xref: /linux/arch/mips/mm/tlbex.c (revision 7056741fd9fc14a65608549a4657cf5178f05f63)
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 completly 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/init.h>
30 #include <linux/cache.h>
31 
32 #include <asm/cacheflush.h>
33 #include <asm/pgtable.h>
34 #include <asm/war.h>
35 #include <asm/uasm.h>
36 #include <asm/setup.h>
37 
38 /*
39  * TLB load/store/modify handlers.
40  *
41  * Only the fastpath gets synthesized at runtime, the slowpath for
42  * do_page_fault remains normal asm.
43  */
44 extern void tlb_do_page_fault_0(void);
45 extern void tlb_do_page_fault_1(void);
46 
47 struct work_registers {
48 	int r1;
49 	int r2;
50 	int r3;
51 };
52 
53 struct tlb_reg_save {
54 	unsigned long a;
55 	unsigned long b;
56 } ____cacheline_aligned_in_smp;
57 
58 static struct tlb_reg_save handler_reg_save[NR_CPUS];
59 
60 static inline int r45k_bvahwbug(void)
61 {
62 	/* XXX: We should probe for the presence of this bug, but we don't. */
63 	return 0;
64 }
65 
66 static inline int r4k_250MHZhwbug(void)
67 {
68 	/* XXX: We should probe for the presence of this bug, but we don't. */
69 	return 0;
70 }
71 
72 static inline int __maybe_unused bcm1250_m3_war(void)
73 {
74 	return BCM1250_M3_WAR;
75 }
76 
77 static inline int __maybe_unused r10000_llsc_war(void)
78 {
79 	return R10000_LLSC_WAR;
80 }
81 
82 static int use_bbit_insns(void)
83 {
84 	switch (current_cpu_type()) {
85 	case CPU_CAVIUM_OCTEON:
86 	case CPU_CAVIUM_OCTEON_PLUS:
87 	case CPU_CAVIUM_OCTEON2:
88 		return 1;
89 	default:
90 		return 0;
91 	}
92 }
93 
94 static int use_lwx_insns(void)
95 {
96 	switch (current_cpu_type()) {
97 	case CPU_CAVIUM_OCTEON2:
98 		return 1;
99 	default:
100 		return 0;
101 	}
102 }
103 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
104     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
105 static bool scratchpad_available(void)
106 {
107 	return true;
108 }
109 static int scratchpad_offset(int i)
110 {
111 	/*
112 	 * CVMSEG starts at address -32768 and extends for
113 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
114 	 */
115 	i += 1; /* Kernel use starts at the top and works down. */
116 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
117 }
118 #else
119 static bool scratchpad_available(void)
120 {
121 	return false;
122 }
123 static int scratchpad_offset(int i)
124 {
125 	BUG();
126 	/* Really unreachable, but evidently some GCC want this. */
127 	return 0;
128 }
129 #endif
130 /*
131  * Found by experiment: At least some revisions of the 4kc throw under
132  * some circumstances a machine check exception, triggered by invalid
133  * values in the index register.  Delaying the tlbp instruction until
134  * after the next branch,  plus adding an additional nop in front of
135  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
136  * why; it's not an issue caused by the core RTL.
137  *
138  */
139 static int __cpuinit m4kc_tlbp_war(void)
140 {
141 	return (current_cpu_data.processor_id & 0xffff00) ==
142 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
143 }
144 
145 /* Handle labels (which must be positive integers). */
146 enum label_id {
147 	label_second_part = 1,
148 	label_leave,
149 	label_vmalloc,
150 	label_vmalloc_done,
151 	label_tlbw_hazard,
152 	label_split,
153 	label_tlbl_goaround1,
154 	label_tlbl_goaround2,
155 	label_nopage_tlbl,
156 	label_nopage_tlbs,
157 	label_nopage_tlbm,
158 	label_smp_pgtable_change,
159 	label_r3000_write_probe_fail,
160 	label_large_segbits_fault,
161 #ifdef CONFIG_HUGETLB_PAGE
162 	label_tlb_huge_update,
163 #endif
164 };
165 
166 UASM_L_LA(_second_part)
167 UASM_L_LA(_leave)
168 UASM_L_LA(_vmalloc)
169 UASM_L_LA(_vmalloc_done)
170 UASM_L_LA(_tlbw_hazard)
171 UASM_L_LA(_split)
172 UASM_L_LA(_tlbl_goaround1)
173 UASM_L_LA(_tlbl_goaround2)
174 UASM_L_LA(_nopage_tlbl)
175 UASM_L_LA(_nopage_tlbs)
176 UASM_L_LA(_nopage_tlbm)
177 UASM_L_LA(_smp_pgtable_change)
178 UASM_L_LA(_r3000_write_probe_fail)
179 UASM_L_LA(_large_segbits_fault)
180 #ifdef CONFIG_HUGETLB_PAGE
181 UASM_L_LA(_tlb_huge_update)
182 #endif
183 
184 /*
185  * For debug purposes.
186  */
187 static inline void dump_handler(const u32 *handler, int count)
188 {
189 	int i;
190 
191 	pr_debug("\t.set push\n");
192 	pr_debug("\t.set noreorder\n");
193 
194 	for (i = 0; i < count; i++)
195 		pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
196 
197 	pr_debug("\t.set pop\n");
198 }
199 
200 /* The only general purpose registers allowed in TLB handlers. */
201 #define K0		26
202 #define K1		27
203 
204 /* Some CP0 registers */
205 #define C0_INDEX	0, 0
206 #define C0_ENTRYLO0	2, 0
207 #define C0_TCBIND	2, 2
208 #define C0_ENTRYLO1	3, 0
209 #define C0_CONTEXT	4, 0
210 #define C0_PAGEMASK	5, 0
211 #define C0_BADVADDR	8, 0
212 #define C0_ENTRYHI	10, 0
213 #define C0_EPC		14, 0
214 #define C0_XCONTEXT	20, 0
215 
216 #ifdef CONFIG_64BIT
217 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
218 #else
219 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
220 #endif
221 
222 /* The worst case length of the handler is around 18 instructions for
223  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
224  * Maximum space available is 32 instructions for R3000 and 64
225  * instructions for R4000.
226  *
227  * We deliberately chose a buffer size of 128, so we won't scribble
228  * over anything important on overflow before we panic.
