xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision 5b628549)
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/export.h>
26 #include <linux/kernel.h>
27 #include <linux/types.h>
28 #include <linux/smp.h>
29 #include <linux/string.h>
30 #include <linux/cache.h>
31 
32 #include <asm/cacheflush.h>
33 #include <asm/cpu-type.h>
34 #include <asm/mmu_context.h>
35 #include <asm/pgtable.h>
36 #include <asm/war.h>
37 #include <asm/uasm.h>
38 #include <asm/setup.h>
39 #include <asm/tlbex.h>
40 
41 static int mips_xpa_disabled;
42 
43 static int __init xpa_disable(char *s)
44 {
45 	mips_xpa_disabled = 1;
46 
47 	return 1;
48 }
49 
50 __setup("noxpa", xpa_disable);
51 
52 /*
53  * TLB load/store/modify handlers.
54  *
55  * Only the fastpath gets synthesized at runtime, the slowpath for
56  * do_page_fault remains normal asm.
57  */
58 extern void tlb_do_page_fault_0(void);
59 extern void tlb_do_page_fault_1(void);
60 
61 struct work_registers {
62 	int r1;
63 	int r2;
64 	int r3;
65 };
66 
67 struct tlb_reg_save {
68 	unsigned long a;
69 	unsigned long b;
70 } ____cacheline_aligned_in_smp;
71 
72 static struct tlb_reg_save handler_reg_save[NR_CPUS];
73 
74 static inline int r45k_bvahwbug(void)
75 {
76 	/* XXX: We should probe for the presence of this bug, but we don't. */
77 	return 0;
78 }
79 
80 static inline int r4k_250MHZhwbug(void)
81 {
82 	/* XXX: We should probe for the presence of this bug, but we don't. */
83 	return 0;
84 }
85 
86 static inline int __maybe_unused bcm1250_m3_war(void)
87 {
88 	return BCM1250_M3_WAR;
89 }
90 
91 static inline int __maybe_unused r10000_llsc_war(void)
92 {
93 	return R10000_LLSC_WAR;
94 }
95 
96 static int use_bbit_insns(void)
97 {
98 	switch (current_cpu_type()) {
99 	case CPU_CAVIUM_OCTEON:
100 	case CPU_CAVIUM_OCTEON_PLUS:
101 	case CPU_CAVIUM_OCTEON2:
102 	case CPU_CAVIUM_OCTEON3:
103 		return 1;
104 	default:
105 		return 0;
106 	}
107 }
108 
109 static int use_lwx_insns(void)
110 {
111 	switch (current_cpu_type()) {
112 	case CPU_CAVIUM_OCTEON2:
113 	case CPU_CAVIUM_OCTEON3:
114 		return 1;
115 	default:
116 		return 0;
117 	}
118 }
119 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
120     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
121 static bool scratchpad_available(void)
122 {
123 	return true;
124 }
125 static int scratchpad_offset(int i)
126 {
127 	/*
128 	 * CVMSEG starts at address -32768 and extends for
129 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
130 	 */
131 	i += 1; /* Kernel use starts at the top and works down. */
132 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
133 }
134 #else
135 static bool scratchpad_available(void)
136 {
137 	return false;
138 }
139 static int scratchpad_offset(int i)
140 {
141 	BUG();
142 	/* Really unreachable, but evidently some GCC want this. */
143 	return 0;
144 }
145 #endif
146 /*
147  * Found by experiment: At least some revisions of the 4kc throw under
148  * some circumstances a machine check exception, triggered by invalid
149  * values in the index register.  Delaying the tlbp instruction until
150  * after the next branch,  plus adding an additional nop in front of
151  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
152  * why; it's not an issue caused by the core RTL.
153  *
154  */
155 static int m4kc_tlbp_war(void)
156 {
157 	return current_cpu_type() == CPU_4KC;
158 }
159 
160 /* Handle labels (which must be positive integers). */
161 enum label_id {
162 	label_second_part = 1,
163 	label_leave,
164 	label_vmalloc,
165 	label_vmalloc_done,
166 	label_tlbw_hazard_0,
167 	label_split = label_tlbw_hazard_0 + 8,
168 	label_tlbl_goaround1,
169 	label_tlbl_goaround2,
170 	label_nopage_tlbl,
171 	label_nopage_tlbs,
172 	label_nopage_tlbm,
173 	label_smp_pgtable_change,
174 	label_r3000_write_probe_fail,
175 	label_large_segbits_fault,
176 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
177 	label_tlb_huge_update,
178 #endif
179 };
180 
181 UASM_L_LA(_second_part)
182 UASM_L_LA(_leave)
183 UASM_L_LA(_vmalloc)
184 UASM_L_LA(_vmalloc_done)
185 /* _tlbw_hazard_x is handled differently.  */
186 UASM_L_LA(_split)
187 UASM_L_LA(_tlbl_goaround1)
188 UASM_L_LA(_tlbl_goaround2)
189 UASM_L_LA(_nopage_tlbl)
190 UASM_L_LA(_nopage_tlbs)
191 UASM_L_LA(_nopage_tlbm)
192 UASM_L_LA(_smp_pgtable_change)
193 UASM_L_LA(_r3000_write_probe_fail)
194 UASM_L_LA(_large_segbits_fault)
195 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
196 UASM_L_LA(_tlb_huge_update)
197 #endif
198 
199 static int hazard_instance;
200 
201 static void uasm_bgezl_hazard(u32 **p, struct uasm_reloc **r, int instance)
202 {
203 	switch (instance) {
204 	case 0 ... 7:
205 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard_0 + instance);
206 		return;
207 	default:
208 		BUG();
209 	}
210 }
211 
212 static void uasm_bgezl_label(struct uasm_label **l, u32 **p, int instance)
213 {
214 	switch (instance) {
215 	case 0 ... 7:
216 		uasm_build_label(l, *p, label_tlbw_hazard_0 + instance);
217 		break;
218 	default:
219 		BUG();
220 	}
221 }
222 
223 /*
224  * pgtable bits are assigned dynamically depending on processor feature
225  * and statically based on kernel configuration.  This spits out the actual
226  * values the kernel is using.	Required to make sense from disassembled
227  * TLB exception handlers.
228  */
229 static void output_pgtable_bits_defines(void)
230 {
231 #define pr_define(fmt, ...)					\
232 	pr_debug("#define " fmt, ##__VA_ARGS__)
233 
234 	pr_debug("#include <asm/asm.h>\n");
235 	pr_debug("#include <asm/regdef.h>\n");
236 	pr_debug("\n");
237 
238 	pr_define("_PAGE_PRESENT_SHIFT %d\n", _PAGE_PRESENT_SHIFT);
239 	pr_define("_PAGE_NO_READ_SHIFT %d\n", _PAGE_NO_READ_SHIFT);
240 	pr_define("_PAGE_WRITE_SHIFT %d\n", _PAGE_WRITE_SHIFT);
241 	pr_define("_PAGE_ACCESSED_SHIFT %d\n", _PAGE_ACCESSED_SHIFT);
242 	pr_define("_PAGE_MODIFIED_SHIFT %d\n", _PAGE_MODIFIED_SHIFT);
243 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
244 	pr_define("_PAGE_HUGE_SHIFT %d\n", _PAGE_HUGE_SHIFT);
245 #endif
246 #ifdef _PAGE_NO_EXEC_SHIFT
247 	if (cpu_has_rixi)
248 		pr_define("_PAGE_NO_EXEC_SHIFT %d\n", _PAGE_NO_EXEC_SHIFT);
249 #endif
250 	pr_define("_PAGE_GLOBAL_SHIFT %d\n", _PAGE_GLOBAL_SHIFT);
251 	pr_define("_PAGE_VALID_SHIFT %d\n", _PAGE_VALID_SHIFT);
252 	pr_define("_PAGE_DIRTY_SHIFT %d\n", _PAGE_DIRTY_SHIFT);
253 	pr_define("_PFN_SHIFT %d\n", _PFN_SHIFT);
254 	pr_debug("\n");
255 }
256 
257 static inline void dump_handler(const char *symbol, const void *start, const void *end)
258 {
259 	unsigned int count = (end - start) / sizeof(u32);
260 	const u32 *handler = start;
261 	int i;
262 
263 	pr_debug("LEAF(%s)\n", symbol);
264 
265 	pr_debug("\t.set push\n");
266 	pr_debug("\t.set noreorder\n");
267 
268 	for (i = 0; i < count; i++)
269 		pr_debug("\t.word\t0x%08x\t\t# %p\n", handler[i], &handler[i]);
270 
271 	pr_debug("\t.set\tpop\n");
272 
273 	pr_debug("\tEND(%s)\n", symbol);
274 }
275 
276 /* The only general purpose registers allowed in TLB handlers. */
277 #define K0		26
278 #define K1		27
279 
280 /* Some CP0 registers */
281 #define C0_INDEX	0, 0
282 #define C0_ENTRYLO0	2, 0
283 #define C0_TCBIND	2, 2
284 #define C0_ENTRYLO1	3, 0
285 #define C0_CONTEXT	4, 0
286 #define C0_PAGEMASK	5, 0
287 #define C0_PWBASE	5, 5
288 #define C0_PWFIELD	5, 6
289 #define C0_PWSIZE	5, 7
290 #define C0_PWCTL	6, 6
291 #define C0_BADVADDR	8, 0
292 #define C0_PGD		9, 7
293 #define C0_ENTRYHI	10, 0
294 #define C0_EPC		14, 0
295 #define C0_XCONTEXT	20, 0
296 
297 #ifdef CONFIG_64BIT
298 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
299 #else
300 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
301 #endif
302 
303 /* The worst case length of the handler is around 18 instructions for
304  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
305  * Maximum space available is 32 instructions for R3000 and 64
306  * instructions for R4000.
307  *
308  * We deliberately chose a buffer size of 128, so we won't scribble
309  * over anything important on overflow before we panic.
310  */
311 static u32 tlb_handler[128];
312 
313 /* simply assume worst case size for labels and relocs */
314 static struct uasm_label labels[128];
315 static struct uasm_reloc relocs[128];
316 
317 static int check_for_high_segbits;
318 static bool fill_includes_sw_bits;
319 
320 static unsigned int kscratch_used_mask;
321 
322 static inline int __maybe_unused c0_kscratch(void)
323 {
324 	switch (current_cpu_type()) {
325 	case CPU_XLP:
326 	case CPU_XLR:
327 		return 22;
328 	default:
329 		return 31;
330 	}
331 }
332 
333 static int allocate_kscratch(void)
334 {
335 	int r;
336 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
337 
338 	r = ffs(a);
339 
340 	if (r == 0)
341 		return -1;
342 
343 	r--; /* make it zero based */
344 
345 	kscratch_used_mask |= (1 << r);
346 
347 	return r;
348 }
349 
350 static int scratch_reg;
351 int pgd_reg;
352 EXPORT_SYMBOL_GPL(pgd_reg);
353 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
354 
355 static struct work_registers build_get_work_registers(u32 **p)
356 {
357 	struct work_registers r;
358 
359 	if (scratch_reg >= 0) {
360 		/* Save in CPU local C0_KScratch? */
361 		UASM_i_MTC0(p, 1, c0_kscratch(), scratch_reg);
362 		r.r1 = K0;
363 		r.r2 = K1;
364 		r.r3 = 1;
365 		return r;
366 	}
367 
368 	if (num_possible_cpus() > 1) {
369 		/* Get smp_processor_id */
370 		UASM_i_CPUID_MFC0(p, K0, SMP_CPUID_REG);
371 		UASM_i_SRL_SAFE(p, K0, K0, SMP_CPUID_REGSHIFT);
372 
373 		/* handler_reg_save index in K0 */
374 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
375 
376 		UASM_i_LA(p, K1, (long)&handler_reg_save);
377 		UASM_i_ADDU(p, K0, K0, K1);
378 	} else {
379 		UASM_i_LA(p, K0, (long)&handler_reg_save);
380 	}
381 	/* K0 now points to save area, save $1 and $2  */
382 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
383 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
384 
385 	r.r1 = K1;
386 	r.r2 = 1;
387 	r.r3 = 2;
388 	return r;
389 }
390 
391 static void build_restore_work_registers(u32 **p)
392 {
393 	if (scratch_reg >= 0) {
394 		UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
395 		return;
396 	}
397 	/* K0 already points to save area, restore $1 and $2  */
398 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
399 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
400 }
401 
402 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
403 
404 /*
405  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
406  * we cannot do r3000 under these circumstances.
