopage_tlbl,
169 	label_nopage_tlbs,
170 	label_nopage_tlbm,
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,
176 #endif
177 };
178 
179 UASM_L_LA(_second_part)
180 UASM_L_LA(_leave)
181 UASM_L_LA(_vmalloc)
182 UASM_L_LA(_vmalloc_done)
183 /* _tlbw_hazard_x is handled differently.  */
184 UASM_L_LA(_split)
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)
195 #endif
196 
197 static int hazard_instance;
198 
199 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
200 {
201 	switch (instance) {
202 	case 0 ... 7:
203 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
204 		return;
205 	default:
206 		BUG();
207 	}
208 }
209 
210 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
211 {
212 	switch (instance) {
213 	case 0 ... 7:
214 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
215 		break;
216 	default:
217 		BUG();
218 	}
219 }
220 
221 /*
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.
226  */
227 static void output_pgtable_bits_defines(void)
228 {
229 #define pr_define(fmt, ...)					\
230 	pr_debug("#define " fmt, ##__VA_ARGS__)
231 
232 	pr_debug("#include <asm/asm.h>\n");
233 	pr_debug("#include <asm/regdef.h>\n");
234 	pr_debug("\n");
235 
236 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
237 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_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 #endif
244 #ifdef _PAGE_NO_EXEC_SHIFT
245 	if (cpu_has_rixi)
246 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
247 #endif
248 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
249 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
250 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
251 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
252 	pr_debug("\n");
253 }
254 
255 static inline void dump_handler(const char *symbol, const u32 *handler, int count)
256 {
257 	int i;
258 
259 	pr_debug("LEAF(%s)\n", symbol);
260 
261 	pr_debug("\t.set push\n");
262 	pr_debug("\t.set noreorder\n");
263 
264 	for (i = 0; i < count; i++)
265 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
266 
267 	pr_debug("\t.set\tpop\n");
268 
269 	pr_debug("\tEND(%s)\n", symbol);
270 }
271 
272 /* The only general purpose registers allowed in TLB handlers. */
273 #define K0		26
274 #define K1		27
275 
276 /* Some CP0 registers */
277 #define C0_INDEX	0, 0
278 #define C0_ENTRYLO0	2, 0
279 #define C0_TCBIND	2, 2
280 #define C0_ENTRYLO1	3, 0
281 #define C0_CONTEXT	4, 0
282 #define C0_PAGEMASK	5, 0
283 #define C0_PWBASE	5, 5
284 #define C0_PWFIELD	5, 6
285 #define C0_PWSIZE	5, 7
286 #define C0_PWCTL	6, 6
287 #define C0_BADVADDR	8, 0
288 #define C0_PGD		9, 7
289 #define C0_ENTRYHI	10, 0
290 #define C0_EPC		14, 0
291 #define C0_XCONTEXT	20, 0
292 
293 #ifdef CONFIG_64BIT
294 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
295 #else
296 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
297 #endif
298 
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.
303  *
304  * We deliberately chose a buffer size of 128, so we won't scribble
305  * over anything important on overflow before we panic.
306  */
307 static u32 tlb_handler[128];
308 
309 /* simply assume worst case size for labels and relocs */
310 static struct uasm_label labels[128];
311 static struct uasm_reloc relocs[128];
312 
313 static int check_for_high_segbits;
314 static bool fill_includes_sw_bits;
315 
316 static unsigned int kscratch_used_mask;
317 
318 static inline int __maybe_unused c0_kscratch(void)
319 {
320 	switch (current_cpu_type()) {
321 	case CPU_XLP:
322 	case CPU_XLR:
323 		return 22;
324 	default:
325 		return 31;
326 	}
327 }
328 
329 static int allocate_kscratch(void)
330 {
331 	int r;
332 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
333 
334 	r = ffs(a);
335 
336 	if (r == 0)
337 		return -1;
338 
339 	r--; /* make it zero based */
340 
341 	kscratch_used_mask |= (1 << r);
342 
343 	return r;
344 }
345 
346 static int scratch_reg;
347 static int pgd_reg;
348 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
349 
350 static struct work_registers build_get_work_registers(u32 **p)
351 {
352 	struct work_registers r;
353 
354 	if (scratch_reg >= 0) {
355 		/* Save in CPU local C0_KScratch? */
356 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
357 		r.r1 = K0;
358 		r.r2 = K1;
359 		r.r3 = 1;
360 		return r;
361 	}
362 
363 	if (num_possible_cpus() > 1) {
364 		/* Get smp_processor_id */
365 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
366 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
367 
368 		/* handler_reg_save index in K0 */
369 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
370 
371 		UASM_i_LA(p, K1, (long)&handler_reg_save);
372 		UASM_i_ADDU(p, K0, K0, K1);
373 	} else {
374 		UASM_i_LA(p, K0, (long)&handler_reg_save);
375 	}
376 	/* K0 now points to save area, save $1 and $2  */
377 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
378 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
379 
380 	r.r1 = K1;
381 	r.r2 = 1;
382 	r.r3 = 2;
383 	return r;
384 }
385 
386 static void build_restore_work_registers(u32 **p)
387 {
388 	if (scratch_reg >= 0) {
389 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
390 		return;
391 	}
392 	/* K0 already points to save area, restore $1 and $2  */
393 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
394 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
395 }
396 
397 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
398 
399 /*
400  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
401  * we cannot do r3000 under these circumstances.
402  *
403  * Declare pgd_current here instead of including mmu_context.h to avoid type
404  * conflicts for tlbmiss_handler_setup_pgd
405  */
406 extern unsigned long pgd_current[];
407 
408 /*
409  * The R3000 TLB handler is simple.
410  */
411 static void build_r3000_tlb_refill_handler(void)
412 {
413 	long pgdc = (long)pgd_current;
414 	u32 *p;
415 
416 	memset(tlb_handler, 0, sizeof(tlb_handler));
417 	p = tlb_handler;
418 
419 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
420 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
421 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
422 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
423 	uasm_i_sll(&p, K0, K0, 2);
424 	uasm_i_addu(&p, K1, K1, K0);
425 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
426 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
427 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
428 	uasm_i_addu(&p, K1, K1, K0);
429 	uasm_i_lw(&p, K0, 0, K1);
430 	uasm_i_nop(&p); /* load delay */
431 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
432 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
433 	uasm_i_tlbwr(&p); /* cp0 delay */
434 	uasm_i_jr(&p, K1);
435 	uasm_i_rfe(&p); /* branch delay */
436 
437 	if (p > tlb_handler + 32)
438 		panic("TLB refill handler space exceeded");
439 
440 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
441 		 (unsigned int)(p - tlb_handler));
442 
443 	memcpy((void *)ebase, tlb_handler, 0x80);
444 	local_flush_icache_range(ebase, ebase + 0x80);
445 
446 	dump_handler("r3000_tlb_refill", (u32 *)ebase, 32);
447 }
448 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
449 
450 /*
451  * The R4000 TLB handler is much more complicated. We have two
452  * consecutive handler areas with 32 instructions space each.
453  * Since they aren't used at the same time, we can overflow in the
454  * other one.To keep things simple, we first assume linear space,
455  * then we relocate it to the final handler layout as needed.
456  */
457 static u32 final_handler[64];
458 
459 /*
460  * Hazards
461  *
462  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
463  * 2. A timing hazard exists for the TLBP instruction.
464  *
465  *	stalling_instruction
466  *	TLBP
467  *
468  * The JTLB is being read for the TLBP throughout the stall generated by the
469  * previous instruction. This is not really correct as the stalling instruction
470  * can modify the address used to access the JTLB.  The failure symptom is that
471  * the TLBP instruction will use an address created for the stalling instruction
472  * and not the address held in C0_ENHI and thus report the wrong results.
473  *
474  * The software work-around is to not allow the instruction preceding the TLBP
475  * to stall - make it an NOP or some other instruction guaranteed not to stall.
476  *
477  * Errata 2 will not be fixed.	This errata is also on the R5000.
478  *
479  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
480  */
481 static void __maybe_unused build_tlb_probe_entry(u32 **p)
482 {
483 	switch (current_cpu_type()) {
484 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
485 	case CPU_R4600:
486 	case CPU_R4700:
487 	case CPU_R5000:
488 	case CPU_NEVADA:
489 		uasm_i_nop(p);
490 		uasm_i_tlbp(p);
491 		break;
492 
493 	default:
494 		uasm_i_tlbp(p);
495 		break;
496 	}
497 }
498 
499 /*
500  * Write random or indexed TLB entry, and care about the hazards from
501  * the preceding mtc0 and for the following eret.