229  */
230 static u32 tlb_handler[128] __cpuinitdata;
231 
232 /* simply assume worst case size for labels and relocs */
233 static struct uasm_label labels[128] __cpuinitdata;
234 static struct uasm_reloc relocs[128] __cpuinitdata;
235 
236 #ifdef CONFIG_64BIT
237 static int check_for_high_segbits __cpuinitdata;
238 #endif
239 
240 static int check_for_high_segbits __cpuinitdata;
241 
242 static unsigned int kscratch_used_mask __cpuinitdata;
243 
244 static int __cpuinit allocate_kscratch(void)
245 {
246 	int r;
247 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
248 
249 	r = ffs(a);
250 
251 	if (r == 0)
252 		return -1;
253 
254 	r--; /* make it zero based */
255 
256 	kscratch_used_mask |= (1 << r);
257 
258 	return r;
259 }
260 
261 static int scratch_reg __cpuinitdata;
262 static int pgd_reg __cpuinitdata;
263 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
264 
265 static struct work_registers __cpuinit build_get_work_registers(u32 **p)
266 {
267 	struct work_registers r;
268 
269 	int smp_processor_id_reg;
270 	int smp_processor_id_sel;
271 	int smp_processor_id_shift;
272 
273 	if (scratch_reg > 0) {
274 		/* Save in CPU local C0_KScratch? */
275 		UASM_i_MTC0(p, 1, 31, scratch_reg);
276 		r.r1 = K0;
277 		r.r2 = K1;
278 		r.r3 = 1;
279 		return r;
280 	}
281 
282 	if (num_possible_cpus() > 1) {
283 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
284 		smp_processor_id_shift = 51;
285 		smp_processor_id_reg = 20; /* XContext */
286 		smp_processor_id_sel = 0;
287 #else
288 # ifdef CONFIG_32BIT
289 		smp_processor_id_shift = 25;
290 		smp_processor_id_reg = 4; /* Context */
291 		smp_processor_id_sel = 0;
292 # endif
293 # ifdef CONFIG_64BIT
294 		smp_processor_id_shift = 26;
295 		smp_processor_id_reg = 4; /* Context */
296 		smp_processor_id_sel = 0;
297 # endif
298 #endif
299 		/* Get smp_processor_id */
300 		UASM_i_MFC0(p, K0, smp_processor_id_reg, smp_processor_id_sel);
301 		UASM_i_SRL_SAFE(p, K0, K0, smp_processor_id_shift);
302 
303 		/* handler_reg_save index in K0 */
304 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
305 
306 		UASM_i_LA(p, K1, (long)&handler_reg_save);
307 		UASM_i_ADDU(p, K0, K0, K1);
308 	} else {
309 		UASM_i_LA(p, K0, (long)&handler_reg_save);
310 	}
311 	/* K0 now points to save area, save $1 and $2  */
312 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
313 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
314 
315 	r.r1 = K1;
316 	r.r2 = 1;
317 	r.r3 = 2;
318 	return r;
319 }
320 
321 static void __cpuinit build_restore_work_registers(u32 **p)
322 {
323 	if (scratch_reg > 0) {
324 		UASM_i_MFC0(p, 1, 31, scratch_reg);
325 		return;
326 	}
327 	/* K0 already points to save area, restore $1 and $2  */
328 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
329 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
330 }
331 
332 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
333 
334 /*
335  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
336  * we cannot do r3000 under these circumstances.
337  *
338  * Declare pgd_current here instead of including mmu_context.h to avoid type
339  * conflicts for tlbmiss_handler_setup_pgd
340  */
341 extern unsigned long pgd_current[];
342 
343 /*
344  * The R3000 TLB handler is simple.
345  */
346 static void __cpuinit build_r3000_tlb_refill_handler(void)
347 {
348 	long pgdc = (long)pgd_current;
349 	u32 *p;
350 
351 	memset(tlb_handler, 0, sizeof(tlb_handler));
352 	p = tlb_handler;
353 
354 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
355 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
356 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
357 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
358 	uasm_i_sll(&p, K0, K0, 2);
359 	uasm_i_addu(&p, K1, K1, K0);
360 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
361 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
362 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
363 	uasm_i_addu(&p, K1, K1, K0);
364 	uasm_i_lw(&p, K0, 0, K1);
365 	uasm_i_nop(&p); /* load delay */
366 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
367 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
368 	uasm_i_tlbwr(&p); /* cp0 delay */
369 	uasm_i_jr(&p, K1);
370 	uasm_i_rfe(&p); /* branch delay */
371 
372 	if (p > tlb_handler + 32)
373 		panic("TLB refill handler space exceeded");
374 
375 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
376 		 (unsigned int)(p - tlb_handler));
377 
378 	memcpy((void *)ebase, tlb_handler, 0x80);
379 
380 	dump_handler((u32 *)ebase, 32);
381 }
382 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
383 
384 /*
385  * The R4000 TLB handler is much more complicated. We have two
386  * consecutive handler areas with 32 instructions space each.
387  * Since they aren't used at the same time, we can overflow in the
388  * other one.To keep things simple, we first assume linear space,
389  * then we relocate it to the final handler layout as needed.
390  */
391 static u32 final_handler[64] __cpuinitdata;
392 
393 /*
394  * Hazards
395  *
396  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
397  * 2. A timing hazard exists for the TLBP instruction.
398  *
399  *      stalling_instruction
400  *      TLBP
401  *
402  * The JTLB is being read for the TLBP throughout the stall generated by the
403  * previous instruction. This is not really correct as the stalling instruction
404  * can modify the address used to access the JTLB.  The failure symptom is that
405  * the TLBP instruction will use an address created for the stalling instruction
406  * and not the address held in C0_ENHI and thus report the wrong results.
407  *
408  * The software work-around is to not allow the instruction preceding the TLBP
409  * to stall - make it an NOP or some other instruction guaranteed not to stall.
410  *
411  * Errata 2 will not be fixed.  This errata is also on the R5000.
412  *
413  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
414  */
415 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
416 {
417 	switch (current_cpu_type()) {
418 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
419 	case CPU_R4600:
420 	case CPU_R4700:
421 	case CPU_R5000:
422 	case CPU_R5000A:
423 	case CPU_NEVADA:
424 		uasm_i_nop(p);
425 		uasm_i_tlbp(p);
426 		break;
427 
428 	default:
429 		uasm_i_tlbp(p);
430 		break;
431 	}
432 }
433 
434 /*
435  * Write random or indexed TLB entry, and care about the hazards from
436  * the preceding mtc0 and for the following eret.
437  */
438 enum tlb_write_entry { tlb_random, tlb_indexed };
439 
440 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
441 					 struct uasm_reloc **r,
442 					 enum tlb_write_entry wmode)
443 {
444 	void(*tlbw)(u32 **) = NULL;
445 
446 	switch (wmode) {
447 	case tlb_random: tlbw = uasm_i_tlbwr; break;
448 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
449 	}
450 
451 	if (cpu_has_mips_r2) {
452 		/*
453 		 * The architecture spec says an ehb is required here,
454 		 * but a number of cores do not have the hazard and
455 		 * using an ehb causes an expensive pipeline stall.
456 		 */
457 		switch (current_cpu_type()) {
458 		case CPU_M14KC:
459 		case CPU_74K:
460 			break;
461 
462 		default:
463 			uasm_i_ehb(p);
464 			break;
465 		}
466 		tlbw(p);
467 		return;
468 	}
469 
470 	switch (current_cpu_type()) {
471 	case CPU_R4000PC:
472 	case CPU_R4000SC:
473 	case CPU_R4000MC:
474 	case CPU_R4400PC:
475 	case CPU_R4400SC:
476 	case CPU_R4400MC:
477 		/*
478 		 * This branch uses up a mtc0 hazard nop slot and saves
479 		 * two nops after the tlbw instruction.