407  *
408  * The R3000 TLB handler is simple.
409  */
410 static void build_r3000_tlb_refill_handler(void)
411 {
412 	long pgdc = (long)pgd_current;
413 	u32 *p;
414 
415 	memset(tlb_handler, 0, sizeof(tlb_handler));
416 	p = tlb_handler;
417 
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 */
433 	uasm_i_jr(&p, K1);
434 	uasm_i_rfe(&p); /* branch delay */
435 
436 	if (p > tlb_handler + 32)
437 		panic("TLB refill handler space exceeded");
438 
439 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
440 		 (unsigned int)(p - tlb_handler));
441 
442 	memcpy((void *)ebase, tlb_handler, 0x80);
443 	local_flush_icache_range(ebase, ebase + 0x80);
444 	dump_handler("r3000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x80));
445 }
446 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
447 
448 /*
449  * The R4000 TLB handler is much more complicated. We have two
450  * consecutive handler areas with 32 instructions space each.
451  * Since they aren't used at the same time, we can overflow in the
452  * other one.To keep things simple, we first assume linear space,
453  * then we relocate it to the final handler layout as needed.
454  */
455 static u32 final_handler[64];
456 
457 /*
458  * Hazards
459  *
460  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
461  * 2. A timing hazard exists for the TLBP instruction.
462  *
463  *	stalling_instruction
464  *	TLBP
465  *
466  * The JTLB is being read for the TLBP throughout the stall generated by the
467  * previous instruction. This is not really correct as the stalling instruction
468  * can modify the address used to access the JTLB.  The failure symptom is that
469  * the TLBP instruction will use an address created for the stalling instruction
470  * and not the address held in C0_ENHI and thus report the wrong results.
471  *
472  * The software work-around is to not allow the instruction preceding the TLBP
473  * to stall - make it an NOP or some other instruction guaranteed not to stall.
474  *
475  * Errata 2 will not be fixed.	This errata is also on the R5000.
476  *
477  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
478  */
479 static void __maybe_unused build_tlb_probe_entry(u32 **p)
480 {
481 	switch (current_cpu_type()) {
482 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
483 	case CPU_R4600:
484 	case CPU_R4700:
485 	case CPU_R5000:
486 	case CPU_NEVADA:
487 		uasm_i_nop(p);
488 		uasm_i_tlbp(p);
489 		break;
490 
491 	default:
492 		uasm_i_tlbp(p);
493 		break;
494 	}
495 }
496 
497 void build_tlb_write_entry(u32 **p, struct uasm_label **l,
498 			   struct uasm_reloc **r,
499 			   enum tlb_write_entry wmode)
500 {
501 	void(*tlbw)(u32 **) = NULL;
502 
503 	switch (wmode) {
504 	case tlb_random: tlbw = uasm_i_tlbwr; break;
505 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
506 	}
507 
508 	if (cpu_has_mips_r2_r6) {
509 		if (cpu_has_mips_r2_exec_hazard)
510 			uasm_i_ehb(p);
511 		tlbw(p);
512 		return;
513 	}
514 
515 	switch (current_cpu_type()) {
516 	case CPU_R4000PC:
517 	case CPU_R4000SC:
518 	case CPU_R4000MC:
519 	case CPU_R4400PC:
520 	case CPU_R4400SC:
521 	case CPU_R4400MC:
522 		/*
523 		 * This branch uses up a mtc0 hazard nop slot and saves
524 		 * two nops after the tlbw instruction.
525 		 */
526 		uasm_bgezl_hazard(p, r, hazard_instance);
527 		tlbw(p);
528 		uasm_bgezl_label(l, p, hazard_instance);
529 		hazard_instance++;
530 		uasm_i_nop(p);
531 		break;
532 
533 	case CPU_R4600:
534 	case CPU_R4700:
535 		uasm_i_nop(p);
536 		tlbw(p);
537 		uasm_i_nop(p);
538 		break;
539 
540 	case CPU_R5000:
541 	case CPU_NEVADA:
542 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
543 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
544 		tlbw(p);
545 		break;
546 
547 	case CPU_R4300:
548 	case CPU_5KC:
549 	case CPU_TX49XX:
550 	case CPU_PR4450:
551 	case CPU_XLR:
552 		uasm_i_nop(p);
553 		tlbw(p);
554 		break;
555 
556 	case CPU_R10000:
557 	case CPU_R12000:
558 	case CPU_R14000:
559 	case CPU_R16000:
560 	case CPU_4KC:
561 	case CPU_4KEC:
562 	case CPU_M14KC:
563 	case CPU_M14KEC:
564 	case CPU_SB1:
565 	case CPU_SB1A:
566 	case CPU_4KSC:
567 	case CPU_20KC:
568 	case CPU_25KF:
569 	case CPU_BMIPS32:
570 	case CPU_BMIPS3300:
571 	case CPU_BMIPS4350:
572 	case CPU_BMIPS4380:
573 	case CPU_BMIPS5000:
574 	case CPU_LOONGSON2:
575 	case CPU_LOONGSON3:
576 	case CPU_R5500:
577 		if (m4kc_tlbp_war())
578 			uasm_i_nop(p);
579 		/* fall through */
580 	case CPU_ALCHEMY:
581 		tlbw(p);
582 		break;
583 
584 	case CPU_RM7000:
585 		uasm_i_nop(p);
586 		uasm_i_nop(p);
587 		uasm_i_nop(p);
588 		uasm_i_nop(p);
589 		tlbw(p);
590 		break;
591 
592 	case CPU_VR4111:
593 	case CPU_VR4121:
594 	case CPU_VR4122:
595 	case CPU_VR4181:
596 	case CPU_VR4181A:
597 		uasm_i_nop(p);
598 		uasm_i_nop(p);
599 		tlbw(p);
600 		uasm_i_nop(p);
601 		uasm_i_nop(p);
602 		break;
603 
604 	case CPU_VR4131:
605 	case CPU_VR4133:
606 	case CPU_R5432:
607 		uasm_i_nop(p);
608 		uasm_i_nop(p);
609 		tlbw(p);
610 		break;
611 
612 	case CPU_JZRISC:
613 		tlbw(p);
614 		uasm_i_nop(p);
615 		break;
616 
617 	default:
618 		panic("No TLB refill handler yet (CPU type: %d)",
619 		      current_cpu_type());
620 		break;
621 	}
622 }
623 EXPORT_SYMBOL_GPL(build_tlb_write_entry);
624 
625 static __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
626 							unsigned int reg)
627 {
628 	if (_PAGE_GLOBAL_SHIFT == 0) {
629 		/* pte_t is already in EntryLo format */
630 		return;
631 	}
632 
633 	if (cpu_has_rixi && _PAGE_NO_EXEC) {
634 		if (fill_includes_sw_bits) {
635 			UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
636 		} else {
637 			UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
638 			UASM_i_ROTR(p, reg, reg,
639 				    ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
640 		}
641 	} else {
642 #ifdef CONFIG_PHYS_ADDR_T_64BIT
643 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
644 #else
645 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
646 #endif
647 	}
648 }
649 
650 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
651 
652 static void build_restore_pagemask(u32 **p, struct uasm_reloc **r,
653 				   unsigned int tmp, enum label_id lid,
654 				   int restore_scratch)
655 {
656 	if (restore_scratch) {
657 		/* Reset default page size */
658 		if (PM_DEFAULT_MASK >> 16) {
659 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
660 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
661 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
662 			uasm_il_b(p, r, lid);
663 		} else if (PM_DEFAULT_MASK) {
664 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
665 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
666 			uasm_il_b(p, r, lid);
667 		} else {
668 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
669 			uasm_il_b(p, r, lid);
670 		}
671 		if (scratch_reg >= 0)
672 			UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
673 		else
674 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
675 	} else {
676 		/* Reset default page size */
677 		if (PM_DEFAULT_MASK >> 16) {
678 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
679 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
680 			uasm_il_b(p, r, lid);
681 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
682 		} else if (PM_DEFAULT_MASK) {
683 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
684 			uasm_il_b(p, r, lid);
685 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
686 		} else {
687 			uasm_il_b(p, r, lid);
688 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
689 		}
690 	}
691 }
692 
693 static void build_huge_tlb_write_entry(u32 **p, struct uasm_label **l,
694 				       struct uasm_reloc **r,
695 				       unsigned int tmp,
696 				       enum tlb_write_entry wmode,
697 				       int restore_scratch)
698 {
699 	/* Set huge page tlb entry size */
700 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
701 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
702 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
703 
704 	build_tlb_write_entry(p, l, r, wmode);
705 
706 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
707 }
708 
709 /*
710  * Check if Huge PTE is present, if so then jump to LABEL.
711  */
712 static void
713 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
714 		  unsigned int pmd, int lid)
715 {
716 	UASM_i_LW(p, tmp, 0, pmd);
717 	if (use_bbit_insns()) {
718 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
719 	} else {
720 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
721 		uasm_il_bnez(p, r, tmp, lid);
722 	}
723 }
724 
725 static void build_huge_update_entries(u32 **p, unsigned int pte,
726 				      unsigned int tmp)
727 {
728 	int small_sequence;
729 
730 	/*
731 	 * A huge PTE describes an area the size of the
732 	 * configured huge page size. This is twice the
733 	 * of the large TLB entry size we intend to use.
734 	 * A TLB entry half the size of the configured
735 	 * huge page size is configured into entrylo0
736 	 * and entrylo1 to cover the contiguous huge PTE
737 	 * address space.
738 	 */
739 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
740 
741 	/* We can clobber tmp.	It isn't used after this.*/
742 	if (!small_sequence)
743 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
744 
745 	build_convert_pte_to_entrylo(p, pte);
746 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
747 	/* convert to entrylo1 */
748 	if (small_sequence)
749 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
750 	else
751 		UASM_i_ADDU(p, pte, pte, tmp);
752 
753 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
754 }
755 
756 static void build_huge_handler_tail(u32 **p, struct uasm_reloc **r,
757 				    struct uasm_label **l,
758 				    unsigned int pte,
759 				    unsigned int ptr,
760 				    unsigned int flush)
761 {
762 #ifdef CONFIG_SMP
763 	UASM_i_SC(p, pte, 0, ptr);
764 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
765 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
766 #else
767 	UASM_i_SW(p, pte, 0, ptr);
768 #endif
769 	if (cpu_has_ftlb && flush) {
770 		BUG_ON(!cpu_has_tlbinv);
771 
772 		UASM_i_MFC0(p, ptr, C0_ENTRYHI);
773 		uasm_i_ori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
774 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
775 		build_tlb_write_entry(p, l, r, tlb_indexed);
776 
777 		uasm_i_xori(p, ptr, ptr, MIPS_ENTRYHI_EHINV);
778 		UASM_i_MTC0(p, ptr, C0_ENTRYHI);
779 		build_huge_update_entries(p, pte, ptr);
780 		build_huge_tlb_write_entry(p, l, r, pte, tlb_random, 0);
781 
782 		return;
783 	}
784 
785 	build_huge_update_entries(p, pte, ptr);
786 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
787 }
788 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
789 
790 #ifdef CONFIG_64BIT
791 /*
792  * TMP and PTR are scratch.