502  */
503 enum tlb_write_entry { tlb_random, tlb_indexed };
504 
505 static void build_tlb_write_entry(u32 **p, struct uasm_label **l,
506 				  struct uasm_reloc **r,
507 				  enum tlb_write_entry wmode)
508 {
509 	void(*tlbw)(u32 **) = NULL;
510 
511 	switch (wmode) {
512 	case tlb_random: tlbw = uasm_i_tlbwr; break;
513 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
514 	}
515 
516 	if (cpu_has_mips_r2_r6) {
517 		if (cpu_has_mips_r2_exec_hazard)
518 			uasm_i_ehb(p);
519 		tlbw(p);
520 		return;
521 	}
522 
523 	switch (current_cpu_type()) {
524 	case CPU_R4000PC:
525 	case CPU_R4000SC:
526 	case CPU_R4000MC:
527 	case CPU_R4400PC:
528 	case CPU_R4400SC:
529 	case CPU_R4400MC:
530 		/*
531 		 * This branch uses up a mtc0 hazard nop slot and saves
532 		 * two nops after the tlbw instruction.
533 		 */
534 		uasm_bgezl_hazard(p, r, hazard_instance);
535 		tlbw(p);
536 		uasm_bgezl_label(l, p, hazard_instance);
537 		hazard_instance++;
538 		uasm_i_nop(p);
539 		break;
540 
541 	case CPU_R4600:
542 	case CPU_R4700:
543 		uasm_i_nop(p);
544 		tlbw(p);
545 		uasm_i_nop(p);
546 		break;
547 
548 	case CPU_R5000:
549 	case CPU_NEVADA:
550 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
551 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
552 		tlbw(p);
553 		break;
554 
555 	case CPU_R4300:
556 	case CPU_5KC:
557 	case CPU_TX49XX:
558 	case CPU_PR4450:
559 	case CPU_XLR:
560 		uasm_i_nop(p);
561 		tlbw(p);
562 		break;
563 
564 	case CPU_R10000:
565 	case CPU_R12000:
566 	case CPU_R14000:
567 	case CPU_R16000:
568 	case CPU_4KC:
569 	case CPU_4KEC:
570 	case CPU_M14KC:
571 	case CPU_M14KEC:
572 	case CPU_SB1:
573 	case CPU_SB1A:
574 	case CPU_4KSC:
575 	case CPU_20KC:
576 	case CPU_25KF:
577 	case CPU_BMIPS32:
578 	case CPU_BMIPS3300:
579 	case CPU_BMIPS4350:
580 	case CPU_BMIPS4380:
581 	case CPU_BMIPS5000:
582 	case CPU_LOONGSON2:
583 	case CPU_LOONGSON3:
584 	case CPU_R5500:
585 		if (m4kc_tlbp_war())
586 			uasm_i_nop(p);
587 	case CPU_ALCHEMY:
588 		tlbw(p);
589 		break;
590 
591 	case CPU_RM7000:
592 		uasm_i_nop(p);
593 		uasm_i_nop(p);
594 		uasm_i_nop(p);
595 		uasm_i_nop(p);
596 		tlbw(p);
597 		break;
598 
599 	case CPU_VR4111:
600 	case CPU_VR4121:
601 	case CPU_VR4122:
602 	case CPU_VR4181:
603 	case CPU_VR4181A:
604 		uasm_i_nop(p);
605 		uasm_i_nop(p);
606 		tlbw(p);
607 		uasm_i_nop(p);
608 		uasm_i_nop(p);
609 		break;
610 
611 	case CPU_VR4131:
612 	case CPU_VR4133:
613 	case CPU_R5432:
614 		uasm_i_nop(p);
615 		uasm_i_nop(p);
616 		tlbw(p);
617 		break;
618 
619 	case CPU_JZRISC:
620 		tlbw(p);
621 		uasm_i_nop(p);
622 		break;
623 
624 	default:
625 		panic("No TLB refill handler yet (CPU type: %d)",
626 		      current_cpu_type());
627 		break;
628 	}
629 }
630 
631 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
632 							unsigned int reg)
633 {
634 	if (_PAGE_GLOBAL_SHIFT == 0) {
635 		/* pte_t is already in EntryLo format */
636 		return;
637 	}
638 
639 	if (cpu_has_rixi && _PAGE_NO_EXEC) {
640 		if (fill_includes_sw_bits) {
641 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
642 		} else {
643 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
644 			UASM_i_ROTR(p, reg, reg,
645 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
646 		}
647 	} else {
648 #ifdef CONFIG_PHYS_ADDR_T_64BIT
649 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
650 #else
651 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
652 #endif
653 	}
654 }
655 
656 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
657 
658 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
659 				   unsigned int tmp, enum label_id lid,
660 				   int restore_scratch)
661 {
662 	if (restore_scratch) {
663 		/* Reset default page size */
664 		if (PM_DEFAULT_MASK >> 16) {
665 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
666 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
667 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
668 			uasm_il_b(p, r, lid);
669 		} else if (PM_DEFAULT_MASK) {
670 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
671 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
672 			uasm_il_b(p, r, lid);
673 		} else {
674 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
675 			uasm_il_b(p, r, lid);
676 		}
677 		if (scratch_reg >= 0)
678 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
679 		else
680 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
681 	} else {
682 		/* Reset default page size */
683 		if (PM_DEFAULT_MASK >> 16) {
684 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
685 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
686 			uasm_il_b(p, r, lid);
687 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
688 		} else if (PM_DEFAULT_MASK) {
689 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
690 			uasm_il_b(p, r, lid);
691 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
692 		} else {
693 			uasm_il_b(p, r, lid);
694 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
695 		}
696 	}
697 }
698 
699 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
700 				       struct uasm_reloc **r,
701 				       unsigned int tmp,
702 				       enum tlb_write_entry wmode,
703 				       int restore_scratch)
704 {
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);
709 
710 	build_tlb_write_entry(p, l, r, wmode);
711 
712 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
713 }
714 
715 /*
716  * Check if Huge PTE is present, if so then jump to LABEL.
717  */
718 static void
719 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
720 		  unsigned int pmd, int lid)
721 {
722 	UASM_i_LW(p, tmp, 0, pmd);
723 	if (use_bbit_insns()) {
724 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
725 	} else {
726 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
727 		uasm_il_bnez(p, r, tmp, lid);
728 	}
729 }
730 
731 static void build_huge_update_entries(u32 **p, unsigned int pte,
732 				      unsigned int tmp)
733 {
734 	int small_sequence;
735 
736 	/*
737 	 * A huge PTE describes an area the size of the
738 	 * configured huge page size. This is twice the
739 	 * of the large TLB entry size we intend to use.
740 	 * A TLB entry half the size of the configured
741 	 * huge page size is configured into entrylo0
742 	 * and entrylo1 to cover the contiguous huge PTE
743 	 * address space.
744 	 */
745 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
746 
747 	/* We can clobber tmp.	It isn't used after this.*/
748 	if (!small_sequence)
749 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
750 
751 	build_convert_pte_to_entrylo(p, pte);
752 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
753 	/* convert to entrylo1 */
754 	if (small_sequence)
755 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
756 	else
757 		UASM_i_ADDU(p, pte, pte, tmp);
758 
759 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
760 }
761 
762 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
763 				    struct uasm_label **l,
764 				    unsigned int pte,
765 				    unsigned int ptr)
766 {
767 #ifdef CONFIG_SMP
768 	UASM_i_SC(p, pte, 0, ptr);
769 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
770 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
771 #else
772 	UASM_i_SW(p, pte, 0, ptr);
773 #endif
774 	build_huge_update_entries(p, pte, ptr);
775 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
776 }
777 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
778 
779 #ifdef CONFIG_64BIT
780 /*
781  * TMP and PTR are scratch.
782  * TMP will be clobbered, PTR will hold the pmd entry.
783  */
784 static void
785 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
786 		 unsigned int tmp, unsigned int ptr)
787 {
788 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
789 	long pgdc = (long)pgd_current;
790 #endif
791 	/*
792 	 * The vmalloc handling is not in the hotpath.
793 	 */
794 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
795 
796 	if (check_for_high_segbits) {
797 		/*
798 		 * The kernel currently implicitely assumes that the
799 		 * MIPS SEGBITS parameter for the processor is
800 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
801 		 * allocate virtual addresses outside the maximum
802 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
803 		 * that doesn't prevent user code from accessing the
804 		 * higher xuseg addresses.  Here, we make sure that
805 		 * everything but the lower xuseg addresses goes down
806 		 * the module_alloc/vmalloc path.
807 		 */
808 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
809 		uasm_il_bnez(p, r, ptr, label_vmalloc);
810 	} else {
811 		uasm_il_bltz(p, r, tmp, label_vmalloc);
812 	}
813 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
814 
815 	if (pgd_reg != -1) {
816 		/* pgd is in pgd_reg */
817 		if (cpu_has_ldpte)
818 			UASM_i_MFC0(p, ptr, C0_PWBASE);
819 		else
820 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
821 	} else {
822 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
823 		/*
824 		 * &pgd << 11 stored in CONTEXT [23..63].