480 		 */
481 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
482 		tlbw(p);
483 		uasm_l_tlbw_hazard(l, *p);
484 		uasm_i_nop(p);
485 		break;
486 
487 	case CPU_R4600:
488 	case CPU_R4700:
489 	case CPU_R5000:
490 	case CPU_R5000A:
491 		uasm_i_nop(p);
492 		tlbw(p);
493 		uasm_i_nop(p);
494 		break;
495 
496 	case CPU_R4300:
497 	case CPU_5KC:
498 	case CPU_TX49XX:
499 	case CPU_PR4450:
500 	case CPU_XLR:
501 		uasm_i_nop(p);
502 		tlbw(p);
503 		break;
504 
505 	case CPU_R10000:
506 	case CPU_R12000:
507 	case CPU_R14000:
508 	case CPU_4KC:
509 	case CPU_4KEC:
510 	case CPU_M14KC:
511 	case CPU_SB1:
512 	case CPU_SB1A:
513 	case CPU_4KSC:
514 	case CPU_20KC:
515 	case CPU_25KF:
516 	case CPU_BMIPS32:
517 	case CPU_BMIPS3300:
518 	case CPU_BMIPS4350:
519 	case CPU_BMIPS4380:
520 	case CPU_BMIPS5000:
521 	case CPU_LOONGSON2:
522 	case CPU_R5500:
523 		if (m4kc_tlbp_war())
524 			uasm_i_nop(p);
525 	case CPU_ALCHEMY:
526 		tlbw(p);
527 		break;
528 
529 	case CPU_NEVADA:
530 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
531 		/*
532 		 * This branch uses up a mtc0 hazard nop slot and saves
533 		 * a nop after the tlbw instruction.
534 		 */
535 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
536 		tlbw(p);
537 		uasm_l_tlbw_hazard(l, *p);
538 		break;
539 
540 	case CPU_RM7000:
541 		uasm_i_nop(p);
542 		uasm_i_nop(p);
543 		uasm_i_nop(p);
544 		uasm_i_nop(p);
545 		tlbw(p);
546 		break;
547 
548 	case CPU_RM9000:
549 		/*
550 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
551 		 * use of the JTLB for instructions should not occur for 4
552 		 * cpu cycles and use for data translations should not occur
553 		 * for 3 cpu cycles.
554 		 */
555 		uasm_i_ssnop(p);
556 		uasm_i_ssnop(p);
557 		uasm_i_ssnop(p);
558 		uasm_i_ssnop(p);
559 		tlbw(p);
560 		uasm_i_ssnop(p);
561 		uasm_i_ssnop(p);
562 		uasm_i_ssnop(p);
563 		uasm_i_ssnop(p);
564 		break;
565 
566 	case CPU_VR4111:
567 	case CPU_VR4121:
568 	case CPU_VR4122:
569 	case CPU_VR4181:
570 	case CPU_VR4181A:
571 		uasm_i_nop(p);
572 		uasm_i_nop(p);
573 		tlbw(p);
574 		uasm_i_nop(p);
575 		uasm_i_nop(p);
576 		break;
577 
578 	case CPU_VR4131:
579 	case CPU_VR4133:
580 	case CPU_R5432:
581 		uasm_i_nop(p);
582 		uasm_i_nop(p);
583 		tlbw(p);
584 		break;
585 
586 	case CPU_JZRISC:
587 		tlbw(p);
588 		uasm_i_nop(p);
589 		break;
590 
591 	default:
592 		panic("No TLB refill handler yet (CPU type: %d)",
593 		      current_cpu_data.cputype);
594 		break;
595 	}
596 }
597 
598 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
599 								  unsigned int reg)
600 {
601 	if (cpu_has_rixi) {
602 		UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL));
603 	} else {
604 #ifdef CONFIG_64BIT_PHYS_ADDR
605 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
606 #else
607 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
608 #endif
609 	}
610 }
611 
612 #ifdef CONFIG_HUGETLB_PAGE
613 
614 static __cpuinit void build_restore_pagemask(u32 **p,
615 					     struct uasm_reloc **r,
616 					     unsigned int tmp,
617 					     enum label_id lid,
618 					     int restore_scratch)
619 {
620 	if (restore_scratch) {
621 		/* Reset default page size */
622 		if (PM_DEFAULT_MASK >> 16) {
623 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
624 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
625 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
626 			uasm_il_b(p, r, lid);
627 		} else if (PM_DEFAULT_MASK) {
628 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
629 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
630 			uasm_il_b(p, r, lid);
631 		} else {
632 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
633 			uasm_il_b(p, r, lid);
634 		}
635 		if (scratch_reg > 0)
636 			UASM_i_MFC0(p, 1, 31, scratch_reg);
637 		else
638 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
639 	} else {
640 		/* Reset default page size */
641 		if (PM_DEFAULT_MASK >> 16) {
642 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
643 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
644 			uasm_il_b(p, r, lid);
645 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
646 		} else if (PM_DEFAULT_MASK) {
647 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
648 			uasm_il_b(p, r, lid);
649 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
650 		} else {
651 			uasm_il_b(p, r, lid);
652 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
653 		}
654 	}
655 }
656 
657 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
658 						 struct uasm_label **l,
659 						 struct uasm_reloc **r,
660 						 unsigned int tmp,
661 						 enum tlb_write_entry wmode,
662 						 int restore_scratch)
663 {
664 	/* Set huge page tlb entry size */
665 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
666 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
667 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
668 
669 	build_tlb_write_entry(p, l, r, wmode);
670 
671 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
672 }
673 
674 /*
675  * Check if Huge PTE is present, if so then jump to LABEL.
676  */
677 static void __cpuinit
678 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
679 		unsigned int pmd, int lid)
680 {
681 	UASM_i_LW(p, tmp, 0, pmd);
682 	if (use_bbit_insns()) {
683 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
684 	} else {
685 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
686 		uasm_il_bnez(p, r, tmp, lid);
687 	}
688 }
689 
690 static __cpuinit void build_huge_update_entries(u32 **p,
691 						unsigned int pte,
692 						unsigned int tmp)
693 {
694 	int small_sequence;
695 
696 	/*
697 	 * A huge PTE describes an area the size of the
698 	 * configured huge page size. This is twice the
699 	 * of the large TLB entry size we intend to use.
700 	 * A TLB entry half the size of the configured
701 	 * huge page size is configured into entrylo0
702 	 * and entrylo1 to cover the contiguous huge PTE
703 	 * address space.