793  * TMP will be clobbered, PTR will hold the pmd entry.
794  */
795 void build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
796 		      unsigned int tmp, unsigned int ptr)
797 {
798 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
799 	long pgdc = (long)pgd_current;
800 #endif
801 	/*
802 	 * The vmalloc handling is not in the hotpath.
803 	 */
804 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
805 
806 	if (check_for_high_segbits) {
807 		/*
808 		 * The kernel currently implicitely assumes that the
809 		 * MIPS SEGBITS parameter for the processor is
810 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
811 		 * allocate virtual addresses outside the maximum
812 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
813 		 * that doesn't prevent user code from accessing the
814 		 * higher xuseg addresses.  Here, we make sure that
815 		 * everything but the lower xuseg addresses goes down
816 		 * the module_alloc/vmalloc path.
817 		 */
818 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
819 		uasm_il_bnez(p, r, ptr, label_vmalloc);
820 	} else {
821 		uasm_il_bltz(p, r, tmp, label_vmalloc);
822 	}
823 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
824 
825 	if (pgd_reg != -1) {
826 		/* pgd is in pgd_reg */
827 		if (cpu_has_ldpte)
828 			UASM_i_MFC0(p, ptr, C0_PWBASE);
829 		else
830 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
831 	} else {
832 #if defined(CONFIG_MIPS_PGD_C0_CONTEXT)
833 		/*
834 		 * &pgd << 11 stored in CONTEXT [23..63].
835 		 */
836 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
837 
838 		/* Clear lower 23 bits of context. */
839 		uasm_i_dins(p, ptr, 0, 0, 23);
840 
841 		/* 1 0	1 0 1  << 6  xkphys cached */
842 		uasm_i_ori(p, ptr, ptr, 0x540);
843 		uasm_i_drotr(p, ptr, ptr, 11);
844 #elif defined(CONFIG_SMP)
845 		UASM_i_CPUID_MFC0(p, ptr, SMP_CPUID_REG);
846 		uasm_i_dsrl_safe(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
847 		UASM_i_LA_mostly(p, tmp, pgdc);
848 		uasm_i_daddu(p, ptr, ptr, tmp);
849 		uasm_i_dmfc0(p, tmp, C0_BADVADDR);
850 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
851 #else
852 		UASM_i_LA_mostly(p, ptr, pgdc);
853 		uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
854 #endif
855 	}
856 
857 	uasm_l_vmalloc_done(l, *p);
858 
859 	/* get pgd offset in bytes */
860 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
861 
862 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
863 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
864 #ifndef __PAGETABLE_PUD_FOLDED
865 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
866 	uasm_i_ld(p, ptr, 0, ptr); /* get pud pointer */
867 	uasm_i_dsrl_safe(p, tmp, tmp, PUD_SHIFT - 3); /* get pud offset in bytes */
868 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PUD - 1) << 3);
869 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pud offset */
870 #endif
871 #ifndef __PAGETABLE_PMD_FOLDED
872 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
873 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
874 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
875 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
876 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
877 #endif
878 }
879 EXPORT_SYMBOL_GPL(build_get_pmde64);
880 
881 /*
882  * BVADDR is the faulting address, PTR is scratch.
883  * PTR will hold the pgd for vmalloc.
884  */
885 static void
886 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
887 			unsigned int bvaddr, unsigned int ptr,
888 			enum vmalloc64_mode mode)
889 {
890 	long swpd = (long)swapper_pg_dir;
891 	int single_insn_swpd;
892 	int did_vmalloc_branch = 0;
893 
894 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
895 
896 	uasm_l_vmalloc(l, *p);
897 
898 	if (mode != not_refill && check_for_high_segbits) {
899 		if (single_insn_swpd) {
900 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
901 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
902 			did_vmalloc_branch = 1;
903 			/* fall through */
904 		} else {
905 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
906 		}
907 	}
908 	if (!did_vmalloc_branch) {
909 		if (single_insn_swpd) {
910 			uasm_il_b(p, r, label_vmalloc_done);
911 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
912 		} else {
913 			UASM_i_LA_mostly(p, ptr, swpd);
914 			uasm_il_b(p, r, label_vmalloc_done);
915 			if (uasm_in_compat_space_p(swpd))
916 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
917 			else
918 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
919 		}
920 	}
921 	if (mode != not_refill && check_for_high_segbits) {
922 		uasm_l_large_segbits_fault(l, *p);
923 		/*
924 		 * We get here if we are an xsseg address, or if we are
925 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
926 		 *
927 		 * Ignoring xsseg (assume disabled so would generate
928 		 * (address errors?), the only remaining possibility
929 		 * is the upper xuseg addresses.  On processors with
930 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
931 		 * addresses would have taken an address error. We try
932 		 * to mimic that here by taking a load/istream page
933 		 * fault.
934 		 */
935 		if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
936 			uasm_i_sync(p, 0);
937 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
938 		uasm_i_jr(p, ptr);
939 
940 		if (mode == refill_scratch) {
941 			if (scratch_reg >= 0)
942 				UASM_i_MFC0(p, 1, c0_kscratch(), scratch_reg);
943 			else
944 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
945 		} else {
946 			uasm_i_nop(p);
947 		}
948 	}
949 }
950 
951 #else /* !CONFIG_64BIT */
952 
953 /*
954  * TMP and PTR are scratch.
955  * TMP will be clobbered, PTR will hold the pgd entry.
956  */
957 void build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
958 {
959 	if (pgd_reg != -1) {
960 		/* pgd is in pgd_reg */
961 		uasm_i_mfc0(p, ptr, c0_kscratch(), pgd_reg);
962 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
963 	} else {
964 		long pgdc = (long)pgd_current;
965 
966 		/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
967 #ifdef CONFIG_SMP
968 		uasm_i_mfc0(p, ptr, SMP_CPUID_REG);
969 		UASM_i_LA_mostly(p, tmp, pgdc);
970 		uasm_i_srl(p, ptr, ptr, SMP_CPUID_PTRSHIFT);
971 		uasm_i_addu(p, ptr, tmp, ptr);
972 #else
973 		UASM_i_LA_mostly(p, ptr, pgdc);
974 #endif
975 		uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
976 		uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
977 	}
978 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
979 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
980 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
981 }
982 EXPORT_SYMBOL_GPL(build_get_pgde32);
983 
984 #endif /* !CONFIG_64BIT */
985 
986 static void build_adjust_context(u32 **p, unsigned int ctx)
987 {
988 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
989 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
990 
991 	switch (current_cpu_type()) {
992 	case CPU_VR41XX:
993 	case CPU_VR4111:
994 	case CPU_VR4121:
995 	case CPU_VR4122:
996 	case CPU_VR4131:
997 	case CPU_VR4181:
998 	case CPU_VR4181A:
999 	case CPU_VR4133:
1000 		shift += 2;
1001 		break;
1002 
1003 	default:
1004 		break;
1005 	}
1006 
1007 	if (shift)
1008 		UASM_i_SRL(p, ctx, ctx, shift);
1009 	uasm_i_andi(p, ctx, ctx, mask);
1010 }
1011 
1012 void build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
1013 {
1014 	/*
1015 	 * Bug workaround for the Nevada. It seems as if under certain
1016 	 * circumstances the move from cp0_context might produce a
1017 	 * bogus result when the mfc0 instruction and its consumer are
1018 	 * in a different cacheline or a load instruction, probably any
1019 	 * memory reference, is between them.
1020 	 */
1021 	switch (current_cpu_type()) {
1022 	case CPU_NEVADA:
1023 		UASM_i_LW(p, ptr, 0, ptr);
1024 		GET_CONTEXT(p, tmp); /* get context reg */
1025 		break;
1026 
1027 	default:
1028 		GET_CONTEXT(p, tmp); /* get context reg */
1029 		UASM_i_LW(p, ptr, 0, ptr);
1030 		break;
1031 	}
1032 
1033 	build_adjust_context(p, tmp);
1034 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1035 }
1036 EXPORT_SYMBOL_GPL(build_get_ptep);
1037 
1038 void build_update_entries(u32 **p, unsigned int tmp, unsigned int ptep)
1039 {
1040 	int pte_off_even = 0;
1041 	int pte_off_odd = sizeof(pte_t);
1042 
1043 #if defined(CONFIG_CPU_MIPS32) && defined(CONFIG_PHYS_ADDR_T_64BIT)
1044 	/* The low 32 bits of EntryLo is stored in pte_high */
1045 	pte_off_even += offsetof(pte_t, pte_high);
1046 	pte_off_odd += offsetof(pte_t, pte_high);
1047 #endif
1048 
1049 	if (IS_ENABLED(CONFIG_XPA)) {
1050 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* even pte */
1051 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1052 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0);
1053 
1054 		if (cpu_has_xpa && !mips_xpa_disabled) {
1055 			uasm_i_lw(p, tmp, 0, ptep);
1056 			uasm_i_ext(p, tmp, tmp, 0, 24);
1057 			uasm_i_mthc0(p, tmp, C0_ENTRYLO0);
1058 		}
1059 
1060 		uasm_i_lw(p, tmp, pte_off_odd, ptep); /* odd pte */
1061 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1062 		UASM_i_MTC0(p, tmp, C0_ENTRYLO1);
1063 
1064 		if (cpu_has_xpa && !mips_xpa_disabled) {
1065 			uasm_i_lw(p, tmp, sizeof(pte_t), ptep);
1066 			uasm_i_ext(p, tmp, tmp, 0, 24);
1067 			uasm_i_mthc0(p, tmp, C0_ENTRYLO1);
1068 		}
1069 		return;
1070 	}
1071 
1072 	UASM_i_LW(p, tmp, pte_off_even, ptep); /* get even pte */
1073 	UASM_i_LW(p, ptep, pte_off_odd, ptep); /* get odd pte */
1074 	if (r45k_bvahwbug())
1075 		build_tlb_probe_entry(p);
1076 	build_convert_pte_to_entrylo(p, tmp);
1077 	if (r4k_250MHZhwbug())
1078 		UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1079 	UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1080 	build_convert_pte_to_entrylo(p, ptep);
1081 	if (r45k_bvahwbug())
1082 		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 */
1086 }
1087 EXPORT_SYMBOL_GPL(build_update_entries);
1088 
1089 struct mips_huge_tlb_info {
1090 	int huge_pte;
1091 	int restore_scratch;
1092 	bool need_reload_pte;
1093 };
1094 
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)
1099 {
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 */
1104 
1105 	rv.huge_pte = scratch;
1106 	rv.restore_scratch = 0;
1107 	rv.need_reload_pte = false;
1108 
1109 	if (check_for_high_segbits) {
1110 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1111 
1112 		if (pgd_reg != -1)
1113 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1114 		else
1115 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
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 		uasm_i_dsrl_safe(p, scratch, tmp,
1123 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1124 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1125 
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);
1130 		}
1131 	} else {
1132 		if (pgd_reg != -1)
1133 			UASM_i_MFC0(p, ptr, c0_kscratch(), pgd_reg);
1134 		else
1135 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1136 
1137 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1138 
1139 		if (c0_scratch_reg >= 0)
1140 			UASM_i_MTC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1141 		else
1142 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1143 
1144 		if (pgd_reg == -1)
1145 			/* Clear lower 23 bits of context. */
1146 			uasm_i_dins(p, ptr, 0, 0, 23);
1147 
1148 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1149 	}
1150 
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);
1156 	}
1157 
1158 #ifdef __PAGETABLE_PMD_FOLDED
1159 #define LOC_PTEP scratch
1160 #else
1161 #define LOC_PTEP ptr
1162 #endif
1163 
1164 	if (!vmalloc_branch_delay_filled)
1165 		/* get pgd offset in bytes */
1166 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1167 
1168 	uasm_l_vmalloc_done(l, *p);
1169 
1170 	/*
1171 	 *			   tmp		ptr
1172 	 * fall-through case =	 badvaddr  *pgd_current
1173 	 * vmalloc case	     =	 badvaddr  swapper_pg_dir
1174 	 */
1175 
1176 	if (vmalloc_branch_delay_filled)
1177 		/* get pgd offset in bytes */
1178 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1179 
1180 #ifdef __PAGETABLE_PMD_FOLDED
1181 	GET_CONTEXT(p, tmp); /* get context reg */
1182 #endif
1183 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1184 
1185 	if (use_lwx_insns()) {
1186 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1187 	} else {
1188 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1189 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1190 	}
1191 
1192 #ifndef __PAGETABLE_PUD_FOLDED
1193 	/* get pud offset in bytes */
1194 	uasm_i_dsrl_safe(p, scratch, tmp, PUD_SHIFT - 3);
1195 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PUD - 1) << 3);
1196 
1197 	if (use_lwx_insns()) {
1198 		UASM_i_LWX(p, ptr, scratch, ptr);
1199 	} else {
1200 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1201 		UASM_i_LW(p, ptr, 0, ptr);
1202 	}
1203 	/* ptr contains a pointer to PMD entry */
1204 	/* tmp contains the address */
1205 #endif
1206 
1207 #ifndef __PAGETABLE_PMD_FOLDED
1208 	/* get pmd offset in bytes */
1209 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1210 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1211 	GET_CONTEXT(p, tmp); /* get context reg */
1212 
1213 	if (use_lwx_insns()) {
1214 		UASM_i_LWX(p, scratch, scratch, ptr);
1215 	} else {
1216 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1217 		UASM_i_LW(p, scratch, 0, ptr);
1218 	}
1219 #endif
1220 	/* Adjust the context during the load latency. */
1221 	build_adjust_context(p, tmp);
1222 
1223 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1224 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1225 	/*
1226 	 * The in the LWX case we don't want to do the load in the
1227 	 * delay slot.	It cannot issue in the same cycle and may be
1228 	 * speculative and unneeded.