825 		 */
826 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
827 
828 		/* Clear lower 23 bits of context. */
829 		uasm_i_dins(p, ptr, 0, 0, 23);
830 
831 		/* 1 0	1 0 1  << 6  xkphys cached */
832 		uasm_i_ori(p, ptr, ptr, 0x540);
833 		uasm_i_drotr(p, ptr, ptr, 11);
834 #elif defined(CONFIG_SMP)
835 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
836 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
837 		UASM_i_LA_mostly(p, tmp, pgdc);
838 		uasm_i_daddu(p, ptr, ptr, tmp);
839 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
840 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
841 #else
842 		UASM_i_LA_mostly(p, ptr, pgdc);
843 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
844 #endif
845 	}
846 
847 	uasm_l_vmalloc_done(l, *p);
848 
849 	/* get pgd offset in bytes */
850 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
851 
852 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
853 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
854 #ifndef __PAGETABLE_PMD_FOLDED
855 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
856 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
857 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
858 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
859 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
860 #endif
861 }
862 
863 /*
864  * BVADDR is the faulting address, PTR is scratch.
865  * PTR will hold the pgd for vmalloc.
866  */
867 static void
868 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
869 			unsigned int bvaddr, unsigned int ptr,
870 			enum vmalloc64_mode mode)
871 {
872 	long swpd = (long)swapper_pg_dir;
873 	int single_insn_swpd;
874 	int did_vmalloc_branch = 0;
875 
876 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
877 
878 	uasm_l_vmalloc(l, *p);
879 
880 	if (mode != not_refill && check_for_high_segbits) {
881 		if (single_insn_swpd) {
882 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
883 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
884 			did_vmalloc_branch = 1;
885 			/* fall through */
886 		} else {
887 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
888 		}
889 	}
890 	if (!did_vmalloc_branch) {
891 		if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
892 			uasm_il_b(p, r, label_vmalloc_done);
893 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
894 		} else {
895 			UASM_i_LA_mostly(p, ptr, swpd);
896 			uasm_il_b(p, r, label_vmalloc_done);
897 			if (uasm_in_compat_space_p(swpd))
898 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
899 			else
900 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
901 		}
902 	}
903 	if (mode != not_refill && check_for_high_segbits) {
904 		uasm_l_large_segbits_fault(l, *p);
905 		/*
906 		 * We get here if we are an xsseg address, or if we are
907 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
908 		 *
909 		 * Ignoring xsseg (assume disabled so would generate
910 		 * (address errors?), the only remaining possibility
911 		 * is the upper xuseg addresses.  On processors with
912 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
913 		 * addresses would have taken an address error. We try
914 		 * to mimic that here by taking a load/istream page
915 		 * fault.
916 		 */
917 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
918 		uasm_i_jr(p, ptr);
919 
920 		if (mode == refill_scratch) {
921 			if (scratch_reg >= 0)
922 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
923 			else
924 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
925 		} else {
926 			uasm_i_nop(p);
927 		}
928 	}
929 }
930 
931 #else /* !CONFIG_64BIT */
932 
933 /*
934  * TMP and PTR are scratch.
935  * TMP will be clobbered, PTR will hold the pgd entry.
936  */
937 static void __maybe_unused
938 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
939 {
940 	if (pgd_reg != -1) {
941 		/* pgd is in pgd_reg */
942 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
943 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
944 	} else {
945 		long pgdc = (long)pgd_current;
946 
947 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
948 #ifdef CONFIG_SMP
949 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
950 		UASM_i_LA_mostly(p, tmp, pgdc);
951 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
952 		uasm_i_addu(p, ptr, tmp, ptr);
953 #else
954 		UASM_i_LA_mostly(p, ptr, pgdc);
955 #endif
956 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
957 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
958 	}
959 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
960 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
961 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
962 }
963 
964 #endif /* !CONFIG_64BIT */
965 
966 static void build_adjust_context(u32 **p, unsigned int ctx)
967 {
968 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
969 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
970 
971 	switch (current_cpu_type()) {
972 	case CPU_VR41XX:
973 	case CPU_VR4111:
974 	case CPU_VR4121:
975 	case CPU_VR4122:
976 	case CPU_VR4131:
977 	case CPU_VR4181:
978 	case CPU_VR4181A:
979 	case CPU_VR4133:
980 		shift += 2;
981 		break;
982 
983 	default:
984 		break;
985 	}
986 
987 	if (shift)
988 		UASM_i_SRL(p, ctx, ctx, shift);
989 	uasm_i_andi(p, ctx, ctx, mask);
990 }
991 
992 static void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
993 {
994 	/*
995 	 * Bug workaround for the Nevada. It seems as if under certain
996 	 * circumstances the move from cp0_context might produce a
997 	 * bogus result when the mfc0 instruction and its consumer are
998 	 * in a different cacheline or a load instruction, probably any
999 	 * memory reference, is between them.
1000 	 */
1001 	switch (current_cpu_type()) {
1002 	case CPU_NEVADA:
1003 		UASM_i_LW(p, ptr, 0, ptr);
1004 		GET_CONTEXT(p, tmp); /* get context reg */
1005 		break;
1006 
1007 	default:
1008 		GET_CONTEXT(p, tmp); /* get context reg */
1009 		UASM_i_LW(p, ptr, 0, ptr);
1010 		break;
1011 	}
1012 
1013 	build_adjust_context(p, tmp);
1014 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1015 }
1016 
1017 static void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1018 {
1019 	int pte_off_even = 0;
1020 	int pte_off_odd = sizeof(pte_t);
1021 
1022 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1023 	/* The low 32 bits of EntryLo is stored in pte_high */
1024 	pte_off_even += offsetof(pte_t, pte_high);
1025 	pte_off_odd += offsetof(pte_t, pte_high);
1026 #endif
1027 
1028 	if (config_enabled(CONFIG_XPA)) {
1029 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1030 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1031 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1032 
1033 		if (cpu_has_xpa && !mips_xpa_disabled) {
1034 			uasm_i_lw(p, tmp, 0, ptep);
1035 			uasm_i_ext(p, tmp, tmp, 0, 24);
1036 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1037 		}
1038 
1039 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1040 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1041 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1042 
1043 		if (cpu_has_xpa && !mips_xpa_disabled) {
1044 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1045 			uasm_i_ext(p, tmp, tmp, 0, 24);
1046 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1047 		}
1048 		return;
1049 	}
1050 
1051 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1052 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1053 	if (r45k_bvahwbug())
1054 		build_tlb_probe_entry(p);
1055 	build_convert_pte_to_entrylo(p, tmp);
1056 	if (r4k_250MHZhwbug())
1057 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1058 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1059 	build_convert_pte_to_entrylo(p, ptep);
1060 	if (r45k_bvahwbug())
1061 		uasm_i_mfc0(p, tmp, C0_INDEX);
1062 	if (r4k_250MHZhwbug())
1063 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1064 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1065 }
1066 
1067 struct mips_huge_tlb_info {
1068 	int huge_pte;
1069 	int restore_scratch;
1070 	bool need_reload_pte;
1071 };
1072 
1073 static struct mips_huge_tlb_info
1074 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1075 			       struct uasm_reloc **r, unsigned int tmp,
1076 			       unsigned int ptr, int c0_scratch_reg)
1077 {
1078 	struct mips_huge_tlb_info rv;
1079 	unsigned int even, odd;
1080 	int vmalloc_branch_delay_filled = 0;
1081 	const int scratch = 1; /* Our extra working register */
1082 
1083 	rv.huge_pte = scratch;
1084 	rv.restore_scratch = 0;
1085 	rv.need_reload_pte = false;
1086 
1087 	if (check_for_high_segbits) {
1088 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1089 
1090 		if (pgd_reg != -1)
1091 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1092 		else
1093 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1094 
1095 		if (c0_scratch_reg >= 0)
1096 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1097 		else
1098 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1099 
1100 		uasm_i_dsrl_safe(p, scratch, tmp,
1101 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1102 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1103 
1104 		if (pgd_reg == -1) {
1105 			vmalloc_branch_delay_filled = 1;
1106 			/* Clear lower 23 bits of context. */
1107 			uasm_i_dins(p, ptr, 0, 0, 23);
1108 		}
1109 	} else {
1110 		if (pgd_reg != -1)
1111 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1112 		else
1113 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1114 
1115 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1116 
1117 		if (c0_scratch_reg >= 0)
1118 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1119 		else
1120 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1121 
1122 		if (pgd_reg == -1)
1123 			/* Clear lower 23 bits of context. */
1124 			uasm_i_dins(p, ptr, 0, 0, 23);
1125 
1126 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1127 	}
1128 
1129 	if (pgd_reg == -1) {
1130 		vmalloc_branch_delay_filled = 1;
1131 		/* 1 0	1 0 1  << 6  xkphys cached */
1132 		uasm_i_ori(p, ptr, ptr, 0x540);
1133 		uasm_i_drotr(p, ptr, ptr, 11);
1134 	}
1135 
1136 #ifdef __PAGETABLE_PMD_FOLDED
1137 #define LOC_PTEP scratch
1138 #else
1139 #define LOC_PTEP ptr
1140 #endif
1141 
1142 	if (!vmalloc_branch_delay_filled)
1143 		/* get pgd offset in bytes */
1144 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1145 
1146 	uasm_l_vmalloc_done(l, *p);
1147 
1148 	/*
1149 	 *			   tmp		ptr
1150 	 * fall-through case =	 badvaddr  *pgd_current
1151 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1152 	 */
1153 
1154 	if (vmalloc_branch_delay_filled)
1155 		/* get pgd offset in bytes */
1156 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1157 
1158 #ifdef __PAGETABLE_PMD_FOLDED
1159 	GET_CONTEXT(p, tmp); /* get context reg */
1160 #endif
1161 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1162 
1163 	if (use_lwx_insns()) {
1164 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1165 	} else {
1166 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1167 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1168 	}
1169 
1170 #ifndef __PAGETABLE_PMD_FOLDED
1171 	/* get pmd offset in bytes */
1172 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1173 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1174 	GET_CONTEXT(p, tmp); /* get context reg */
1175 
1176 	if (use_lwx_insns()) {
1177 		UASM_i_LWX(p, scratch, scratch, ptr);
1178 	} else {
1179 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1180 		UASM_i_LW(p, scratch, 0, ptr);
1181 	}
1182 #endif
1183 	/* Adjust the context during the load latency. */
1184 	build_adjust_context(p, tmp);
1185 
1186 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1187 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1188 	/*
1189 	 * The in the LWX case we don't want to do the load in the
1190 	 * delay slot.	It cannot issue in the same cycle and may be
1191 	 * speculative and unneeded.