704 	 */
705 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
706 
707 	/* We can clobber tmp.  It isn't used after this.*/
708 	if (!small_sequence)
709 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
710 
711 	build_convert_pte_to_entrylo(p, pte);
712 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
713 	/* convert to entrylo1 */
714 	if (small_sequence)
715 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
716 	else
717 		UASM_i_ADDU(p, pte, pte, tmp);
718 
719 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
720 }
721 
722 static __cpuinit void build_huge_handler_tail(u32 **p,
723 					      struct uasm_reloc **r,
724 					      struct uasm_label **l,
725 					      unsigned int pte,
726 					      unsigned int ptr)
727 {
728 #ifdef CONFIG_SMP
729 	UASM_i_SC(p, pte, 0, ptr);
730 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
731 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
732 #else
733 	UASM_i_SW(p, pte, 0, ptr);
734 #endif
735 	build_huge_update_entries(p, pte, ptr);
736 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
737 }
738 #endif /* CONFIG_HUGETLB_PAGE */
739 
740 #ifdef CONFIG_64BIT
741 /*
742  * TMP and PTR are scratch.
743  * TMP will be clobbered, PTR will hold the pmd entry.
744  */
745 static void __cpuinit
746 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
747 		 unsigned int tmp, unsigned int ptr)
748 {
749 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
750 	long pgdc = (long)pgd_current;
751 #endif
752 	/*
753 	 * The vmalloc handling is not in the hotpath.
754 	 */
755 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
756 
757 	if (check_for_high_segbits) {
758 		/*
759 		 * The kernel currently implicitely assumes that the
760 		 * MIPS SEGBITS parameter for the processor is
761 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
762 		 * allocate virtual addresses outside the maximum
763 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
764 		 * that doesn't prevent user code from accessing the
765 		 * higher xuseg addresses.  Here, we make sure that
766 		 * everything but the lower xuseg addresses goes down
767 		 * the module_alloc/vmalloc path.
768 		 */
769 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
770 		uasm_il_bnez(p, r, ptr, label_vmalloc);
771 	} else {
772 		uasm_il_bltz(p, r, tmp, label_vmalloc);
773 	}
774 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
775 
776 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
777 	if (pgd_reg != -1) {
778 		/* pgd is in pgd_reg */
779 		UASM_i_MFC0(p, ptr, 31, pgd_reg);
780 	} else {
781 		/*
782 		 * &pgd << 11 stored in CONTEXT [23..63].
783 		 */
784 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
785 
786 		/* Clear lower 23 bits of context. */
787 		uasm_i_dins(p, ptr, 0, 0, 23);
788 
789 		/* 1 0  1 0 1  << 6  xkphys cached */
790 		uasm_i_ori(p, ptr, ptr, 0x540);
791 		uasm_i_drotr(p, ptr, ptr, 11);
792 	}
793 #elif defined(CONFIG_SMP)
794 # ifdef  CONFIG_MIPS_MT_SMTC
795 	/*
796 	 * SMTC uses TCBind value as "CPU" index
797 	 */
798 	uasm_i_mfc0(p, ptr, C0_TCBIND);
799 	uasm_i_dsrl_safe(p, ptr, ptr, 19);
800 # else
801 	/*
802 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
803 	 * stored in CONTEXT.
804 	 */
805 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
806 	uasm_i_dsrl_safe(p, ptr, ptr, 23);
807 # endif
808 	UASM_i_LA_mostly(p, tmp, pgdc);
809 	uasm_i_daddu(p, ptr, ptr, tmp);
810 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
811 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
812 #else
813 	UASM_i_LA_mostly(p, ptr, pgdc);
814 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
815 #endif
816 
817 	uasm_l_vmalloc_done(l, *p);
818 
819 	/* get pgd offset in bytes */
820 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
821 
822 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
823 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
824 #ifndef __PAGETABLE_PMD_FOLDED
825 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
826 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
827 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
828 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
829 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
830 #endif
831 }
832 
833 /*
834  * BVADDR is the faulting address, PTR is scratch.
835  * PTR will hold the pgd for vmalloc.
836  */
837 static void __cpuinit
838 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
839 			unsigned int bvaddr, unsigned int ptr,
840 			enum vmalloc64_mode mode)
841 {
842 	long swpd = (long)swapper_pg_dir;
843 	int single_insn_swpd;
844 	int did_vmalloc_branch = 0;
845 
846 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
847 
848 	uasm_l_vmalloc(l, *p);
849 
850 	if (mode != not_refill && check_for_high_segbits) {
851 		if (single_insn_swpd) {
852 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
853 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
854 			did_vmalloc_branch = 1;
855 			/* fall through */
856 		} else {
857 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
858 		}
859 	}
860 	if (!did_vmalloc_branch) {
861 		if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
862 			uasm_il_b(p, r, label_vmalloc_done);
863 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
864 		} else {
865 			UASM_i_LA_mostly(p, ptr, swpd);
866 			uasm_il_b(p, r, label_vmalloc_done);
867 			if (uasm_in_compat_space_p(swpd))
868 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
869 			else
870 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
871 		}
872 	}
873 	if (mode != not_refill && check_for_high_segbits) {
874 		uasm_l_large_segbits_fault(l, *p);
875 		/*
876 		 * We get here if we are an xsseg address, or if we are
877 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
878 		 *
879 		 * Ignoring xsseg (assume disabled so would generate
880 		 * (address errors?), the only remaining possibility
881 		 * is the upper xuseg addresses.  On processors with
882 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
883 		 * addresses would have taken an address error. We try
884 		 * to mimic that here by taking a load/istream page
885 		 * fault.
886 		 */
887 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
888 		uasm_i_jr(p, ptr);
889 
890 		if (mode == refill_scratch) {
891 			if (scratch_reg > 0)
892 				UASM_i_MFC0(p, 1, 31, scratch_reg);
893 			else
894 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
895 		} else {
896 			uasm_i_nop(p);
897 		}
898 	}
899 }
900 
901 #else /* !CONFIG_64BIT */
902 
903 /*
904  * TMP and PTR are scratch.
905  * TMP will be clobbered, PTR will hold the pgd entry.
906  */
907 static void __cpuinit __maybe_unused
908 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
909 {
910 	long pgdc = (long)pgd_current;
911 
912 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
913 #ifdef CONFIG_SMP
914 #ifdef  CONFIG_MIPS_MT_SMTC
915 	/*
916 	 * SMTC uses TCBind value as "CPU" index
917 	 */
918 	uasm_i_mfc0(p, ptr, C0_TCBIND);
919 	UASM_i_LA_mostly(p, tmp, pgdc);
920 	uasm_i_srl(p, ptr, ptr, 19);
921 #else
922 	/*
923 	 * smp_processor_id() << 3 is stored in CONTEXT.