1229 	 */
1230 	if (use_lwx_insns())
1231 		uasm_i_nop(p);
1232 #endif /* CONFIG_MIPS_HUGE_TLB_SUPPORT */
1233 
1234 
1235 	/* build_update_entries */
1236 	if (use_lwx_insns()) {
1237 		even = ptr;
1238 		odd = tmp;
1239 		UASM_i_LWX(p, even, scratch, tmp);
1240 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1241 		UASM_i_LWX(p, odd, scratch, tmp);
1242 	} else {
1243 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1244 		even = tmp;
1245 		odd = ptr;
1246 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1247 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1248 	}
1249 	if (cpu_has_rixi) {
1250 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1251 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1252 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1253 	} else {
1254 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1255 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1256 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1257 	}
1258 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1259 
1260 	if (c0_scratch_reg >= 0) {
1261 		UASM_i_MFC0(p, scratch, c0_kscratch(), c0_scratch_reg);
1262 		build_tlb_write_entry(p, l, r, tlb_random);
1263 		uasm_l_leave(l, *p);
1264 		rv.restore_scratch = 1;
1265 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1266 		build_tlb_write_entry(p, l, r, tlb_random);
1267 		uasm_l_leave(l, *p);
1268 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1269 	} else {
1270 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1271 		build_tlb_write_entry(p, l, r, tlb_random);
1272 		uasm_l_leave(l, *p);
1273 		rv.restore_scratch = 1;
1274 	}
1275 
1276 	uasm_i_eret(p); /* return from trap */
1277 
1278 	return rv;
1279 }
1280 
1281 /*
1282  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1283  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1284  * slots before the XTLB refill exception handler which belong to the
1285  * unused TLB refill exception.
1286  */
1287 #define MIPS64_REFILL_INSNS 32
1288 
1289 static void build_r4000_tlb_refill_handler(void)
1290 {
1291 	u32 *p = tlb_handler;
1292 	struct uasm_label *l = labels;
1293 	struct uasm_reloc *r = relocs;
1294 	u32 *f;
1295 	unsigned int final_len;
1296 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1297 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1298 
1299 	memset(tlb_handler, 0, sizeof(tlb_handler));
1300 	memset(labels, 0, sizeof(labels));
1301 	memset(relocs, 0, sizeof(relocs));
1302 	memset(final_handler, 0, sizeof(final_handler));
1303 
1304 	if (IS_ENABLED(CONFIG_64BIT) && (scratch_reg >= 0 || scratchpad_available()) && use_bbit_insns()) {
1305 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1306 							  scratch_reg);
1307 		vmalloc_mode = refill_scratch;
1308 	} else {
1309 		htlb_info.huge_pte = K0;
1310 		htlb_info.restore_scratch = 0;
1311 		htlb_info.need_reload_pte = true;
1312 		vmalloc_mode = refill_noscratch;
1313 		/*
1314 		 * create the plain linear handler
1315 		 */
1316 		if (bcm1250_m3_war()) {
1317 			unsigned int segbits = 44;
1318 
1319 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1320 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1321 			uasm_i_xor(&p, K0, K0, K1);
1322 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1323 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1324 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1325 			uasm_i_or(&p, K0, K0, K1);
1326 			uasm_il_bnez(&p, &r, K0, label_leave);
1327 			/* No need for uasm_i_nop */
1328 		}
1329 
1330 #ifdef CONFIG_64BIT
1331 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1332 #else
1333 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1334 #endif
1335 
1336 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1337 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1338 #endif
1339 
1340 		build_get_ptep(&p, K0, K1);
1341 		build_update_entries(&p, K0, K1);
1342 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1343 		uasm_l_leave(&l, p);
1344 		uasm_i_eret(&p); /* return from trap */
1345 	}
1346 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1347 	uasm_l_tlb_huge_update(&l, p);
1348 	if (htlb_info.need_reload_pte)
1349 		UASM_i_LW(&p, htlb_info.huge_pte, 0, K1);
1350 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1351 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1352 				   htlb_info.restore_scratch);
1353 #endif
1354 
1355 #ifdef CONFIG_64BIT
1356 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1357 #endif
1358 
1359 	/*
1360 	 * Overflow check: For the 64bit handler, we need at least one
1361 	 * free instruction slot for the wrap-around branch. In worst
1362 	 * case, if the intended insertion point is a delay slot, we
1363 	 * need three, with the second nop'ed and the third being
1364 	 * unused.
1365 	 */
1366 	switch (boot_cpu_type()) {
1367 	default:
1368 		if (sizeof(long) == 4) {
1369 	case CPU_LOONGSON2:
1370 		/* Loongson2 ebase is different than r4k, we have more space */
1371 			if ((p - tlb_handler) > 64)
1372 				panic("TLB refill handler space exceeded");
1373 			/*
1374 			 * Now fold the handler in the TLB refill handler space.
1375 			 */
1376 			f = final_handler;
1377 			/* Simplest case, just copy the handler. */
1378 			uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1379 			final_len = p - tlb_handler;
1380 			break;
1381 		} else {
1382 			if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1383 			    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1384 				&& uasm_insn_has_bdelay(relocs,
1385 							tlb_handler + MIPS64_REFILL_INSNS - 3)))
1386 				panic("TLB refill handler space exceeded");
1387 			/*
1388 			 * Now fold the handler in the TLB refill handler space.
1389 			 */
1390 			f = final_handler + MIPS64_REFILL_INSNS;
1391 			if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1392 				/* Just copy the handler. */
1393 				uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1394 				final_len = p - tlb_handler;
1395 			} else {
1396 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1397 				const enum label_id ls = label_tlb_huge_update;
1398 #else
1399 				const enum label_id ls = label_vmalloc;
1400 #endif
1401 				u32 *split;
1402 				int ov = 0;
1403 				int i;
1404 
1405 				for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1406 					;
1407 				BUG_ON(i == ARRAY_SIZE(labels));
1408 				split = labels[i].addr;
1409 
1410 				/*
1411 				 * See if we have overflown one way or the other.
1412 				 */
1413 				if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1414 				    split < p - MIPS64_REFILL_INSNS)
1415 					ov = 1;
1416 
1417 				if (ov) {
1418 					/*
1419 					 * Split two instructions before the end.  One
1420 					 * for the branch and one for the instruction
1421 					 * in the delay slot.
1422 					 */
1423 					split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1424 
1425 					/*
1426 					 * If the branch would fall in a delay slot,
1427 					 * we must back up an additional instruction
1428 					 * so that it is no longer in a delay slot.