1192 	 */
1193 	if (use_lwx_insns())
1194 		uasm_i_nop(p);
1195 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1196 
1197 
1198 	/* build_update_entries */
1199 	if (use_lwx_insns()) {
1200 		even = ptr;
1201 		odd = tmp;
1202 		UASM_i_LWX(p, even, scratch, tmp);
1203 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1204 		UASM_i_LWX(p, odd, scratch, tmp);
1205 	} else {
1206 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1207 		even = tmp;
1208 		odd = ptr;
1209 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1210 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1211 	}
1212 	if (cpu_has_rixi) {
1213 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1214 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1215 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1216 	} else {
1217 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1218 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1219 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1220 	}
1221 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1222 
1223 	if (c0_scratch_reg >= 0) {
1224 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1225 		build_tlb_write_entry(p, l, r, tlb_random);
1226 		uasm_l_leave(l, *p);
1227 		rv.restore_scratch = 1;
1228 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1229 		build_tlb_write_entry(p, l, r, tlb_random);
1230 		uasm_l_leave(l, *p);
1231 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1232 	} else {
1233 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1234 		build_tlb_write_entry(p, l, r, tlb_random);
1235 		uasm_l_leave(l, *p);
1236 		rv.restore_scratch = 1;
1237 	}
1238 
1239 	uasm_i_eret(p); /* return from trap */
1240 
1241 	return rv;
1242 }
1243 
1244 /*
1245  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1246  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1247  * slots before the XTLB refill exception handler which belong to the
1248  * unused TLB refill exception.
1249  */
1250 #define MIPS64_REFILL_INSNS 32
1251 
1252 static void build_r4000_tlb_refill_handler(void)
1253 {
1254 	u32 *p = tlb_handler;
1255 	struct uasm_label *l = labels;
1256 	struct uasm_reloc *r = relocs;
1257 	u32 *f;
1258 	unsigned int final_len;
1259 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1260 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1261 
1262 	memset(tlb_handler, 0, sizeof(tlb_handler));
1263 	memset(labels, 0, sizeof(labels));
1264 	memset(relocs, 0, sizeof(relocs));
1265 	memset(final_handler, 0, sizeof(final_handler));
1266 
1267 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1268 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1269 							  scratch_reg);
1270 		vmalloc_mode = refill_scratch;
1271 	} else {
1272 		htlb_info.huge_pte = K0;
1273 		htlb_info.restore_scratch = 0;
1274 		htlb_info.need_reload_pte = true;
1275 		vmalloc_mode = refill_noscratch;
1276 		/*
1277 		 * create the plain linear handler
1278 		 */
1279 		if (bcm1250_m3_war()) {
1280 			unsigned int segbits = 44;
1281 
1282 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1283 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1284 			uasm_i_xor(&p, K0, K0, K1);
1285 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1286 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1287 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1288 			uasm_i_or(&p, K0, K0, K1);
1289 			uasm_il_bnez(&p, &r, K0, label_leave);
1290 			/* No need for uasm_i_nop */
1291 		}
1292 
1293 #ifdef CONFIG_64BIT
1294 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1295 #else
1296 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1297 #endif
1298 
1299 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1300 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1301 #endif
1302 
1303 		build_get_ptep(&p, K0, K1);
1304 		build_update_entries(&p, K0, K1);
1305 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1306 		uasm_l_leave(&l, p);
1307 		uasm_i_eret(&p); /* return from trap */
1308 	}
1309 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1310 	uasm_l_tlb_huge_update(&l, p);
1311 	if (htlb_info.need_reload_pte)
1312 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1313 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1314 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1315 				   htlb_info.restore_scratch);
1316 #endif
1317 
1318 #ifdef CONFIG_64BIT
1319 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1320 #endif
1321 
1322 	/*
1323 	 * Overflow check: For the 64bit handler, we need at least one
1324 	 * free instruction slot for the wrap-around branch. In worst
1325 	 * case, if the intended insertion point is a delay slot, we
1326 	 * need three, with the second nop'ed and the third being
1327 	 * unused.
1328 	 */
1329 	switch (boot_cpu_type()) {
1330 	default:
1331 		if (sizeof(long) == 4) {
1332 	case CPU_LOONGSON2:
1333 		/* Loongson2 ebase is different than r4k, we have more space */
1334 			if ((p - tlb_handler) > 64)
1335 				panic("TLB refill handler space exceeded");
1336 			/*
1337 			 * Now fold the handler in the TLB refill handler space.
1338 			 */
1339 			f = final_handler;
1340 			/* Simplest case, just copy the handler. */
1341 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1342 			final_len = p - tlb_handler;
1343 			break;
1344 		} else {
1345 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1346 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1347 				&& uasm_insn_has_bdelay(relocs,
1348 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1349 				panic("TLB refill handler space exceeded");
1350 			/*
1351 			 * Now fold the handler in the TLB refill handler space.
1352 			 */
1353 			f = final_handler + MIPS64_REFILL_INSNS;
1354 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1355 				/* Just copy the handler. */
1356 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1357 				final_len = p - tlb_handler;
1358 			} else {
1359 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1360 				const enum label_id ls = label_tlb_huge_update;
1361 #else
1362 				const enum label_id ls = label_vmalloc;
1363 #endif
1364 				u32 *split;
1365 				int ov = 0;
1366 				int i;
1367 
1368 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1369 					;
1370 				BUG_ON(i == ARRAY_SIZE(labels));
1371 				split = labels[i].addr;
1372 
1373 				/*
1374 				 * See if we have overflown one way or the other.
1375 				 */
1376 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1377 				    split < p - MIPS64_REFILL_INSNS)
1378 					ov = 1;
1379 
1380 				if (ov) {
1381 					/*
1382 					 * Split two instructions before the end.  One
1383 					 * for the branch and one for the instruction
1384 					 * in the delay slot.
1385 					 */
1386 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1387 
1388 					/*
1389 					 * If the branch would fall in a delay slot,
1390 					 * we must back up an additional instruction
1391 					 * so that it is no longer in a delay slot.