924          */
925 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
926 	UASM_i_LA_mostly(p, tmp, pgdc);
927 	uasm_i_srl(p, ptr, ptr, 23);
928 #endif
929 	uasm_i_addu(p, ptr, tmp, ptr);
930 #else
931 	UASM_i_LA_mostly(p, ptr, pgdc);
932 #endif
933 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
934 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
935 
936 	if (cpu_has_mips_r2) {
937 		uasm_i_ext(p, tmp, tmp, PGDIR_SHIFT, (32 - PGDIR_SHIFT));
938 		uasm_i_ins(p, ptr, tmp, PGD_T_LOG2, (32 - PGDIR_SHIFT));
939 		return;
940 	}
941 
942 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
943 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
944 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
945 }
946 
947 #endif /* !CONFIG_64BIT */
948 
949 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
950 {
951 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
952 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
953 
954 	switch (current_cpu_type()) {
955 	case CPU_VR41XX:
956 	case CPU_VR4111:
957 	case CPU_VR4121:
958 	case CPU_VR4122:
959 	case CPU_VR4131:
960 	case CPU_VR4181:
961 	case CPU_VR4181A:
962 	case CPU_VR4133:
963 		shift += 2;
964 		break;
965 
966 	default:
967 		break;
968 	}
969 
970 	if (shift)
971 		UASM_i_SRL(p, ctx, ctx, shift);
972 	uasm_i_andi(p, ctx, ctx, mask);
973 }
974 
975 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
976 {
977 	if (cpu_has_mips_r2) {
978 		/* PTE ptr offset is obtained from BadVAddr */
979 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
980 		UASM_i_LW(p, ptr, 0, ptr);
981 		uasm_i_ext(p, tmp, tmp, PAGE_SHIFT+1, PGDIR_SHIFT-PAGE_SHIFT-1);
982 		uasm_i_ins(p, ptr, tmp, PTE_T_LOG2+1, PGDIR_SHIFT-PAGE_SHIFT-1);
983 		return;
984 	}
985 
986 	/*
987 	 * Bug workaround for the Nevada. It seems as if under certain
988 	 * circumstances the move from cp0_context might produce a
989 	 * bogus result when the mfc0 instruction and its consumer are
990 	 * in a different cacheline or a load instruction, probably any
991 	 * memory reference, is between them.
992 	 */
993 	switch (current_cpu_type()) {
994 	case CPU_NEVADA:
995 		UASM_i_LW(p, ptr, 0, ptr);
996 		GET_CONTEXT(p, tmp); /* get context reg */
997 		break;
998 
999 	default:
1000 		GET_CONTEXT(p, tmp); /* get context reg */
1001 		UASM_i_LW(p, ptr, 0, ptr);
1002 		break;
1003 	}
1004 
1005 	build_adjust_context(p, tmp);
1006 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
1007 }
1008 
1009 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
1010 					unsigned int ptep)
1011 {
1012 	/*
1013 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
1014 	 * Kernel is a special case. Only a few CPUs use it.
1015 	 */
1016 #ifdef CONFIG_64BIT_PHYS_ADDR
1017 	if (cpu_has_64bits) {
1018 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
1019 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1020 		if (cpu_has_rixi) {
1021 			UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1022 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1023 			UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1024 		} else {
1025 			uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1026 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1027 			uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1028 		}
1029 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1030 	} else {
1031 		int pte_off_even = sizeof(pte_t) / 2;
1032 		int pte_off_odd = pte_off_even + sizeof(pte_t);
1033 
1034 		/* The pte entries are pre-shifted */
1035 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1036 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1037 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1038 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1039 	}
1040 #else
1041 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1042 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1043 	if (r45k_bvahwbug())
1044 		build_tlb_probe_entry(p);
1045 	if (cpu_has_rixi) {
1046 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL));
1047 		if (r4k_250MHZhwbug())
1048 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1049 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1050 		UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL));
1051 	} else {
1052 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1053 		if (r4k_250MHZhwbug())
1054 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1055 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1056 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1057 		if (r45k_bvahwbug())
1058 			uasm_i_mfc0(p, tmp, C0_INDEX);
1059 	}
1060 	if (r4k_250MHZhwbug())
1061 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1062 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1063 #endif
1064 }
1065 
1066 struct mips_huge_tlb_info {
1067 	int huge_pte;
1068 	int restore_scratch;
1069 };
1070 
1071 static struct mips_huge_tlb_info __cpuinit
1072 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1073 			       struct uasm_reloc **r, unsigned int tmp,
1074 			       unsigned int ptr, int c0_scratch)
1075 {
1076 	struct mips_huge_tlb_info rv;
1077 	unsigned int even, odd;
1078 	int vmalloc_branch_delay_filled = 0;
1079 	const int scratch = 1; /* Our extra working register */
1080 
1081 	rv.huge_pte = scratch;
1082 	rv.restore_scratch = 0;
1083 
1084 	if (check_for_high_segbits) {
1085 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1086 
1087 		if (pgd_reg != -1)
1088 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1089 		else
1090 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1091 
1092 		if (c0_scratch >= 0)
1093 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1094 		else
1095 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1096 
1097 		uasm_i_dsrl_safe(p, scratch, tmp,
1098 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1099 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1100 
1101 		if (pgd_reg == -1) {
1102 			vmalloc_branch_delay_filled = 1;
1103 			/* Clear lower 23 bits of context. */
1104 			uasm_i_dins(p, ptr, 0, 0, 23);
1105 		}
1106 	} else {
1107 		if (pgd_reg != -1)
1108 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1109 		else
1110 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1111 
1112 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1113 
1114 		if (c0_scratch >= 0)
1115 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1116 		else
1117 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1118 
1119 		if (pgd_reg == -1)
1120 			/* Clear lower 23 bits of context. */
1121 			uasm_i_dins(p, ptr, 0, 0, 23);
1122 
1123 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1124 	}
1125 
1126 	if (pgd_reg == -1) {
1127 		vmalloc_branch_delay_filled = 1;
1128 		/* 1 0  1 0 1  << 6  xkphys cached */
1129 		uasm_i_ori(p, ptr, ptr, 0x540);
1130 		uasm_i_drotr(p, ptr, ptr, 11);
1131 	}
1132 
1133 #ifdef __PAGETABLE_PMD_FOLDED
1134 #define LOC_PTEP scratch
1135 #else
1136 #define LOC_PTEP ptr
1137 #endif
1138 
1139 	if (!vmalloc_branch_delay_filled)
1140 		/* get pgd offset in bytes */
1141 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1142 
1143 	uasm_l_vmalloc_done(l, *p);
1144 
1145 	/*
1146 	 *                         tmp          ptr
1147 	 * fall-through case =   badvaddr  *pgd_current
1148 	 * vmalloc case      =   badvaddr  swapper_pg_dir
1149 	 */
1150 
1151 	if (vmalloc_branch_delay_filled)
1152 		/* get pgd offset in bytes */
1153 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1154 
1155 #ifdef __PAGETABLE_PMD_FOLDED
1156 	GET_CONTEXT(p, tmp); /* get context reg */
1157 #endif
1158 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1159 
1160 	if (use_lwx_insns()) {
1161 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1162 	} else {
1163 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1164 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1165 	}
1166 
1167 #ifndef __PAGETABLE_PMD_FOLDED
1168 	/* get pmd offset in bytes */
1169 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1170 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1171 	GET_CONTEXT(p, tmp); /* get context reg */
1172 
1173 	if (use_lwx_insns()) {
1174 		UASM_i_LWX(p, scratch, scratch, ptr);
1175 	} else {
1176 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1177 		UASM_i_LW(p, scratch, 0, ptr);
1178 	}
1179 #endif
1180 	/* Adjust the context during the load latency. */
1181 	build_adjust_context(p, tmp);
1182 
1183 #ifdef CONFIG_HUGETLB_PAGE
1184 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1185 	/*
1186 	 * The in the LWX case we don't want to do the load in the
1187 	 * delay slot.  It cannot issue in the same cycle and may be
1188 	 * speculative and unneeded.