1429 					 */
1430 					if (uasm_insn_has_bdelay(relocs, split - 1))
1431 						split--;
1432 				}
1433 				/* Copy first part of the handler. */
1434 				uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1435 				f += split - tlb_handler;
1436 
1437 				if (ov) {
1438 					/* Insert branch. */
1439 					uasm_l_split(&l, final_handler);
1440 					uasm_il_b(&f, &r, label_split);
1441 					if (uasm_insn_has_bdelay(relocs, split))
1442 						uasm_i_nop(&f);
1443 					else {
1444 						uasm_copy_handler(relocs, labels,
1445 								  split, split + 1, f);
1446 						uasm_move_labels(labels, f, f + 1, -1);
1447 						f++;
1448 						split++;
1449 					}
1450 				}
1451 
1452 				/* Copy the rest of the handler. */
1453 				uasm_copy_handler(relocs, labels, split, p, final_handler);
1454 				final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1455 					    (p - split);
1456 			}
1457 		}
1458 		break;
1459 	}
1460 
1461 	uasm_resolve_relocs(relocs, labels);
1462 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1463 		 final_len);
1464 
1465 	memcpy((void *)ebase, final_handler, 0x100);
1466 	local_flush_icache_range(ebase, ebase + 0x100);
1467 	dump_handler("r4000_tlb_refill", (u32 *)ebase, (u32 *)(ebase + 0x100));
1468 }
1469 
1470 static void setup_pw(void)
1471 {
1472 	unsigned long pgd_i, pgd_w;
1473 #ifndef __PAGETABLE_PMD_FOLDED
1474 	unsigned long pmd_i, pmd_w;
1475 #endif
1476 	unsigned long pt_i, pt_w;
1477 	unsigned long pte_i, pte_w;
1478 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1479 	unsigned long psn;
1480 
1481 	psn = ilog2(_PAGE_HUGE);     /* bit used to indicate huge page */
1482 #endif
1483 	pgd_i = PGDIR_SHIFT;  /* 1st level PGD */
1484 #ifndef __PAGETABLE_PMD_FOLDED
1485 	pgd_w = PGDIR_SHIFT - PMD_SHIFT + PGD_ORDER;
1486 
1487 	pmd_i = PMD_SHIFT;    /* 2nd level PMD */
1488 	pmd_w = PMD_SHIFT - PAGE_SHIFT;
1489 #else
1490 	pgd_w = PGDIR_SHIFT - PAGE_SHIFT + PGD_ORDER;
1491 #endif
1492 
1493 	pt_i  = PAGE_SHIFT;    /* 3rd level PTE */
1494 	pt_w  = PAGE_SHIFT - 3;
1495 
1496 	pte_i = ilog2(_PAGE_GLOBAL);
1497 	pte_w = 0;
1498 
1499 #ifndef __PAGETABLE_PMD_FOLDED
1500 	write_c0_pwfield(pgd_i << 24 | pmd_i << 12 | pt_i << 6 | pte_i);
1501 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pmd_w << 12 | pt_w << 6 | pte_w);
1502 #else
1503 	write_c0_pwfield(pgd_i << 24 | pt_i << 6 | pte_i);
1504 	write_c0_pwsize(1 << 30 | pgd_w << 24 | pt_w << 6 | pte_w);
1505 #endif
1506 
1507 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
1508 	write_c0_pwctl(1 << 6 | psn);
1509 #endif
1510 	write_c0_kpgd((long)swapper_pg_dir);
1511 	kscratch_used_mask |= (1 << 7); /* KScratch6 is used for KPGD */
1512 }
1513 
1514 static void build_loongson3_tlb_refill_handler(void)
1515 {
1516 	u32 *p = tlb_handler;
1517 	struct uasm_label *l = labels;
1518 	struct uasm_reloc *r = relocs;
1519 
1520 	memset(labels, 0, sizeof(labels));
1521 	memset(relocs, 0, sizeof(relocs));
1522 	memset(tlb_handler, 0, sizeof(tlb_handler));
1523 
1524 	if (check_for_high_segbits) {
1525 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1526 		uasm_i_dsrl_safe(&p, K1, K0, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1527 		uasm_il_beqz(&p, &r, K1, label_vmalloc);
1528 		uasm_i_nop(&p);
1529 
1530 		uasm_il_bgez(&p, &r, K0, label_large_segbits_fault);
1531 		uasm_i_nop(&p);
1532 		uasm_l_vmalloc(&l, p);
1533 	}
1534 
1535 	uasm_i_dmfc0(&p, K1, C0_PGD);
1536 
1537 	uasm_i_lddir(&p, K0, K1, 3);  /* global page dir */
1538 #ifndef __PAGETABLE_PMD_FOLDED
1539 	uasm_i_lddir(&p, K1, K0, 1);  /* middle page dir */
1540 #endif
1541 	uasm_i_ldpte(&p, K1, 0);      /* even */
1542 	uasm_i_ldpte(&p, K1, 1);      /* odd */
1543 	uasm_i_tlbwr(&p);
1544 
1545 	/* restore page mask */
1546 	if (PM_DEFAULT_MASK >> 16) {
1547 		uasm_i_lui(&p, K0, PM_DEFAULT_MASK >> 16);
1548 		uasm_i_ori(&p, K0, K0, PM_DEFAULT_MASK & 0xffff);
1549 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1550 	} else if (PM_DEFAULT_MASK) {
1551 		uasm_i_ori(&p, K0, 0, PM_DEFAULT_MASK);
1552 		uasm_i_mtc0(&p, K0, C0_PAGEMASK);
1553 	} else {
1554 		uasm_i_mtc0(&p, 0, C0_PAGEMASK);
1555 	}
1556 
1557 	uasm_i_eret(&p);
1558 
1559 	if (check_for_high_segbits) {
1560 		uasm_l_large_segbits_fault(&l, p);
1561 		UASM_i_LA(&p, K1, (unsigned long)tlb_do_page_fault_0);
1562 		uasm_i_jr(&p, K1);
1563 		uasm_i_nop(&p);
1564 	}
1565 
1566 	uasm_resolve_relocs(relocs, labels);
1567 	memcpy((void *)(ebase + 0x80), tlb_handler, 0x80);
1568 	local_flush_icache_range(ebase + 0x80, ebase + 0x100);
1569 	dump_handler("loongson3_tlb_refill",
1570 		     (u32 *)(ebase + 0x80), (u32 *)(ebase + 0x100));
1571 }
1572 
1573 static void build_setup_pgd(void)
1574 {
1575 	const int a0 = 4;
1576 	const int __maybe_unused a1 = 5;
1577 	const int __maybe_unused a2 = 6;
1578 	u32 *p = (u32 *)msk_isa16_mode((ulong)tlbmiss_handler_setup_pgd);
1579 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1580 	long pgdc = (long)pgd_current;
1581 #endif
1582 
1583 	memset(p, 0, tlbmiss_handler_setup_pgd_end - (char *)p);
1584 	memset(labels, 0, sizeof(labels));
1585 	memset(relocs, 0, sizeof(relocs));
1586 	pgd_reg = allocate_kscratch();
1587 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1588 	if (pgd_reg == -1) {
1589 		struct uasm_label *l = labels;
1590 		struct uasm_reloc *r = relocs;
1591 
1592 		/* PGD << 11 in c0_Context */
1593 		/*
1594 		 * If it is a ckseg0 address, convert to a physical
1595 		 * address.  Shifting right by 29 and adding 4 will
1596 		 * result in zero for these addresses.
1597 		 *
1598 		 */
1599 		UASM_i_SRA(&p, a1, a0, 29);
1600 		UASM_i_ADDIU(&p, a1, a1, 4);
1601 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1602 		uasm_i_nop(&p);
1603 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1604 		uasm_l_tlbl_goaround1(&l, p);
1605 		UASM_i_SLL(&p, a0, a0, 11);
1606 		uasm_i_jr(&p, 31);
1607 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1608 	} else {
1609 		/* PGD in c0_KScratch */
1610 		uasm_i_jr(&p, 31);
1611 		if (cpu_has_ldpte)
1612 			UASM_i_MTC0(&p, a0, C0_PWBASE);
1613 		else
1614 			UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1615 	}
1616 #else
1617 #ifdef CONFIG_SMP
1618 	/* Save PGD to pgd_current[smp_processor_id()] */
1619 	UASM_i_CPUID_MFC0(&p, a1, SMP_CPUID_REG);
1620 	UASM_i_SRL_SAFE(&p, a1, a1, SMP_CPUID_PTRSHIFT);
1621 	UASM_i_LA_mostly(&p, a2, pgdc);
1622 	UASM_i_ADDU(&p, a2, a2, a1);
1623 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1624 #else
1625 	UASM_i_LA_mostly(&p, a2, pgdc);
1626 	UASM_i_SW(&p, a0, uasm_rel_lo(pgdc), a2);
1627 #endif /* SMP */
1628 	uasm_i_jr(&p, 31);
1629 
1630 	/* if pgd_reg is allocated, save PGD also to scratch register */
1631 	if (pgd_reg != -1)
1632 		UASM_i_MTC0(&p, a0, c0_kscratch(), pgd_reg);
1633 	else
1634 		uasm_i_nop(&p);
1635 #endif
1636 	if (p >= (u32 *)tlbmiss_handler_setup_pgd_end)
1637 		panic("tlbmiss_handler_setup_pgd space exceeded");
1638 
1639 	uasm_resolve_relocs(relocs, labels);
1640 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1641 		 (unsigned int)(p - (u32 *)tlbmiss_handler_setup_pgd));
1642 
1643 	dump_handler("tlbmiss_handler", tlbmiss_handler_setup_pgd,
1644 					tlbmiss_handler_setup_pgd_end);
1645 }
1646 
1647 static void
1648 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1649 {
1650 #ifdef CONFIG_SMP
1651 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
1652 		uasm_i_sync(p, 0);
1653 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1654 	if (cpu_has_64bits)
1655 		uasm_i_lld(p, pte, 0, ptr);
1656 	else
1657 # endif
1658 		UASM_i_LL(p, pte, 0, ptr);
1659 #else
1660 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1661 	if (cpu_has_64bits)
1662 		uasm_i_ld(p, pte, 0, ptr);
1663 	else
1664 # endif
1665 		UASM_i_LW(p, pte, 0, ptr);
1666 #endif
1667 }
1668 
1669 static void
1670 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1671 	unsigned int mode, unsigned int scratch)
1672 {
1673 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1674 	unsigned int swmode = mode & ~hwmode;
1675 
1676 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_64bits) {
1677 		uasm_i_lui(p, scratch, swmode >> 16);
1678 		uasm_i_or(p, pte, pte, scratch);
1679 		BUG_ON(swmode & 0xffff);
1680 	} else {
1681 		uasm_i_ori(p, pte, pte, mode);
1682 	}
1683 
1684 #ifdef CONFIG_SMP
1685 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1686 	if (cpu_has_64bits)
1687 		uasm_i_scd(p, pte, 0, ptr);
1688 	else
1689 # endif
1690 		UASM_i_SC(p, pte, 0, ptr);
1691 
1692 	if (r10000_llsc_war())
1693 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1694 	else
1695 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1696 
1697 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1698 	if (!cpu_has_64bits) {
1699 		/* no uasm_i_nop needed */
1700 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1701 		uasm_i_ori(p, pte, pte, hwmode);
1702 		BUG_ON(hwmode & ~0xffff);
1703 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1704 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1705 		/* no uasm_i_nop needed */
1706 		uasm_i_lw(p, pte, 0, ptr);
1707 	} else
1708 		uasm_i_nop(p);
1709 # else
1710 	uasm_i_nop(p);
1711 # endif
1712 #else
1713 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1714 	if (cpu_has_64bits)
1715 		uasm_i_sd(p, pte, 0, ptr);
1716 	else
1717 # endif
1718 		UASM_i_SW(p, pte, 0, ptr);
1719 
1720 # ifdef CONFIG_PHYS_ADDR_T_64BIT
1721 	if (!cpu_has_64bits) {
1722 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1723 		uasm_i_ori(p, pte, pte, hwmode);
1724 		BUG_ON(hwmode & ~0xffff);
1725 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1726 		uasm_i_lw(p, pte, 0, ptr);
1727 	}
1728 # endif
1729 #endif
1730 }
1731 
1732 /*
1733  * Check if PTE is present, if not then jump to LABEL. PTR points to
1734  * the page table where this PTE is located, PTE will be re-loaded
1735  * with it's original value.
1736  */
1737 static void
1738 build_pte_present(u32 **p, struct uasm_reloc **r,
1739 		  int pte, int ptr, int scratch, enum label_id lid)
1740 {
1741 	int t = scratch >= 0 ? scratch : pte;
1742 	int cur = pte;
1743 
1744 	if (cpu_has_rixi) {
1745 		if (use_bbit_insns()) {
1746 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1747 			uasm_i_nop(p);
1748 		} else {
1749 			if (_PAGE_PRESENT_SHIFT) {
1750 				uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1751 				cur = t;
1752 			}
1753 			uasm_i_andi(p, t, cur, 1);
1754 			uasm_il_beqz(p, r, t, lid);
1755 			if (pte == t)
1756 				/* You lose the SMP race :-(*/
1757 				iPTE_LW(p, pte, ptr);
1758 		}
1759 	} else {
1760 		if (_PAGE_PRESENT_SHIFT) {
1761 			uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1762 			cur = t;
1763 		}
1764 		uasm_i_andi(p, t, cur,
1765 			(_PAGE_PRESENT | _PAGE_NO_READ) >> _PAGE_PRESENT_SHIFT);
1766 		uasm_i_xori(p, t, t, _PAGE_PRESENT >> _PAGE_PRESENT_SHIFT);
1767 		uasm_il_bnez(p, r, t, lid);
1768 		if (pte == t)
1769 			/* You lose the SMP race :-(*/
1770 			iPTE_LW(p, pte, ptr);
1771 	}
1772 }
1773 
1774 /* Make PTE valid, store result in PTR. */
1775 static void
1776 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1777 		 unsigned int ptr, unsigned int scratch)
1778 {
1779 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1780 
1781 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1782 }
1783 
1784 /*
1785  * Check if PTE can be written to, if not branch to LABEL. Regardless
1786  * restore PTE with value from PTR when done.