1392 					 */
1393 					if (uasm_insn_has_bdelay(relocs, split - 1))
1394 						split--;
1395 				}
1396 				/* Copy first part of the handler. */
1397 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1398 				f += split - tlb_handler;
1399 
1400 				if (ov) {
1401 					/* Insert branch. */
1402 					uasm_l_split(&l, final_handler);
1403 					uasm_il_b(&f, &r, label_split);
1404 					if (uasm_insn_has_bdelay(relocs, split))
1405 						uasm_i_nop(&f);
1406 					else {
1407 						uasm_copy_handler(relocs, labels,
1408 								  split, split + 1, f);
1409 						uasm_move_labels(labels, f, f + 1, -1);
1410 						f++;
1411 						split++;
1412 					}
1413 				}
1414 
1415 				/* Copy the rest of the handler. */
1416 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1417 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1418 					    (p - split);
1419 			}
1420 		}
1421 		break;
1422 	}
1423 
1424 	uasm_resolve_relocs(relocs, labels);
1425 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1426 		 final_len);
1427 
1428 	memcpy((void *)ebase, final_handler, 0x100);
1429 	local_flush_icache_range(ebase, ebase + 0x100);
1430 
1431 	dump_handler("r4000_tlb_refill", (u32 *)ebase, 64);
1432 }
1433 
1434 static void setup_pw(void)
1435 {
1436 	unsigned long pgd_i, pgd_w;
1437 #ifndef __PAGETABLE_PMD_FOLDED
1438 	unsigned long pmd_i, pmd_w;
1439 #endif
1440 	unsigned long pt_i, pt_w;
1441 	unsigned long pte_i, pte_w;
1442 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1443 	unsigned long psn;
1444 
1445 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1446 #endif
1447 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1448 #ifndef __PAGETABLE_PMD_FOLDED
1449 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1450 
1451 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1452 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1453 #else
1454 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1455 #endif
1456 
1457 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1458 	pt_w  = PAGE_SHIFT - 3;
1459 
1460 	pte_i = ilog2(_PAGE_GLOBAL);
1461 	pte_w = 0;
1462 
1463 #ifndef __PAGETABLE_PMD_FOLDED
1464 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1465 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1466 #else
1467 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1468 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1469 #endif
1470 
1471 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1472 	write_c0_pwctl(1 << 6 | psn);
1473 #endif
1474 	write_c0_kpgd(swapper_pg_dir);
1475 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1476 }
1477 
1478 static void build_loongson3_tlb_refill_handler(void)
1479 {
1480 	u32 *p = tlb_handler;
1481 	struct uasm_label *l = labels;
1482 	struct uasm_reloc *r = relocs;
1483 
1484 	memset(labels, 0, sizeof(labels));
1485 	memset(relocs, 0, sizeof(relocs));
1486 	memset(tlb_handler, 0, sizeof(tlb_handler));
1487 
1488 	if (check_for_high_segbits) {
1489 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1490 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1491 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1492 		uasm_i_nop(&p);
1493 
1494 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1495 		uasm_i_nop(&p);
1496 		uasm_l_vmalloc(&l, p);
1497 	}
1498 
1499 	uasm_i_dmfc0(&p, K1, C0_PGD);
1500 
1501 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1502 #ifndef __PAGETABLE_PMD_FOLDED
1503 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1504 #endif
1505 	uasm_i_ldpte(&p, K1, 0);      /* even */
1506 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1507 	uasm_i_tlbwr(&p);
1508 
1509 	/* restore page mask */
1510 	if (PM_DEFAULT_MASK >> 16) {
1511 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1512 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1513 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1514 	} else if (PM_DEFAULT_MASK) {
1515 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1516 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1517 	} else {
1518 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1519 	}
1520 
1521 	uasm_i_eret(&p);
1522 
1523 	if (check_for_high_segbits) {
1524 		uasm_l_large_segbits_fault(&l, p);
1525 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1526 		uasm_i_jr(&p, K1);
1527 		uasm_i_nop(&p);
1528 	}
1529 
1530 	uasm_resolve_relocs(relocs, labels);
1531 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1532 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1533 	dump_handler("loongson3_tlb_refill", (u32 *)(ebase + 0x80), 32);
1534 }
1535 
1536 extern u32 handle_tlbl[], handle_tlbl_end[];
1537 extern u32 handle_tlbs[], handle_tlbs_end[];
1538 extern u32 handle_tlbm[], handle_tlbm_end[];
1539 extern u32 tlbmiss_handler_setup_pgd_start[], tlbmiss_handler_setup_pgd[];
1540 extern u32 tlbmiss_handler_setup_pgd_end[];
1541 
1542 static void build_setup_pgd(void)
1543 {
1544 	const int a0 = 4;
1545 	const int __maybe_unused a1 = 5;
1546 	const int __maybe_unused a2 = 6;
1547 	u32 *p = tlbmiss_handler_setup_pgd_start;
1548 	const int tlbmiss_handler_setup_pgd_size =
1549 		tlbmiss_handler_setup_pgd_end - tlbmiss_handler_setup_pgd_start;
1550 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1551 	long pgdc = (long)pgd_current;
1552 #endif
1553 
1554 	memset(tlbmiss_handler_setup_pgd, 0, tlbmiss_handler_setup_pgd_size *
1555 					sizeof(tlbmiss_handler_setup_pgd[0]));
1556 	memset(labels, 0, sizeof(labels));
1557 	memset(relocs, 0, sizeof(relocs));
1558 	pgd_reg = allocate_kscratch();
1559 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1560 	if (pgd_reg == -1) {
1561 		struct uasm_label *l = labels;
1562 		struct uasm_reloc *r = relocs;
1563 
1564 		/* PGD << 11 in c0_Context */
1565 		/*
1566 		 * If it is a ckseg0 address, convert to a physical
1567 		 * address.  Shifting right by 29 and adding 4 will
1568 		 * result in zero for these addresses.
1569 		 *
1570 		 */
1571 		UASM_i_SRA(&p, a1, a0, 29);
1572 		UASM_i_ADDIU(&p, a1, a1, 4);
1573 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1574 		uasm_i_nop(&p);
1575 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1576 		uasm_l_tlbl_goaround1(&l, p);
1577 		UASM_i_SLL(&p, a0, a0, 11);
1578 		uasm_i_jr(&p, 31);
1579 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1580 	} else {
1581 		/* PGD in c0_KScratch */
1582 		uasm_i_jr(&p, 31);
1583 		if (cpu_has_ldpte)
1584 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1585 		else
1586 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1587 	}
1588 #else
1589 #ifdef CONFIG_SMP
1590 	/* Save PGD to pgd_current[smp_processor_id()] */
1591 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1592 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1593 	UASM_i_LA_mostly(&p, a2, pgdc);
1594 	UASM_i_ADDU(&p, a2, a2, a1);
1595 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1596 #else
1597 	UASM_i_LA_mostly(&p, a2, pgdc);
1598 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1599 #endif /* SMP */
1600 	uasm_i_jr(&p, 31);
1601 
1602 	/* if pgd_reg is allocated, save PGD also to scratch register */
1603 	if (pgd_reg != -1)
1604 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1605 	else
1606 		uasm_i_nop(&p);
1607 #endif
1608 	if (p >= tlbmiss_handler_setup_pgd_end)
1609 		panic("tlbmiss_handler_setup_pgd space exceeded");
1610 
1611 	uasm_resolve_relocs(relocs, labels);
1612 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1613 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1614 
1615 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1616 					tlbmiss_handler_setup_pgd_size);
1617 }
1618 
1619 static void
1620 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1621 {
1622 #ifdef CONFIG_SMP
1623 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1624 	if (cpu_has_64bits)
1625 		uasm_i_lld(p, pte, 0, ptr);
1626 	else
1627 # endif
1628 		UASM_i_LL(p, pte, 0, ptr);
1629 #else
1630 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1631 	if (cpu_has_64bits)
1632 		uasm_i_ld(p, pte, 0, ptr);
1633 	else
1634 # endif
1635 		UASM_i_LW(p, pte, 0, ptr);
1636 #endif
1637 }
1638 
1639 static void
1640 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1641 	unsigned int mode, unsigned int scratch)
1642 {
1643 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1644 	unsigned int swmode = mode & ~hwmode;
1645 
1646 	if (config_enabled(CONFIG_XPA) && !cpu_has_64bits) {
1647 		uasm_i_lui(p, scratch, swmode >> 16);
1648 		uasm_i_or(p, pte, pte, scratch);
1649 		BUG_ON(swmode & 0xffff);
1650 	} else {
1651 		uasm_i_ori(p, pte, pte, mode);
1652 	}
1653 
1654 #ifdef CONFIG_SMP
1655 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1656 	if (cpu_has_64bits)
1657 		uasm_i_scd(p, pte, 0, ptr);
1658 	else
1659 # endif
1660 		UASM_i_SC(p, pte, 0, ptr);
1661 
1662 	if (r10000_llsc_war())
1663 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1664 	else
1665 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1666 
1667 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1668 	if (!cpu_has_64bits) {
1669 		/* no uasm_i_nop needed */
1670 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1671 		uasm_i_ori(p, pte, pte, hwmode);
1672 		BUG_ON(hwmode & ~0xffff);
1673 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1674 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1675 		/* no uasm_i_nop needed */
1676 		uasm_i_lw(p, pte, 0, ptr);
1677 	} else
1678 		uasm_i_nop(p);
1679 # else
1680 	uasm_i_nop(p);
1681 # endif
1682 #else
1683 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1684 	if (cpu_has_64bits)
1685 		uasm_i_sd(p, pte, 0, ptr);
1686 	else
1687 # endif
1688 		UASM_i_SW(p, pte, 0, ptr);
1689 
1690 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1691 	if (!cpu_has_64bits) {
1692 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1693 		uasm_i_ori(p, pte, pte, hwmode);
1694 		BUG_ON(hwmode & ~0xffff);
1695 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1696 		uasm_i_lw(p, pte, 0, ptr);
1697 	}
1698 # endif
1699 #endif
1700 }
1701 
1702 /*
1703  * Check if PTE is present, if not then jump to LABEL. PTR points to
1704  * the page table where this PTE is located, PTE will be re-loaded
1705  * with it's original value.