1189 	 */
1190 	if (use_lwx_insns())
1191 		uasm_i_nop(p);
1192 #endif /* CONFIG_HUGETLB_PAGE */
1193 
1194 
1195 	/* build_update_entries */
1196 	if (use_lwx_insns()) {
1197 		even = ptr;
1198 		odd = tmp;
1199 		UASM_i_LWX(p, even, scratch, tmp);
1200 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1201 		UASM_i_LWX(p, odd, scratch, tmp);
1202 	} else {
1203 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1204 		even = tmp;
1205 		odd = ptr;
1206 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1207 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1208 	}
1209 	if (cpu_has_rixi) {
1210 		uasm_i_drotr(p, even, even, ilog2(_PAGE_GLOBAL));
1211 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1212 		uasm_i_drotr(p, odd, odd, ilog2(_PAGE_GLOBAL));
1213 	} else {
1214 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1215 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1216 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1217 	}
1218 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1219 
1220 	if (c0_scratch >= 0) {
1221 		UASM_i_MFC0(p, scratch, 31, c0_scratch);
1222 		build_tlb_write_entry(p, l, r, tlb_random);
1223 		uasm_l_leave(l, *p);
1224 		rv.restore_scratch = 1;
1225 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1226 		build_tlb_write_entry(p, l, r, tlb_random);
1227 		uasm_l_leave(l, *p);
1228 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1229 	} else {
1230 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1231 		build_tlb_write_entry(p, l, r, tlb_random);
1232 		uasm_l_leave(l, *p);
1233 		rv.restore_scratch = 1;
1234 	}
1235 
1236 	uasm_i_eret(p); /* return from trap */
1237 
1238 	return rv;
1239 }
1240 
1241 /*
1242  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1243  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1244  * slots before the XTLB refill exception handler which belong to the
1245  * unused TLB refill exception.
1246  */
1247 #define MIPS64_REFILL_INSNS 32
1248 
1249 static void __cpuinit build_r4000_tlb_refill_handler(void)
1250 {
1251 	u32 *p = tlb_handler;
1252 	struct uasm_label *l = labels;
1253 	struct uasm_reloc *r = relocs;
1254 	u32 *f;
1255 	unsigned int final_len;
1256 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1257 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1258 
1259 	memset(tlb_handler, 0, sizeof(tlb_handler));
1260 	memset(labels, 0, sizeof(labels));
1261 	memset(relocs, 0, sizeof(relocs));
1262 	memset(final_handler, 0, sizeof(final_handler));
1263 
1264 	if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1265 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1266 							  scratch_reg);
1267 		vmalloc_mode = refill_scratch;
1268 	} else {
1269 		htlb_info.huge_pte = K0;
1270 		htlb_info.restore_scratch = 0;
1271 		vmalloc_mode = refill_noscratch;
1272 		/*
1273 		 * create the plain linear handler
1274 		 */
1275 		if (bcm1250_m3_war()) {
1276 			unsigned int segbits = 44;
1277 
1278 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1279 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1280 			uasm_i_xor(&p, K0, K0, K1);
1281 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1282 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1283 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1284 			uasm_i_or(&p, K0, K0, K1);
1285 			uasm_il_bnez(&p, &r, K0, label_leave);
1286 			/* No need for uasm_i_nop */
1287 		}
1288 
1289 #ifdef CONFIG_64BIT
1290 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1291 #else
1292 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1293 #endif
1294 
1295 #ifdef CONFIG_HUGETLB_PAGE
1296 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1297 #endif
1298 
1299 		build_get_ptep(&p, K0, K1);
1300 		build_update_entries(&p, K0, K1);
1301 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1302 		uasm_l_leave(&l, p);
1303 		uasm_i_eret(&p); /* return from trap */
1304 	}
1305 #ifdef CONFIG_HUGETLB_PAGE
1306 	uasm_l_tlb_huge_update(&l, p);
1307 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1308 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1309 				   htlb_info.restore_scratch);
1310 #endif
1311 
1312 #ifdef CONFIG_64BIT
1313 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1314 #endif
1315 
1316 	/*
1317 	 * Overflow check: For the 64bit handler, we need at least one
1318 	 * free instruction slot for the wrap-around branch. In worst
1319 	 * case, if the intended insertion point is a delay slot, we
1320 	 * need three, with the second nop'ed and the third being
1321 	 * unused.
1322 	 */
1323 	/* Loongson2 ebase is different than r4k, we have more space */
1324 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1325 	if ((p - tlb_handler) > 64)
1326 		panic("TLB refill handler space exceeded");
1327 #else
1328 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1329 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1330 		&& uasm_insn_has_bdelay(relocs,
1331 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
1332 		panic("TLB refill handler space exceeded");
1333 #endif
1334 
1335 	/*
1336 	 * Now fold the handler in the TLB refill handler space.
1337 	 */
1338 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
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 #else /* CONFIG_64BIT */
1344 	f = final_handler + MIPS64_REFILL_INSNS;
1345 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1346 		/* Just copy the handler. */
1347 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1348 		final_len = p - tlb_handler;
1349 	} else {
1350 #if defined(CONFIG_HUGETLB_PAGE)
1351 		const enum label_id ls = label_tlb_huge_update;
1352 #else
1353 		const enum label_id ls = label_vmalloc;
1354 #endif
1355 		u32 *split;
1356 		int ov = 0;
1357 		int i;
1358 
1359 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1360 			;
1361 		BUG_ON(i == ARRAY_SIZE(labels));
1362 		split = labels[i].addr;
1363 
1364 		/*
1365 		 * See if we have overflown one way or the other.
1366 		 */
1367 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1368 		    split < p - MIPS64_REFILL_INSNS)
1369 			ov = 1;
1370 
1371 		if (ov) {
1372 			/*
1373 			 * Split two instructions before the end.  One
1374 			 * for the branch and one for the instruction
1375 			 * in the delay slot.
1376 			 */
1377 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1378 
1379 			/*
1380 			 * If the branch would fall in a delay slot,
1381 			 * we must back up an additional instruction
1382 			 * so that it is no longer in a delay slot.