1787  */
1788 static void
1789 build_pte_writable(u32 **p, struct uasm_reloc **r,
1790 		   unsigned int pte, unsigned int ptr, int scratch,
1791 		   enum label_id lid)
1792 {
1793 	int t = scratch >= 0 ? scratch : pte;
1794 	int cur = pte;
1795 
1796 	if (_PAGE_PRESENT_SHIFT) {
1797 		uasm_i_srl(p, t, cur, _PAGE_PRESENT_SHIFT);
1798 		cur = t;
1799 	}
1800 	uasm_i_andi(p, t, cur,
1801 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1802 	uasm_i_xori(p, t, t,
1803 		    (_PAGE_PRESENT | _PAGE_WRITE) >> _PAGE_PRESENT_SHIFT);
1804 	uasm_il_bnez(p, r, t, lid);
1805 	if (pte == t)
1806 		/* You lose the SMP race :-(*/
1807 		iPTE_LW(p, pte, ptr);
1808 	else
1809 		uasm_i_nop(p);
1810 }
1811 
1812 /* Make PTE writable, update software status bits as well, then store
1813  * at PTR.
1814  */
1815 static void
1816 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1817 		 unsigned int ptr, unsigned int scratch)
1818 {
1819 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1820 			     | _PAGE_DIRTY);
1821 
1822 	iPTE_SW(p, r, pte, ptr, mode, scratch);
1823 }
1824 
1825 /*
1826  * Check if PTE can be modified, if not branch to LABEL. Regardless
1827  * restore PTE with value from PTR when done.
1828  */
1829 static void
1830 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1831 		     unsigned int pte, unsigned int ptr, int scratch,
1832 		     enum label_id lid)
1833 {
1834 	if (use_bbit_insns()) {
1835 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1836 		uasm_i_nop(p);
1837 	} else {
1838 		int t = scratch >= 0 ? scratch : pte;
1839 		uasm_i_srl(p, t, pte, _PAGE_WRITE_SHIFT);
1840 		uasm_i_andi(p, t, t, 1);
1841 		uasm_il_beqz(p, r, t, lid);
1842 		if (pte == t)
1843 			/* You lose the SMP race :-(*/
1844 			iPTE_LW(p, pte, ptr);
1845 	}
1846 }
1847 
1848 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1849 
1850 
1851 /*
1852  * R3000 style TLB load/store/modify handlers.
1853  */
1854 
1855 /*
1856  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1857  * Then it returns.
1858  */
1859 static void
1860 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1861 {
1862 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1863 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1864 	uasm_i_tlbwi(p);
1865 	uasm_i_jr(p, tmp);
1866 	uasm_i_rfe(p); /* branch delay */
1867 }
1868 
1869 /*
1870  * This places the pte into ENTRYLO0 and writes it with tlbwi
1871  * or tlbwr as appropriate.  This is because the index register
1872  * may have the probe fail bit set as a result of a trap on a
1873  * kseg2 access, i.e. without refill.  Then it returns.
1874  */
1875 static void
1876 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1877 			     struct uasm_reloc **r, unsigned int pte,
1878 			     unsigned int tmp)
1879 {
1880 	uasm_i_mfc0(p, tmp, C0_INDEX);
1881 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1882 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1883 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1884 	uasm_i_tlbwi(p); /* cp0 delay */
1885 	uasm_i_jr(p, tmp);
1886 	uasm_i_rfe(p); /* branch delay */
1887 	uasm_l_r3000_write_probe_fail(l, *p);
1888 	uasm_i_tlbwr(p); /* cp0 delay */
1889 	uasm_i_jr(p, tmp);
1890 	uasm_i_rfe(p); /* branch delay */
1891 }
1892 
1893 static void
1894 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1895 				   unsigned int ptr)
1896 {
1897 	long pgdc = (long)pgd_current;
1898 
1899 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1900 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1901 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1902 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1903 	uasm_i_sll(p, pte, pte, 2);
1904 	uasm_i_addu(p, ptr, ptr, pte);
1905 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1906 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1907 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1908 	uasm_i_addu(p, ptr, ptr, pte);
1909 	uasm_i_lw(p, pte, 0, ptr);
1910 	uasm_i_tlbp(p); /* load delay */
1911 }
1912 
1913 static void build_r3000_tlb_load_handler(void)
1914 {
1915 	u32 *p = (u32 *)handle_tlbl;
1916 	struct uasm_label *l = labels;
1917 	struct uasm_reloc *r = relocs;
1918 
1919 	memset(p, 0, handle_tlbl_end - (char *)p);
1920 	memset(labels, 0, sizeof(labels));
1921 	memset(relocs, 0, sizeof(relocs));
1922 
1923 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1924 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1925 	uasm_i_nop(&p); /* load delay */
1926 	build_make_valid(&p, &r, K0, K1, -1);
1927 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1928 
1929 	uasm_l_nopage_tlbl(&l, p);
1930 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1931 	uasm_i_nop(&p);
1932 
1933 	if (p >= (u32 *)handle_tlbl_end)
1934 		panic("TLB load handler fastpath space exceeded");
1935 
1936 	uasm_resolve_relocs(relocs, labels);
1937 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1938 		 (unsigned int)(p - (u32 *)handle_tlbl));
1939 
1940 	dump_handler("r3000_tlb_load", handle_tlbl, handle_tlbl_end);
1941 }
1942 
1943 static void build_r3000_tlb_store_handler(void)
1944 {
1945 	u32 *p = (u32 *)handle_tlbs;
1946 	struct uasm_label *l = labels;
1947 	struct uasm_reloc *r = relocs;
1948 
1949 	memset(p, 0, handle_tlbs_end - (char *)p);
1950 	memset(labels, 0, sizeof(labels));
1951 	memset(relocs, 0, sizeof(relocs));
1952 
1953 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1954 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1955 	uasm_i_nop(&p); /* load delay */
1956 	build_make_write(&p, &r, K0, K1, -1);
1957 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1958 
1959 	uasm_l_nopage_tlbs(&l, p);
1960 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1961 	uasm_i_nop(&p);
1962 
1963 	if (p >= (u32 *)handle_tlbs_end)
1964 		panic("TLB store handler fastpath space exceeded");
1965 
1966 	uasm_resolve_relocs(relocs, labels);
1967 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1968 		 (unsigned int)(p - (u32 *)handle_tlbs));
1969 
1970 	dump_handler("r3000_tlb_store", handle_tlbs, handle_tlbs_end);
1971 }
1972 
1973 static void build_r3000_tlb_modify_handler(void)
1974 {
1975 	u32 *p = (u32 *)handle_tlbm;
1976 	struct uasm_label *l = labels;
1977 	struct uasm_reloc *r = relocs;
1978 
1979 	memset(p, 0, handle_tlbm_end - (char *)p);
1980 	memset(labels, 0, sizeof(labels));
1981 	memset(relocs, 0, sizeof(relocs));
1982 
1983 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1984 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1985 	uasm_i_nop(&p); /* load delay */
1986 	build_make_write(&p, &r, K0, K1, -1);
1987 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1988 
1989 	uasm_l_nopage_tlbm(&l, p);
1990 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1991 	uasm_i_nop(&p);
1992 
1993 	if (p >= (u32 *)handle_tlbm_end)
1994 		panic("TLB modify handler fastpath space exceeded");
1995 
1996 	uasm_resolve_relocs(relocs, labels);
1997 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1998 		 (unsigned int)(p - (u32 *)handle_tlbm));
1999 
2000 	dump_handler("r3000_tlb_modify", handle_tlbm, handle_tlbm_end);
2001 }
2002 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
2003 
2004 static bool cpu_has_tlbex_tlbp_race(void)
2005 {
2006 	/*
2007 	 * When a Hardware Table Walker is running it can replace TLB entries
2008 	 * at any time, leading to a race between it & the CPU.
2009 	 */
2010 	if (cpu_has_htw)
2011 		return true;
2012 
2013 	/*
2014 	 * If the CPU shares FTLB RAM with its siblings then our entry may be
2015 	 * replaced at any time by a sibling performing a write to the FTLB.
2016 	 */
2017 	if (cpu_has_shared_ftlb_ram)
2018 		return true;
2019 
2020 	/* In all other cases there ought to be no race condition to handle */
2021 	return false;
2022 }
2023 
2024 /*
2025  * R4000 style TLB load/store/modify handlers.
2026  */
2027 static struct work_registers
2028 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
2029 				   struct uasm_reloc **r)
2030 {
2031 	struct work_registers wr = build_get_work_registers(p);
2032 
2033 #ifdef CONFIG_64BIT
2034 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
2035 #else
2036 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
2037 #endif
2038 
2039 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2040 	/*
2041 	 * For huge tlb entries, pmd doesn't contain an address but
2042 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
2043 	 * see if we need to jump to huge tlb processing.
2044 	 */
2045 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
2046 #endif
2047 
2048 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
2049 	UASM_i_LW(p, wr.r2, 0, wr.r2);
2050 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
2051 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
2052 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
2053 
2054 #ifdef CONFIG_SMP
2055 	uasm_l_smp_pgtable_change(l, *p);
2056 #endif
2057 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
2058 	if (!m4kc_tlbp_war()) {
2059 		build_tlb_probe_entry(p);
2060 		if (cpu_has_tlbex_tlbp_race()) {
2061 			/* race condition happens, leaving */
2062 			uasm_i_ehb(p);
2063 			uasm_i_mfc0(p, wr.r3, C0_INDEX);
2064 			uasm_il_bltz(p, r, wr.r3, label_leave);
2065 			uasm_i_nop(p);
2066 		}
2067 	}
2068 	return wr;
2069 }
2070 
2071 static void
2072 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
2073 				   struct uasm_reloc **r, unsigned int tmp,
2074 				   unsigned int ptr)
2075 {
2076 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
2077 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
2078 	build_update_entries(p, tmp, ptr);
2079 	build_tlb_write_entry(p, l, r, tlb_indexed);
2080 	uasm_l_leave(l, *p);
2081 	build_restore_work_registers(p);
2082 	uasm_i_eret(p); /* return from trap */
2083 
2084 #ifdef CONFIG_64BIT
2085 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
2086 #endif
2087 }
2088 
2089 static void build_r4000_tlb_load_handler(void)
2090 {
2091 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbl);
2092 	struct uasm_label *l = labels;
2093 	struct uasm_reloc *r = relocs;
2094 	struct work_registers wr;
2095 
2096 	memset(p, 0, handle_tlbl_end - (char *)p);
2097 	memset(labels, 0, sizeof(labels));
2098 	memset(relocs, 0, sizeof(relocs));
2099 
2100 	if (bcm1250_m3_war()) {
2101 		unsigned int segbits = 44;
2102 
2103 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
2104 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
2105 		uasm_i_xor(&p, K0, K0, K1);
2106 		uasm_i_dsrl_safe(&p, K1, K0, 62);
2107 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
2108 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
2109 		uasm_i_or(&p, K0, K0, K1);
2110 		uasm_il_bnez(&p, &r, K0, label_leave);
2111 		/* No need for uasm_i_nop */
2112 	}
2113 
2114 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2115 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2116 	if (m4kc_tlbp_war())
2117 		build_tlb_probe_entry(&p);
2118 
2119 	if (cpu_has_rixi && !cpu_has_rixiex) {
2120 		/*
2121 		 * If the page is not _PAGE_VALID, RI or XI could not
2122 		 * have triggered it.  Skip the expensive test..