1706  */
1707 static void
1708 build_pte_present(u32 **p, struct uasm_reloc **r,
1709 		  int pte, int ptr, int scratch, enum label_id lid)
1710 {
1711 	int t = scratch >= 0 ? scratch : pte;
1712 	int cur = pte;
1713 
1714 	if (cpu_has_rixi) {
1715 		if (use_bbit_insns()) {
1716 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1717 			uasm_i_nop(p);
1718 		} else {
1719 			if (_PAGE_PRESENT_SHIFT) {
1720 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1721 				cur = t;
1722 			}
1723 			uasm_i_andi(p, t, cur, 1);
1724 			uasm_il_beqz(p, r, t, lid);
1725 			if (pte == t)
1726 				/* You lose the SMP race :-(*/
1727 				iPTE_LW(p, pte, ptr);
1728 		}
1729 	} else {
1730 		if (_PAGE_PRESENT_SHIFT) {
1731 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1732 			cur = t;
1733 		}
1734 		uasm_i_andi(p, t, cur,
1735 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1736 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1737 		uasm_il_bnez(p, r, t, lid);
1738 		if (pte == t)
1739 			/* You lose the SMP race :-(*/
1740 			iPTE_LW(p, pte, ptr);
1741 	}
1742 }
1743 
1744 /* Make PTE valid, store result in PTR. */
1745 static void
1746 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1747 		 unsigned int ptr, unsigned int scratch)
1748 {
1749 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1750 
1751 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1752 }
1753 
1754 /*
1755  * Check if PTE can be written to, if not branch to LABEL. Regardless
1756  * restore PTE with value from PTR when done.
1757  */
1758 static void
1759 build_pte_writable(u32 **p, struct uasm_reloc **r,
1760 		   unsigned int pte, unsigned int ptr, int scratch,
1761 		   enum label_id lid)
1762 {
1763 	int t = scratch >= 0 ? scratch : pte;
1764 	int cur = pte;
1765 
1766 	if (_PAGE_PRESENT_SHIFT) {
1767 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1768 		cur = t;
1769 	}
1770 	uasm_i_andi(p, t, cur,
1771 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1772 	uasm_i_xori(p, t, t,
1773 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1774 	uasm_il_bnez(p, r, t, lid);
1775 	if (pte == t)
1776 		/* You lose the SMP race :-(*/
1777 		iPTE_LW(p, pte, ptr);
1778 	else
1779 		uasm_i_nop(p);
1780 }
1781 
1782 /* Make PTE writable, update software status bits as well, then store
1783  * at PTR.
1784  */
1785 static void
1786 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1787 		 unsigned int ptr, unsigned int scratch)
1788 {
1789 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1790 			     | _PAGE_DIRTY);
1791 
1792 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1793 }
1794 
1795 /*
1796  * Check if PTE can be modified, if not branch to LABEL. Regardless
1797  * restore PTE with value from PTR when done.
1798  */
1799 static void
1800 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1801 		     unsigned int pte, unsigned int ptr, int scratch,
1802 		     enum label_id lid)
1803 {
1804 	if (use_bbit_insns()) {
1805 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1806 		uasm_i_nop(p);
1807 	} else {
1808 		int t = scratch >= 0 ? scratch : pte;
1809 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1810 		uasm_i_andi(p, t, t, 1);
1811 		uasm_il_beqz(p, r, t, lid);
1812 		if (pte == t)
1813 			/* You lose the SMP race :-(*/
1814 			iPTE_LW(p, pte, ptr);
1815 	}
1816 }
1817 
1818 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1819 
1820 
1821 /*
1822  * R3000 style TLB load/store/modify handlers.
1823  */
1824 
1825 /*
1826  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1827  * Then it returns.
1828  */
1829 static void
1830 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1831 {
1832 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1833 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1834 	uasm_i_tlbwi(p);
1835 	uasm_i_jr(p, tmp);
1836 	uasm_i_rfe(p); /* branch delay */
1837 }
1838 
1839 /*
1840  * This places the pte into ENTRYLO0 and writes it with tlbwi
1841  * or tlbwr as appropriate.  This is because the index register
1842  * may have the probe fail bit set as a result of a trap on a
1843  * kseg2 access, i.e. without refill.  Then it returns.
1844  */
1845 static void
1846 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1847 			     struct uasm_reloc **r, unsigned int pte,
1848 			     unsigned int tmp)
1849 {
1850 	uasm_i_mfc0(p, tmp, C0_INDEX);
1851 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1852 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1853 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1854 	uasm_i_tlbwi(p); /* cp0 delay */
1855 	uasm_i_jr(p, tmp);
1856 	uasm_i_rfe(p); /* branch delay */
1857 	uasm_l_r3000_write_probe_fail(l, *p);
1858 	uasm_i_tlbwr(p); /* cp0 delay */
1859 	uasm_i_jr(p, tmp);
1860 	uasm_i_rfe(p); /* branch delay */
1861 }
1862 
1863 static void
1864 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1865 				   unsigned int ptr)
1866 {
1867 	long pgdc = (long)pgd_current;
1868 
1869 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1870 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1871 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1872 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1873 	uasm_i_sll(p, pte, pte, 2);
1874 	uasm_i_addu(p, ptr, ptr, pte);
1875 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1876 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1877 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1878 	uasm_i_addu(p, ptr, ptr, pte);
1879 	uasm_i_lw(p, pte, 0, ptr);
1880 	uasm_i_tlbp(p); /* load delay */
1881 }
1882 
1883 static void build_r3000_tlb_load_handler(void)
1884 {
1885 	u32 *p = handle_tlbl;
1886 	const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
1887 	struct uasm_label *l = labels;
1888 	struct uasm_reloc *r = relocs;
1889 
1890 	memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
1891 	memset(labels, 0, sizeof(labels));
1892 	memset(relocs, 0, sizeof(relocs));
1893 
1894 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1895 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1896 	uasm_i_nop(&p); /* load delay */
1897 	build_make_valid(&p, &r, K0, K1, -1);
1898 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1899 
1900 	uasm_l_nopage_tlbl(&l, p);
1901 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1902 	uasm_i_nop(&p);
1903 
1904 	if (p >= handle_tlbl_end)
1905 		panic("TLB load handler fastpath space exceeded");
1906 
1907 	uasm_resolve_relocs(relocs, labels);
1908 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1909 		 (unsigned int)(p - handle_tlbl));
1910 
1911 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_size);
1912 }
1913 
1914 static void build_r3000_tlb_store_handler(void)
1915 {
1916 	u32 *p = handle_tlbs;
1917 	const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
1918 	struct uasm_label *l = labels;
1919 	struct uasm_reloc *r = relocs;
1920 
1921 	memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
1922 	memset(labels, 0, sizeof(labels));
1923 	memset(relocs, 0, sizeof(relocs));
1924 
1925 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1926 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1927 	uasm_i_nop(&p); /* load delay */
1928 	build_make_write(&p, &r, K0, K1, -1);
1929 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1930 
1931 	uasm_l_nopage_tlbs(&l, p);
1932 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1933 	uasm_i_nop(&p);
1934 
1935 	if (p >= handle_tlbs_end)
1936 		panic("TLB store handler fastpath space exceeded");
1937 
1938 	uasm_resolve_relocs(relocs, labels);
1939 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1940 		 (unsigned int)(p - handle_tlbs));
1941 
1942 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_size);
1943 }
1944 
1945 static void build_r3000_tlb_modify_handler(void)
1946 {
1947 	u32 *p = handle_tlbm;
1948 	const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
1949 	struct uasm_label *l = labels;
1950 	struct uasm_reloc *r = relocs;
1951 
1952 	memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
1953 	memset(labels, 0, sizeof(labels));
1954 	memset(relocs, 0, sizeof(relocs));
1955 
1956 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1957 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1958 	uasm_i_nop(&p); /* load delay */
1959 	build_make_write(&p, &r, K0, K1, -1);
1960 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1961 
1962 	uasm_l_nopage_tlbm(&l, p);
1963 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1964 	uasm_i_nop(&p);
1965 
1966 	if (p >= handle_tlbm_end)
1967 		panic("TLB modify handler fastpath space exceeded");
1968 
1969 	uasm_resolve_relocs(relocs, labels);
1970 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1971 		 (unsigned int)(p - handle_tlbm));
1972 
1973 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_size);
1974 }
1975 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1976 
1977 /*
1978  * R4000 style TLB load/store/modify handlers.
1979  */
1980 static struct work_registers
1981 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1982 				   struct uasm_reloc **r)
1983 {
1984 	struct work_registers wr = build_get_work_registers(p);
1985 
1986 #ifdef CONFIG_64BIT
1987 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1988 #else
1989 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1990 #endif
1991 
1992 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1993 	/*
1994 	 * For huge tlb entries, pmd doesn't contain an address but
1995 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1996 	 * see if we need to jump to huge tlb processing.