1383 			 */
1384 			if (uasm_insn_has_bdelay(relocs, split - 1))
1385 				split--;
1386 		}
1387 		/* Copy first part of the handler. */
1388 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1389 		f += split - tlb_handler;
1390 
1391 		if (ov) {
1392 			/* Insert branch. */
1393 			uasm_l_split(&l, final_handler);
1394 			uasm_il_b(&f, &r, label_split);
1395 			if (uasm_insn_has_bdelay(relocs, split))
1396 				uasm_i_nop(&f);
1397 			else {
1398 				uasm_copy_handler(relocs, labels,
1399 						  split, split + 1, f);
1400 				uasm_move_labels(labels, f, f + 1, -1);
1401 				f++;
1402 				split++;
1403 			}
1404 		}
1405 
1406 		/* Copy the rest of the handler. */
1407 		uasm_copy_handler(relocs, labels, split, p, final_handler);
1408 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1409 			    (p - split);
1410 	}
1411 #endif /* CONFIG_64BIT */
1412 
1413 	uasm_resolve_relocs(relocs, labels);
1414 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1415 		 final_len);
1416 
1417 	memcpy((void *)ebase, final_handler, 0x100);
1418 
1419 	dump_handler((u32 *)ebase, 64);
1420 }
1421 
1422 /*
1423  * 128 instructions for the fastpath handler is generous and should
1424  * never be exceeded.
1425  */
1426 #define FASTPATH_SIZE 128
1427 
1428 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1429 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1430 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1431 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1432 u32 tlbmiss_handler_setup_pgd[16] __cacheline_aligned;
1433 
1434 static void __cpuinit build_r4000_setup_pgd(void)
1435 {
1436 	const int a0 = 4;
1437 	const int a1 = 5;
1438 	u32 *p = tlbmiss_handler_setup_pgd;
1439 	struct uasm_label *l = labels;
1440 	struct uasm_reloc *r = relocs;
1441 
1442 	memset(tlbmiss_handler_setup_pgd, 0, sizeof(tlbmiss_handler_setup_pgd));
1443 	memset(labels, 0, sizeof(labels));
1444 	memset(relocs, 0, sizeof(relocs));
1445 
1446 	pgd_reg = allocate_kscratch();
1447 
1448 	if (pgd_reg == -1) {
1449 		/* PGD << 11 in c0_Context */
1450 		/*
1451 		 * If it is a ckseg0 address, convert to a physical
1452 		 * address.  Shifting right by 29 and adding 4 will
1453 		 * result in zero for these addresses.
1454 		 *
1455 		 */
1456 		UASM_i_SRA(&p, a1, a0, 29);
1457 		UASM_i_ADDIU(&p, a1, a1, 4);
1458 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1459 		uasm_i_nop(&p);
1460 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1461 		uasm_l_tlbl_goaround1(&l, p);
1462 		UASM_i_SLL(&p, a0, a0, 11);
1463 		uasm_i_jr(&p, 31);
1464 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1465 	} else {
1466 		/* PGD in c0_KScratch */
1467 		uasm_i_jr(&p, 31);
1468 		UASM_i_MTC0(&p, a0, 31, pgd_reg);
1469 	}
1470 	if (p - tlbmiss_handler_setup_pgd > ARRAY_SIZE(tlbmiss_handler_setup_pgd))
1471 		panic("tlbmiss_handler_setup_pgd space exceeded");
1472 	uasm_resolve_relocs(relocs, labels);
1473 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1474 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1475 
1476 	dump_handler(tlbmiss_handler_setup_pgd,
1477 		     ARRAY_SIZE(tlbmiss_handler_setup_pgd));
1478 }
1479 #endif
1480 
1481 static void __cpuinit
1482 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1483 {
1484 #ifdef CONFIG_SMP
1485 # ifdef CONFIG_64BIT_PHYS_ADDR
1486 	if (cpu_has_64bits)
1487 		uasm_i_lld(p, pte, 0, ptr);
1488 	else
1489 # endif
1490 		UASM_i_LL(p, pte, 0, ptr);
1491 #else
1492 # ifdef CONFIG_64BIT_PHYS_ADDR
1493 	if (cpu_has_64bits)
1494 		uasm_i_ld(p, pte, 0, ptr);
1495 	else
1496 # endif
1497 		UASM_i_LW(p, pte, 0, ptr);
1498 #endif
1499 }
1500 
1501 static void __cpuinit
1502 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1503 	unsigned int mode)
1504 {
1505 #ifdef CONFIG_64BIT_PHYS_ADDR
1506 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1507 #endif
1508 
1509 	uasm_i_ori(p, pte, pte, mode);
1510 #ifdef CONFIG_SMP
1511 # ifdef CONFIG_64BIT_PHYS_ADDR
1512 	if (cpu_has_64bits)
1513 		uasm_i_scd(p, pte, 0, ptr);
1514 	else
1515 # endif
1516 		UASM_i_SC(p, pte, 0, ptr);
1517 
1518 	if (r10000_llsc_war())
1519 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1520 	else
1521 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1522 
1523 # ifdef CONFIG_64BIT_PHYS_ADDR
1524 	if (!cpu_has_64bits) {
1525 		/* no uasm_i_nop needed */
1526 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1527 		uasm_i_ori(p, pte, pte, hwmode);
1528 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1529 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1530 		/* no uasm_i_nop needed */
1531 		uasm_i_lw(p, pte, 0, ptr);
1532 	} else
1533 		uasm_i_nop(p);
1534 # else
1535 	uasm_i_nop(p);
1536 # endif
1537 #else
1538 # ifdef CONFIG_64BIT_PHYS_ADDR
1539 	if (cpu_has_64bits)
1540 		uasm_i_sd(p, pte, 0, ptr);
1541 	else
1542 # endif
1543 		UASM_i_SW(p, pte, 0, ptr);
1544 
1545 # ifdef CONFIG_64BIT_PHYS_ADDR
1546 	if (!cpu_has_64bits) {
1547 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1548 		uasm_i_ori(p, pte, pte, hwmode);
1549 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1550 		uasm_i_lw(p, pte, 0, ptr);
1551 	}
1552 # endif
1553 #endif
1554 }
1555 
1556 /*
1557  * Check if PTE is present, if not then jump to LABEL. PTR points to
1558  * the page table where this PTE is located, PTE will be re-loaded
1559  * with it's original value.
1560  */
1561 static void __cpuinit
1562 build_pte_present(u32 **p, struct uasm_reloc **r,
1563 		  int pte, int ptr, int scratch, enum label_id lid)
1564 {
1565 	int t = scratch >= 0 ? scratch : pte;
1566 
1567 	if (cpu_has_rixi) {
1568 		if (use_bbit_insns()) {
1569 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1570 			uasm_i_nop(p);
1571 		} else {
1572 			uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1573 			uasm_il_beqz(p, r, t, lid);
1574 			if (pte == t)
1575 				/* You lose the SMP race :-(*/
1576 				iPTE_LW(p, pte, ptr);
1577 		}
1578 	} else {
1579 		uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1580 		uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1581 		uasm_il_bnez(p, r, t, lid);
1582 		if (pte == t)
1583 			/* You lose the SMP race :-(*/
1584 			iPTE_LW(p, pte, ptr);
1585 	}
1586 }
1587 
1588 /* Make PTE valid, store result in PTR. */
1589 static void __cpuinit
1590 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1591 		 unsigned int ptr)
1592 {
1593 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1594 
1595 	iPTE_SW(p, r, pte, ptr, mode);
1596 }
1597 
1598 /*
1599  * Check if PTE can be written to, if not branch to LABEL. Regardless
1600  * restore PTE with value from PTR when done.