2123 		 */
2124 		if (use_bbit_insns()) {
2125 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2126 				      label_tlbl_goaround1);
2127 		} else {
2128 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2129 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
2130 		}
2131 		uasm_i_nop(&p);
2132 
2133 		/*
2134 		 * Warn if something may race with us & replace the TLB entry
2135 		 * before we read it here. Everything with such races should
2136 		 * also have dedicated RiXi exception handlers, so this
2137 		 * shouldn't be hit.
2138 		 */
2139 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2140 
2141 		uasm_i_tlbr(&p);
2142 
2143 		switch (current_cpu_type()) {
2144 		default:
2145 			if (cpu_has_mips_r2_exec_hazard) {
2146 				uasm_i_ehb(&p);
2147 
2148 		case CPU_CAVIUM_OCTEON:
2149 		case CPU_CAVIUM_OCTEON_PLUS:
2150 		case CPU_CAVIUM_OCTEON2:
2151 				break;
2152 			}
2153 		}
2154 
2155 		/* Examine  entrylo 0 or 1 based on ptr. */
2156 		if (use_bbit_insns()) {
2157 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2158 		} else {
2159 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2160 			uasm_i_beqz(&p, wr.r3, 8);
2161 		}
2162 		/* load it in the delay slot*/
2163 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2164 		/* load it if ptr is odd */
2165 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2166 		/*
2167 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2168 		 * XI must have triggered it.
2169 		 */
2170 		if (use_bbit_insns()) {
2171 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
2172 			uasm_i_nop(&p);
2173 			uasm_l_tlbl_goaround1(&l, p);
2174 		} else {
2175 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2176 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
2177 			uasm_i_nop(&p);
2178 		}
2179 		uasm_l_tlbl_goaround1(&l, p);
2180 	}
2181 	build_make_valid(&p, &r, wr.r1, wr.r2, wr.r3);
2182 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2183 
2184 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2185 	/*
2186 	 * This is the entry point when build_r4000_tlbchange_handler_head
2187 	 * spots a huge page.
2188 	 */
2189 	uasm_l_tlb_huge_update(&l, p);
2190 	iPTE_LW(&p, wr.r1, wr.r2);
2191 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
2192 	build_tlb_probe_entry(&p);
2193 
2194 	if (cpu_has_rixi && !cpu_has_rixiex) {
2195 		/*
2196 		 * If the page is not _PAGE_VALID, RI or XI could not
2197 		 * have triggered it.  Skip the expensive test..
2198 		 */
2199 		if (use_bbit_insns()) {
2200 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
2201 				      label_tlbl_goaround2);
2202 		} else {
2203 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
2204 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2205 		}
2206 		uasm_i_nop(&p);
2207 
2208 		/*
2209 		 * Warn if something may race with us & replace the TLB entry
2210 		 * before we read it here. Everything with such races should
2211 		 * also have dedicated RiXi exception handlers, so this
2212 		 * shouldn't be hit.
2213 		 */
2214 		WARN(cpu_has_tlbex_tlbp_race(), "Unhandled race in RiXi path");
2215 
2216 		uasm_i_tlbr(&p);
2217 
2218 		switch (current_cpu_type()) {
2219 		default:
2220 			if (cpu_has_mips_r2_exec_hazard) {
2221 				uasm_i_ehb(&p);
2222 
2223 		case CPU_CAVIUM_OCTEON:
2224 		case CPU_CAVIUM_OCTEON_PLUS:
2225 		case CPU_CAVIUM_OCTEON2:
2226 				break;
2227 			}
2228 		}
2229 
2230 		/* Examine  entrylo 0 or 1 based on ptr. */
2231 		if (use_bbit_insns()) {
2232 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
2233 		} else {
2234 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
2235 			uasm_i_beqz(&p, wr.r3, 8);
2236 		}
2237 		/* load it in the delay slot*/
2238 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
2239 		/* load it if ptr is odd */
2240 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
2241 		/*
2242 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
2243 		 * XI must have triggered it.
2244 		 */
2245 		if (use_bbit_insns()) {
2246 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
2247 		} else {
2248 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
2249 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
2250 		}
2251 		if (PM_DEFAULT_MASK == 0)
2252 			uasm_i_nop(&p);
2253 		/*
2254 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
2255 		 * it is restored in build_huge_tlb_write_entry.
2256 		 */
2257 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
2258 
2259 		uasm_l_tlbl_goaround2(&l, p);
2260 	}
2261 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
2262 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2263 #endif
2264 
2265 	uasm_l_nopage_tlbl(&l, p);
2266 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2267 		uasm_i_sync(&p, 0);
2268 	build_restore_work_registers(&p);
2269 #ifdef CONFIG_CPU_MICROMIPS
2270 	if ((unsigned long)tlb_do_page_fault_0 & 1) {
2271 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_0));
2272 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_0));
2273 		uasm_i_jr(&p, K0);
2274 	} else
2275 #endif
2276 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2277 	uasm_i_nop(&p);
2278 
2279 	if (p >= (u32 *)handle_tlbl_end)
2280 		panic("TLB load handler fastpath space exceeded");
2281 
2282 	uasm_resolve_relocs(relocs, labels);
2283 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2284 		 (unsigned int)(p - (u32 *)handle_tlbl));
2285 
2286 	dump_handler("r4000_tlb_load", handle_tlbl, handle_tlbl_end);
2287 }
2288 
2289 static void build_r4000_tlb_store_handler(void)
2290 {
2291 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbs);
2292 	struct uasm_label *l = labels;
2293 	struct uasm_reloc *r = relocs;
2294 	struct work_registers wr;
2295 
2296 	memset(p, 0, handle_tlbs_end - (char *)p);
2297 	memset(labels, 0, sizeof(labels));
2298 	memset(relocs, 0, sizeof(relocs));
2299 
2300 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2301 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2302 	if (m4kc_tlbp_war())
2303 		build_tlb_probe_entry(&p);
2304 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2305 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2306 
2307 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2308 	/*
2309 	 * This is the entry point when
2310 	 * build_r4000_tlbchange_handler_head spots a huge page.
2311 	 */
2312 	uasm_l_tlb_huge_update(&l, p);
2313 	iPTE_LW(&p, wr.r1, wr.r2);
2314 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2315 	build_tlb_probe_entry(&p);
2316 	uasm_i_ori(&p, wr.r1, wr.r1,
2317 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2318 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 1);
2319 #endif
2320 
2321 	uasm_l_nopage_tlbs(&l, p);
2322 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2323 		uasm_i_sync(&p, 0);
2324 	build_restore_work_registers(&p);
2325 #ifdef CONFIG_CPU_MICROMIPS
2326 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2327 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2328 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2329 		uasm_i_jr(&p, K0);
2330 	} else
2331 #endif
2332 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2333 	uasm_i_nop(&p);
2334 
2335 	if (p >= (u32 *)handle_tlbs_end)
2336 		panic("TLB store handler fastpath space exceeded");
2337 
2338 	uasm_resolve_relocs(relocs, labels);
2339 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2340 		 (unsigned int)(p - (u32 *)handle_tlbs));
2341 
2342 	dump_handler("r4000_tlb_store", handle_tlbs, handle_tlbs_end);
2343 }
2344 
2345 static void build_r4000_tlb_modify_handler(void)
2346 {
2347 	u32 *p = (u32 *)msk_isa16_mode((ulong)handle_tlbm);
2348 	struct uasm_label *l = labels;
2349 	struct uasm_reloc *r = relocs;
2350 	struct work_registers wr;
2351 
2352 	memset(p, 0, handle_tlbm_end - (char *)p);
2353 	memset(labels, 0, sizeof(labels));
2354 	memset(relocs, 0, sizeof(relocs));
2355 
2356 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2357 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2358 	if (m4kc_tlbp_war())
2359 		build_tlb_probe_entry(&p);
2360 	/* Present and writable bits set, set accessed and dirty bits. */
2361 	build_make_write(&p, &r, wr.r1, wr.r2, wr.r3);
2362 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2363 
2364 #ifdef CONFIG_MIPS_HUGE_TLB_SUPPORT
2365 	/*
2366 	 * This is the entry point when
2367 	 * build_r4000_tlbchange_handler_head spots a huge page.