1997 	 */
1998 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1999 #endif
2000 
2001 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2002 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2003 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2004 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2005 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2006 
2007 #ifdef CONFIG_SMP
2008 	uasm_l_smp_pgtable_change(l, *p);
2009 #endif
2010 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2011 	if (!m4kc_tlbp_war()) {
2012 		build_tlb_probe_entry(p);
2013 		if (cpu_has_htw) {
2014 			/* race condition happens, leaving */
2015 			uasm_i_ehb(p);
2016 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2017 			uasm_il_bltz(p, r, wr.r3, label_leave);
2018 			uasm_i_nop(p);
2019 		}
2020 	}
2021 	return wr;
2022 }
2023 
2024 static void
2025 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2026 				   struct uasm_reloc **r, unsigned int tmp,
2027 				   unsigned int ptr)
2028 {
2029 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2030 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2031 	build_update_entries(p, tmp, ptr);
2032 	build_tlb_write_entry(p, l, r, tlb_indexed);
2033 	uasm_l_leave(l, *p);
2034 	build_restore_work_registers(p);
2035 	uasm_i_eret(p); /* return from trap */
2036 
2037 #ifdef CONFIG_64BIT
2038 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2039 #endif
2040 }
2041 
2042 static void build_r4000_tlb_load_handler(void)
2043 {
2044 	u32 *p = handle_tlbl;
2045 	const int handle_tlbl_size = handle_tlbl_end - handle_tlbl;
2046 	struct uasm_label *l = labels;
2047 	struct uasm_reloc *r = relocs;
2048 	struct work_registers wr;
2049 
2050 	memset(handle_tlbl, 0, handle_tlbl_size * sizeof(handle_tlbl[0]));
2051 	memset(labels, 0, sizeof(labels));
2052 	memset(relocs, 0, sizeof(relocs));
2053 
2054 	if (bcm1250_m3_war()) {
2055 		unsigned int segbits = 44;
2056 
2057 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2058 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2059 		uasm_i_xor(&p, K0, K0, K1);
2060 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2061 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2062 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2063 		uasm_i_or(&p, K0, K0, K1);
2064 		uasm_il_bnez(&p, &r, K0, label_leave);
2065 		/* No need for uasm_i_nop */
2066 	}
2067 
2068 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2069 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2070 	if (m4kc_tlbp_war())
2071 		build_tlb_probe_entry(&p);
2072 
2073 	if (cpu_has_rixi && !cpu_has_rixiex) {
2074 		/*
2075 		 * If the page is not _PAGE_VALID, RI or XI could not
2076 		 * have triggered it.  Skip the expensive test..
2077 		 */
2078 		if (use_bbit_insns()) {
2079 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2080 				      label_tlbl_goaround1);
2081 		} else {
2082 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2083 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2084 		}
2085 		uasm_i_nop(&p);
2086 
2087 		uasm_i_tlbr(&p);
2088 
2089 		switch (current_cpu_type()) {
2090 		default:
2091 			if (cpu_has_mips_r2_exec_hazard) {
2092 				uasm_i_ehb(&p);
2093 
2094 		case CPU_CAVIUM_OCTEON:
2095 		case CPU_CAVIUM_OCTEON_PLUS:
2096 		case CPU_CAVIUM_OCTEON2:
2097 				break;
2098 			}
2099 		}
2100 
2101 		/* Examine  entrylo 0 or 1 based on ptr. */
2102 		if (use_bbit_insns()) {
2103 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2104 		} else {
2105 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2106 			uasm_i_beqz(&p, wr.r3, 8);
2107 		}
2108 		/* load it in the delay slot*/
2109 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2110 		/* load it if ptr is odd */
2111 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2112 		/*
2113 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2114 		 * XI must have triggered it.
2115 		 */
2116 		if (use_bbit_insns()) {
2117 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2118 			uasm_i_nop(&p);
2119 			uasm_l_tlbl_goaround1(&l, p);
2120 		} else {
2121 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2122 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2123 			uasm_i_nop(&p);
2124 		}
2125 		uasm_l_tlbl_goaround1(&l, p);
2126 	}
2127 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2128 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2129 
2130 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2131 	/*
2132 	 * This is the entry point when build_r4000_tlbchange_handler_head
2133 	 * spots a huge page.
2134 	 */
2135 	uasm_l_tlb_huge_update(&l, p);
2136 	iPTE_LW(&p, wr.r1, wr.r2);
2137 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2138 	build_tlb_probe_entry(&p);
2139 
2140 	if (cpu_has_rixi && !cpu_has_rixiex) {
2141 		/*
2142 		 * If the page is not _PAGE_VALID, RI or XI could not
2143 		 * have triggered it.  Skip the expensive test..
2144 		 */
2145 		if (use_bbit_insns()) {
2146 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2147 				      label_tlbl_goaround2);
2148 		} else {
2149 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2150 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2151 		}
2152 		uasm_i_nop(&p);
2153 
2154 		uasm_i_tlbr(&p);
2155 
2156 		switch (current_cpu_type()) {
2157 		default:
2158 			if (cpu_has_mips_r2_exec_hazard) {
2159 				uasm_i_ehb(&p);
2160 
2161 		case CPU_CAVIUM_OCTEON:
2162 		case CPU_CAVIUM_OCTEON_PLUS:
2163 		case CPU_CAVIUM_OCTEON2:
2164 				break;
2165 			}
2166 		}
2167 
2168 		/* Examine  entrylo 0 or 1 based on ptr. */
2169 		if (use_bbit_insns()) {
2170 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2171 		} else {
2172 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2173 			uasm_i_beqz(&p, wr.r3, 8);
2174 		}
2175 		/* load it in the delay slot*/
2176 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2177 		/* load it if ptr is odd */
2178 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2179 		/*
2180 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2181 		 * XI must have triggered it.
2182 		 */
2183 		if (use_bbit_insns()) {
2184 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2185 		} else {
2186 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2187 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2188 		}
2189 		if (PM_DEFAULT_MASK == 0)
2190 			uasm_i_nop(&p);
2191 		/*
2192 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2193 		 * it is restored in build_huge_tlb_write_entry.
2194 		 */
2195 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2196 
2197 		uasm_l_tlbl_goaround2(&l, p);
2198 	}
2199 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2200 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2201 #endif
2202 
2203 	uasm_l_nopage_tlbl(&l, p);
2204 	build_restore_work_registers(&p);
2205 #ifdef CONFIG_CPU_MICROMIPS
2206 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2207 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2208 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2209 		uasm_i_jr(&p, K0);
2210 	} else
2211 #endif
2212 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2213 	uasm_i_nop(&p);
2214 
2215 	if (p >= handle_tlbl_end)
2216 		panic("TLB load handler fastpath space exceeded");
2217 
2218 	uasm_resolve_relocs(relocs, labels);
2219 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2220 		 (unsigned int)(p - handle_tlbl));
2221 
2222 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_size);
2223 }
2224 
2225 static void build_r4000_tlb_store_handler(void)
2226 {
2227 	u32 *p = handle_tlbs;
2228 	const int handle_tlbs_size = handle_tlbs_end - handle_tlbs;
2229 	struct uasm_label *l = labels;
2230 	struct uasm_reloc *r = relocs;
2231 	struct work_registers wr;
2232 
2233 	memset(handle_tlbs, 0, handle_tlbs_size * sizeof(handle_tlbs[0]));
2234 	memset(labels, 0, sizeof(labels));
2235 	memset(relocs, 0, sizeof(relocs));
2236 
2237 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2238 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2239 	if (m4kc_tlbp_war())
2240 		build_tlb_probe_entry(&p);
2241 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2242 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2243 
2244 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2245 	/*
2246 	 * This is the entry point when
2247 	 * build_r4000_tlbchange_handler_head spots a huge page.
2248 	 */
2249 	uasm_l_tlb_huge_update(&l, p);
2250 	iPTE_LW(&p, wr.r1, wr.r2);
2251 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2252 	build_tlb_probe_entry(&p);
2253 	uasm_i_ori(&p, wr.r1, wr.r1,
2254 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2255 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2256 #endif
2257 
2258 	uasm_l_nopage_tlbs(&l, p);
2259 	build_restore_work_registers(&p);
2260 #ifdef CONFIG_CPU_MICROMIPS
2261 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2262 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2263 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2264 		uasm_i_jr(&p, K0);
2265 	} else
2266 #endif
2267 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2268 	uasm_i_nop(&p);
2269 
2270 	if (p >= handle_tlbs_end)
2271 		panic("TLB store handler fastpath space exceeded");
2272 
2273 	uasm_resolve_relocs(relocs, labels);
2274 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2275 		 (unsigned int)(p - handle_tlbs));
2276 
2277 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_size);
2278 }
2279 
2280 static void build_r4000_tlb_modify_handler(void)
2281 {
2282 	u32 *p = handle_tlbm;
2283 	const int handle_tlbm_size = handle_tlbm_end - handle_tlbm;
2284 	struct uasm_label *l = labels;
2285 	struct uasm_reloc *r = relocs;
2286 	struct work_registers wr;
2287 
2288 	memset(handle_tlbm, 0, handle_tlbm_size * sizeof(handle_tlbm[0]));
2289 	memset(labels, 0, sizeof(labels));
2290 	memset(relocs, 0, sizeof(relocs));
2291 
2292 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2293 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2294 	if (m4kc_tlbp_war())
2295 		build_tlb_probe_entry(&p);
2296 	/* Present and writable bits set, set accessed and dirty bits. */
2297 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2298 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2299 
2300 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2301 	/*
2302 	 * This is the entry point when
2303 	 * build_r4000_tlbchange_handler_head spots a huge page.
2304 	 */
2305 	uasm_l_tlb_huge_update(&l, p);
2306 	iPTE_LW(&p, wr.r1, wr.r2);
2307 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2308 	build_tlb_probe_entry(&p);
2309 	uasm_i_ori(&p, wr.r1, wr.r1,
2310 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2311 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2312 #endif
2313 
2314 	uasm_l_nopage_tlbm(&l, p);
2315 	build_restore_work_registers(&p);
2316 #ifdef CONFIG_CPU_MICROMIPS
2317 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2318 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2319 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2320 		uasm_i_jr(&p, K0);
2321 	} else
2322 #endif
2323 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2324 	uasm_i_nop(&p);
2325 
2326 	if (p >= handle_tlbm_end)
2327 		panic("TLB modify handler fastpath space exceeded");
2328 
2329 	uasm_resolve_relocs(relocs, labels);
2330 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2331 		 (unsigned int)(p - handle_tlbm));
2332 
2333 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_size);
2334 }
2335 
2336 static void flush_tlb_handlers(void)
2337 {
2338 	local_flush_icache_range((unsigned long)handle_tlbl,
2339 			   (unsigned long)handle_tlbl_end);
2340 	local_flush_icache_range((unsigned long)handle_tlbs,
2341 			   (unsigned long)handle_tlbs_end);
2342 	local_flush_icache_range((unsigned long)handle_tlbm,
2343 			   (unsigned long)handle_tlbm_end);
2344 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2345 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2346 }
2347 
2348 static void print_htw_config(void)
2349 {
2350 	unsigned long config;
2351 	unsigned int pwctl;
2352 	const int field = 2 * sizeof(unsigned long);
2353 
2354 	config = read_c0_pwfield();
2355 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2356 		field, config,
2357 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2358 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2359 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2360 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2361 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2362 
2363 	config = read_c0_pwsize();
2364 	pr_debug("PWSize  (0x%0*lx): PS: 0x%lx  GDW: 0x%02lx  UDW: 0x%02lx  MDW: 0x%02lx  PTW: 0x%02lx  PTEW: 0x%02lx\n",
2365 		field, config,
2366 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2367 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2368 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2369 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2370 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2371 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2372 
2373 	pwctl = read_c0_pwctl();
2374 	pr_debug("PWCtl   (0x%x): PWEn: 0x%x  XK: 0x%x  XS: 0x%x  XU: 0x%x  DPH: 0x%x  HugePg: 0x%x  Psn: 0x%x\n",
2375 		pwctl,
2376 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2377 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2378 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2379 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2380 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2381 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2382 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2383 }
2384 
2385 static void config_htw_params(void)
2386 {
2387 	unsigned long pwfield, pwsize, ptei;
2388 	unsigned int config;
2389 
2390 	/*
2391 	 * We are using 2-level page tables, so we only need to
2392 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2393 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2394 	 * write values less than 0xc in these fields because the entire
2395 	 * write will be dropped. As a result of which, we must preserve
2396 	 * the original reset values and overwrite only what we really want.
2397 	 */
2398 
2399 	pwfield = read_c0_pwfield();
2400 	/* re-initialize the GDI field */
2401 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2402 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2403 	/* re-initialize the PTI field including the even/odd bit */
2404 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2405 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2406 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2407 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2408 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2409 	}
2410 	/* Set the PTEI right shift */
2411 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2412 	pwfield |= ptei;
2413 	write_c0_pwfield(pwfield);
2414 	/* Check whether the PTEI value is supported */
2415 	back_to_back_c0_hazard();
2416 	pwfield = read_c0_pwfield();
2417 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2418 		!= ptei) {
2419 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2420 			ptei);
2421 		/*
2422 		 * Drop option to avoid HTW being enabled via another path
2423 		 * (eg htw_reset())
2424 		 */
2425 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2426 		return;
2427 	}
2428 
2429 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2430 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2431 	if (CONFIG_PGTABLE_LEVELS >= 3)
2432 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2433 
2434 	/* Set pointer size to size of directory pointers */
2435 	if (config_enabled(CONFIG_64BIT))
2436 		pwsize |= MIPS_PWSIZE_PS_MASK;
2437 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2438 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2439 			& MIPS_PWSIZE_PTEW_MASK;
2440 
2441 	write_c0_pwsize(pwsize);
2442 
2443 	/* Make sure everything is set before we enable the HTW */
2444 	back_to_back_c0_hazard();
2445 
2446 	/*
2447 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2448 	 * the pwctl fields.
2449 	 */
2450 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2451 	if (config_enabled(CONFIG_64BIT))
2452 		config |= MIPS_PWCTL_XU_MASK;
2453 	write_c0_pwctl(config);
2454 	pr_info("Hardware Page Table Walker enabled\n");
2455 
2456 	print_htw_config();
2457 }
2458 
2459 static void config_xpa_params(void)
2460 {
2461 #ifdef CONFIG_XPA
2462 	unsigned int pagegrain;
2463 
2464 	if (mips_xpa_disabled) {
2465 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2466 		return;
2467 	}
2468 
2469 	pagegrain = read_c0_pagegrain();
2470 	write_c0_pagegrain(pagegrain | PG_ELPA);
2471 	back_to_back_c0_hazard();
2472 	pagegrain = read_c0_pagegrain();
2473 
2474 	if (pagegrain & PG_ELPA)
2475 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2476 	else
2477 		panic("Extended Physical Addressing (XPA) disabled");
2478 #endif
2479 }
2480 
2481 static void check_pabits(void)
2482 {
2483 	unsigned long entry;
2484 	unsigned pabits, fillbits;
2485 
2486 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2487 		/*
2488 		 * We'll only be making use of the fact that we can rotate bits
2489 		 * into the fill if the CPU supports RIXI, so don't bother
2490 		 * probing this for CPUs which don't.
2491 		 */
2492 		return;
2493 	}
2494 
2495 	write_c0_entrylo0(~0ul);
2496 	back_to_back_c0_hazard();
2497 	entry = read_c0_entrylo0();
2498 
2499 	/* clear all non-PFN bits */
2500 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2501 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2502 
2503 	/* find a lower bound on PABITS, and upper bound on fill bits */
2504 	pabits = fls_long(entry) + 6;
2505 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2506 
2507 	/* minus the RI & XI bits */
2508 	fillbits -= min_t(unsigned, fillbits, 2);
2509 
2510 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2511 		fill_includes_sw_bits = true;
2512 
2513 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2514 }
2515 
2516 void build_tlb_refill_handler(void)
2517 {
2518 	/*
2519 	 * The refill handler is generated per-CPU, multi-node systems
2520 	 * may have local storage for it. The other handlers are only
2521 	 * needed once.
2522 	 */
2523 	static int run_once = 0;
2524 
2525 	if (config_enabled(CONFIG_XPA) && !cpu_has_rixi)
2526 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2527 
2528 	output_pgtable_bits_defines();
2529 	check_pabits();
2530 
2531 #ifdef CONFIG_64BIT
2532 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2533 #endif
2534 
2535 	switch (current_cpu_type()) {
2536 	case CPU_R2000:
2537 	case CPU_R3000:
2538 	case CPU_R3000A:
2539 	case CPU_R3081E:
2540 	case CPU_TX3912:
2541 	case CPU_TX3922:
2542 	case CPU_TX3927:
2543 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2544 		if (cpu_has_local_ebase)
2545 			build_r3000_tlb_refill_handler();
2546 		if (!run_once) {
2547 			if (!cpu_has_local_ebase)
2548 				build_r3000_tlb_refill_handler();
2549 			build_setup_pgd();
2550 			build_r3000_tlb_load_handler();
2551 			build_r3000_tlb_store_handler();
2552 			build_r3000_tlb_modify_handler();
2553 			flush_tlb_handlers();
2554 			run_once++;
2555 		}
2556 #else
2557 		panic("No R3000 TLB refill handler");
2558 #endif
2559 		break;
2560 
2561 	case CPU_R6000:
2562 	case CPU_R6000A:
2563 		panic("No R6000 TLB refill handler yet");
2564 		break;
2565 
2566 	case CPU_R8000:
2567 		panic("No R8000 TLB refill handler yet");
2568 		break;
2569 
2570 	default:
2571 		if (cpu_has_ldpte)
2572 			setup_pw();
2573 
2574 		if (!run_once) {
2575 			scratch_reg = allocate_kscratch();
2576 			build_setup_pgd();
2577 			build_r4000_tlb_load_handler();
2578 			build_r4000_tlb_store_handler();
2579 			build_r4000_tlb_modify_handler();
2580 			if (cpu_has_ldpte)
2581 				build_loongson3_tlb_refill_handler();
2582 			else if (!cpu_has_local_ebase)
2583 				build_r4000_tlb_refill_handler();
2584 			flush_tlb_handlers();
2585 			run_once++;
2586 		}
2587 		if (cpu_has_local_ebase)
2588 			build_r4000_tlb_refill_handler();
2589 		if (cpu_has_xpa)
2590 			config_xpa_params();
2591 		if (cpu_has_htw)
2592 			config_htw_params();
2593 	}
2594 }
2595