2368 	 */
2369 	uasm_l_tlb_huge_update(&l, p);
2370 	iPTE_LW(&p, wr.r1, wr.r2);
2371 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2372 	build_tlb_probe_entry(&p);
2373 	uasm_i_ori(&p, wr.r1, wr.r1,
2374 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2375 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2, 0);
2376 #endif
2377 
2378 	uasm_l_nopage_tlbm(&l, p);
2379 	if (IS_ENABLED(CONFIG_CPU_LOONGSON3_WORKAROUNDS))
2380 		uasm_i_sync(&p, 0);
2381 	build_restore_work_registers(&p);
2382 #ifdef CONFIG_CPU_MICROMIPS
2383 	if ((unsigned long)tlb_do_page_fault_1 & 1) {
2384 		uasm_i_lui(&p, K0, uasm_rel_hi((long)tlb_do_page_fault_1));
2385 		uasm_i_addiu(&p, K0, K0, uasm_rel_lo((long)tlb_do_page_fault_1));
2386 		uasm_i_jr(&p, K0);
2387 	} else
2388 #endif
2389 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2390 	uasm_i_nop(&p);
2391 
2392 	if (p >= (u32 *)handle_tlbm_end)
2393 		panic("TLB modify handler fastpath space exceeded");
2394 
2395 	uasm_resolve_relocs(relocs, labels);
2396 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2397 		 (unsigned int)(p - (u32 *)handle_tlbm));
2398 
2399 	dump_handler("r4000_tlb_modify", handle_tlbm, handle_tlbm_end);
2400 }
2401 
2402 static void flush_tlb_handlers(void)
2403 {
2404 	local_flush_icache_range((unsigned long)handle_tlbl,
2405 			   (unsigned long)handle_tlbl_end);
2406 	local_flush_icache_range((unsigned long)handle_tlbs,
2407 			   (unsigned long)handle_tlbs_end);
2408 	local_flush_icache_range((unsigned long)handle_tlbm,
2409 			   (unsigned long)handle_tlbm_end);
2410 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2411 			   (unsigned long)tlbmiss_handler_setup_pgd_end);
2412 }
2413 
2414 static void print_htw_config(void)
2415 {
2416 	unsigned long config;
2417 	unsigned int pwctl;
2418 	const int field = 2 * sizeof(unsigned long);
2419 
2420 	config = read_c0_pwfield();
2421 	pr_debug("PWField (0x%0*lx): GDI: 0x%02lx  UDI: 0x%02lx  MDI: 0x%02lx  PTI: 0x%02lx  PTEI: 0x%02lx\n",
2422 		field, config,
2423 		(config & MIPS_PWFIELD_GDI_MASK) >> MIPS_PWFIELD_GDI_SHIFT,
2424 		(config & MIPS_PWFIELD_UDI_MASK) >> MIPS_PWFIELD_UDI_SHIFT,
2425 		(config & MIPS_PWFIELD_MDI_MASK) >> MIPS_PWFIELD_MDI_SHIFT,
2426 		(config & MIPS_PWFIELD_PTI_MASK) >> MIPS_PWFIELD_PTI_SHIFT,
2427 		(config & MIPS_PWFIELD_PTEI_MASK) >> MIPS_PWFIELD_PTEI_SHIFT);
2428 
2429 	config = read_c0_pwsize();
2430 	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",
2431 		field, config,
2432 		(config & MIPS_PWSIZE_PS_MASK) >> MIPS_PWSIZE_PS_SHIFT,
2433 		(config & MIPS_PWSIZE_GDW_MASK) >> MIPS_PWSIZE_GDW_SHIFT,
2434 		(config & MIPS_PWSIZE_UDW_MASK) >> MIPS_PWSIZE_UDW_SHIFT,
2435 		(config & MIPS_PWSIZE_MDW_MASK) >> MIPS_PWSIZE_MDW_SHIFT,
2436 		(config & MIPS_PWSIZE_PTW_MASK) >> MIPS_PWSIZE_PTW_SHIFT,
2437 		(config & MIPS_PWSIZE_PTEW_MASK) >> MIPS_PWSIZE_PTEW_SHIFT);
2438 
2439 	pwctl = read_c0_pwctl();
2440 	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",
2441 		pwctl,
2442 		(pwctl & MIPS_PWCTL_PWEN_MASK) >> MIPS_PWCTL_PWEN_SHIFT,
2443 		(pwctl & MIPS_PWCTL_XK_MASK) >> MIPS_PWCTL_XK_SHIFT,
2444 		(pwctl & MIPS_PWCTL_XS_MASK) >> MIPS_PWCTL_XS_SHIFT,
2445 		(pwctl & MIPS_PWCTL_XU_MASK) >> MIPS_PWCTL_XU_SHIFT,
2446 		(pwctl & MIPS_PWCTL_DPH_MASK) >> MIPS_PWCTL_DPH_SHIFT,
2447 		(pwctl & MIPS_PWCTL_HUGEPG_MASK) >> MIPS_PWCTL_HUGEPG_SHIFT,
2448 		(pwctl & MIPS_PWCTL_PSN_MASK) >> MIPS_PWCTL_PSN_SHIFT);
2449 }
2450 
2451 static void config_htw_params(void)
2452 {
2453 	unsigned long pwfield, pwsize, ptei;
2454 	unsigned int config;
2455 
2456 	/*
2457 	 * We are using 2-level page tables, so we only need to
2458 	 * setup GDW and PTW appropriately. UDW and MDW will remain 0.
2459 	 * The default value of GDI/UDI/MDI/PTI is 0xc. It is illegal to
2460 	 * write values less than 0xc in these fields because the entire
2461 	 * write will be dropped. As a result of which, we must preserve
2462 	 * the original reset values and overwrite only what we really want.
2463 	 */
2464 
2465 	pwfield = read_c0_pwfield();
2466 	/* re-initialize the GDI field */
2467 	pwfield &= ~MIPS_PWFIELD_GDI_MASK;
2468 	pwfield |= PGDIR_SHIFT << MIPS_PWFIELD_GDI_SHIFT;
2469 	/* re-initialize the PTI field including the even/odd bit */
2470 	pwfield &= ~MIPS_PWFIELD_PTI_MASK;
2471 	pwfield |= PAGE_SHIFT << MIPS_PWFIELD_PTI_SHIFT;
2472 	if (CONFIG_PGTABLE_LEVELS >= 3) {
2473 		pwfield &= ~MIPS_PWFIELD_MDI_MASK;
2474 		pwfield |= PMD_SHIFT << MIPS_PWFIELD_MDI_SHIFT;
2475 	}
2476 	/* Set the PTEI right shift */
2477 	ptei = _PAGE_GLOBAL_SHIFT << MIPS_PWFIELD_PTEI_SHIFT;
2478 	pwfield |= ptei;
2479 	write_c0_pwfield(pwfield);
2480 	/* Check whether the PTEI value is supported */
2481 	back_to_back_c0_hazard();
2482 	pwfield = read_c0_pwfield();
2483 	if (((pwfield & MIPS_PWFIELD_PTEI_MASK) << MIPS_PWFIELD_PTEI_SHIFT)
2484 		!= ptei) {
2485 		pr_warn("Unsupported PTEI field value: 0x%lx. HTW will not be enabled",
2486 			ptei);
2487 		/*
2488 		 * Drop option to avoid HTW being enabled via another path
2489 		 * (eg htw_reset())
2490 		 */
2491 		current_cpu_data.options &= ~MIPS_CPU_HTW;
2492 		return;
2493 	}
2494 
2495 	pwsize = ilog2(PTRS_PER_PGD) << MIPS_PWSIZE_GDW_SHIFT;
2496 	pwsize |= ilog2(PTRS_PER_PTE) << MIPS_PWSIZE_PTW_SHIFT;
2497 	if (CONFIG_PGTABLE_LEVELS >= 3)
2498 		pwsize |= ilog2(PTRS_PER_PMD) << MIPS_PWSIZE_MDW_SHIFT;
2499 
2500 	/* Set pointer size to size of directory pointers */
2501 	if (IS_ENABLED(CONFIG_64BIT))
2502 		pwsize |= MIPS_PWSIZE_PS_MASK;
2503 	/* PTEs may be multiple pointers long (e.g. with XPA) */
2504 	pwsize |= ((PTE_T_LOG2 - PGD_T_LOG2) << MIPS_PWSIZE_PTEW_SHIFT)
2505 			& MIPS_PWSIZE_PTEW_MASK;
2506 
2507 	write_c0_pwsize(pwsize);
2508 
2509 	/* Make sure everything is set before we enable the HTW */
2510 	back_to_back_c0_hazard();
2511 
2512 	/*
2513 	 * Enable HTW (and only for XUSeg on 64-bit), and disable the rest of
2514 	 * the pwctl fields.
2515 	 */
2516 	config = 1 << MIPS_PWCTL_PWEN_SHIFT;
2517 	if (IS_ENABLED(CONFIG_64BIT))
2518 		config |= MIPS_PWCTL_XU_MASK;
2519 	write_c0_pwctl(config);
2520 	pr_info("Hardware Page Table Walker enabled\n");
2521 
2522 	print_htw_config();
2523 }
2524 
2525 static void config_xpa_params(void)
2526 {
2527 #ifdef CONFIG_XPA
2528 	unsigned int pagegrain;
2529 
2530 	if (mips_xpa_disabled) {
2531 		pr_info("Extended Physical Addressing (XPA) disabled\n");
2532 		return;
2533 	}
2534 
2535 	pagegrain = read_c0_pagegrain();
2536 	write_c0_pagegrain(pagegrain | PG_ELPA);
2537 	back_to_back_c0_hazard();
2538 	pagegrain = read_c0_pagegrain();
2539 
2540 	if (pagegrain & PG_ELPA)
2541 		pr_info("Extended Physical Addressing (XPA) enabled\n");
2542 	else
2543 		panic("Extended Physical Addressing (XPA) disabled");
2544 #endif
2545 }
2546 
2547 static void check_pabits(void)
2548 {
2549 	unsigned long entry;
2550 	unsigned pabits, fillbits;
2551 
2552 	if (!cpu_has_rixi || !_PAGE_NO_EXEC) {
2553 		/*
2554 		 * We'll only be making use of the fact that we can rotate bits
2555 		 * into the fill if the CPU supports RIXI, so don't bother
2556 		 * probing this for CPUs which don't.
2557 		 */
2558 		return;
2559 	}
2560 
2561 	write_c0_entrylo0(~0ul);
2562 	back_to_back_c0_hazard();
2563 	entry = read_c0_entrylo0();
2564 
2565 	/* clear all non-PFN bits */
2566 	entry &= ~((1 << MIPS_ENTRYLO_PFN_SHIFT) - 1);
2567 	entry &= ~(MIPS_ENTRYLO_RI | MIPS_ENTRYLO_XI);
2568 
2569 	/* find a lower bound on PABITS, and upper bound on fill bits */
2570 	pabits = fls_long(entry) + 6;
2571 	fillbits = max_t(int, (int)BITS_PER_LONG - pabits, 0);
2572 
2573 	/* minus the RI & XI bits */
2574 	fillbits -= min_t(unsigned, fillbits, 2);
2575 
2576 	if (fillbits >= ilog2(_PAGE_NO_EXEC))
2577 		fill_includes_sw_bits = true;
2578 
2579 	pr_debug("Entry* registers contain %u fill bits\n", fillbits);
2580 }
2581 
2582 void build_tlb_refill_handler(void)
2583 {
2584 	/*
2585 	 * The refill handler is generated per-CPU, multi-node systems
2586 	 * may have local storage for it. The other handlers are only
2587 	 * needed once.
2588 	 */
2589 	static int run_once = 0;
2590 
2591 	if (IS_ENABLED(CONFIG_XPA) && !cpu_has_rixi)
2592 		panic("Kernels supporting XPA currently require CPUs with RIXI");
2593 
2594 	output_pgtable_bits_defines();
2595 	check_pabits();
2596 
2597 #ifdef CONFIG_64BIT
2598 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2599 #endif
2600 
2601 	switch (current_cpu_type()) {
2602 	case CPU_R2000:
2603 	case CPU_R3000:
2604 	case CPU_R3000A:
2605 	case CPU_R3081E:
2606 	case CPU_TX3912:
2607 	case CPU_TX3922:
2608 	case CPU_TX3927:
2609 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2610 		if (cpu_has_local_ebase)
2611 			build_r3000_tlb_refill_handler();
2612 		if (!run_once) {
2613 			if (!cpu_has_local_ebase)
2614 				build_r3000_tlb_refill_handler();
2615 			build_setup_pgd();
2616 			build_r3000_tlb_load_handler();
2617 			build_r3000_tlb_store_handler();
2618 			build_r3000_tlb_modify_handler();
2619 			flush_tlb_handlers();
2620 			run_once++;
2621 		}
2622 #else
2623 		panic("No R3000 TLB refill handler");
2624 #endif
2625 		break;
2626 
2627 	case CPU_R8000:
2628 		panic("No R8000 TLB refill handler yet");
2629 		break;
2630 
2631 	default:
2632 		if (cpu_has_ldpte)
2633 			setup_pw();
2634 
2635 		if (!run_once) {
2636 			scratch_reg = allocate_kscratch();
2637 			build_setup_pgd();
2638 			build_r4000_tlb_load_handler();
2639 			build_r4000_tlb_store_handler();
2640 			build_r4000_tlb_modify_handler();
2641 			if (cpu_has_ldpte)
2642 				build_loongson3_tlb_refill_handler();
2643 			else if (!cpu_has_local_ebase)
2644 				build_r4000_tlb_refill_handler();
2645 			flush_tlb_handlers();
2646 			run_once++;
2647 		}
2648 		if (cpu_has_local_ebase)
2649 			build_r4000_tlb_refill_handler();
2650 		if (cpu_has_xpa)
2651 			config_xpa_params();
2652 		if (cpu_has_htw)
2653 			config_htw_params();
2654 	}
2655 }
2656