1601  */
1602 static void __cpuinit
1603 build_pte_writable(u32 **p, struct uasm_reloc **r,
1604 		   unsigned int pte, unsigned int ptr, int scratch,
1605 		   enum label_id lid)
1606 {
1607 	int t = scratch >= 0 ? scratch : pte;
1608 
1609 	uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1610 	uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1611 	uasm_il_bnez(p, r, t, lid);
1612 	if (pte == t)
1613 		/* You lose the SMP race :-(*/
1614 		iPTE_LW(p, pte, ptr);
1615 	else
1616 		uasm_i_nop(p);
1617 }
1618 
1619 /* Make PTE writable, update software status bits as well, then store
1620  * at PTR.
1621  */
1622 static void __cpuinit
1623 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1624 		 unsigned int ptr)
1625 {
1626 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1627 			     | _PAGE_DIRTY);
1628 
1629 	iPTE_SW(p, r, pte, ptr, mode);
1630 }
1631 
1632 /*
1633  * Check if PTE can be modified, if not branch to LABEL. Regardless
1634  * restore PTE with value from PTR when done.
1635  */
1636 static void __cpuinit
1637 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1638 		     unsigned int pte, unsigned int ptr, int scratch,
1639 		     enum label_id lid)
1640 {
1641 	if (use_bbit_insns()) {
1642 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1643 		uasm_i_nop(p);
1644 	} else {
1645 		int t = scratch >= 0 ? scratch : pte;
1646 		uasm_i_andi(p, t, pte, _PAGE_WRITE);
1647 		uasm_il_beqz(p, r, t, lid);
1648 		if (pte == t)
1649 			/* You lose the SMP race :-(*/
1650 			iPTE_LW(p, pte, ptr);
1651 	}
1652 }
1653 
1654 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1655 
1656 
1657 /*
1658  * R3000 style TLB load/store/modify handlers.
1659  */
1660 
1661 /*
1662  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1663  * Then it returns.
1664  */
1665 static void __cpuinit
1666 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1667 {
1668 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1669 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1670 	uasm_i_tlbwi(p);
1671 	uasm_i_jr(p, tmp);
1672 	uasm_i_rfe(p); /* branch delay */
1673 }
1674 
1675 /*
1676  * This places the pte into ENTRYLO0 and writes it with tlbwi
1677  * or tlbwr as appropriate.  This is because the index register
1678  * may have the probe fail bit set as a result of a trap on a
1679  * kseg2 access, i.e. without refill.  Then it returns.
1680  */
1681 static void __cpuinit
1682 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1683 			     struct uasm_reloc **r, unsigned int pte,
1684 			     unsigned int tmp)
1685 {
1686 	uasm_i_mfc0(p, tmp, C0_INDEX);
1687 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1688 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1689 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1690 	uasm_i_tlbwi(p); /* cp0 delay */
1691 	uasm_i_jr(p, tmp);
1692 	uasm_i_rfe(p); /* branch delay */
1693 	uasm_l_r3000_write_probe_fail(l, *p);
1694 	uasm_i_tlbwr(p); /* cp0 delay */
1695 	uasm_i_jr(p, tmp);
1696 	uasm_i_rfe(p); /* branch delay */
1697 }
1698 
1699 static void __cpuinit
1700 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1701 				   unsigned int ptr)
1702 {
1703 	long pgdc = (long)pgd_current;
1704 
1705 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1706 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1707 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1708 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1709 	uasm_i_sll(p, pte, pte, 2);
1710 	uasm_i_addu(p, ptr, ptr, pte);
1711 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1712 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1713 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1714 	uasm_i_addu(p, ptr, ptr, pte);
1715 	uasm_i_lw(p, pte, 0, ptr);
1716 	uasm_i_tlbp(p); /* load delay */
1717 }
1718 
1719 static void __cpuinit build_r3000_tlb_load_handler(void)
1720 {
1721 	u32 *p = handle_tlbl;
1722 	struct uasm_label *l = labels;
1723 	struct uasm_reloc *r = relocs;
1724 
1725 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1726 	memset(labels, 0, sizeof(labels));
1727 	memset(relocs, 0, sizeof(relocs));
1728 
1729 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1730 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1731 	uasm_i_nop(&p); /* load delay */
1732 	build_make_valid(&p, &r, K0, K1);
1733 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1734 
1735 	uasm_l_nopage_tlbl(&l, p);
1736 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1737 	uasm_i_nop(&p);
1738 
1739 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1740 		panic("TLB load handler fastpath space exceeded");
1741 
1742 	uasm_resolve_relocs(relocs, labels);
1743 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1744 		 (unsigned int)(p - handle_tlbl));
1745 
1746 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1747 }
1748 
1749 static void __cpuinit build_r3000_tlb_store_handler(void)
1750 {
1751 	u32 *p = handle_tlbs;
1752 	struct uasm_label *l = labels;
1753 	struct uasm_reloc *r = relocs;
1754 
1755 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1756 	memset(labels, 0, sizeof(labels));
1757 	memset(relocs, 0, sizeof(relocs));
1758 
1759 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1760 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1761 	uasm_i_nop(&p); /* load delay */
1762 	build_make_write(&p, &r, K0, K1);
1763 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1764 
1765 	uasm_l_nopage_tlbs(&l, p);
1766 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1767 	uasm_i_nop(&p);
1768 
1769 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1770 		panic("TLB store handler fastpath space exceeded");
1771 
1772 	uasm_resolve_relocs(relocs, labels);
1773 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1774 		 (unsigned int)(p - handle_tlbs));
1775 
1776 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1777 }
1778 
1779 static void __cpuinit build_r3000_tlb_modify_handler(void)
1780 {
1781 	u32 *p = handle_tlbm;
1782 	struct uasm_label *l = labels;
1783 	struct uasm_reloc *r = relocs;
1784 
1785 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1786 	memset(labels, 0, sizeof(labels));
1787 	memset(relocs, 0, sizeof(relocs));
1788 
1789 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1790 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1791 	uasm_i_nop(&p); /* load delay */
1792 	build_make_write(&p, &r, K0, K1);
1793 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1794 
1795 	uasm_l_nopage_tlbm(&l, p);
1796 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1797 	uasm_i_nop(&p);
1798 
1799 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1800 		panic("TLB modify handler fastpath space exceeded");
1801 
1802 	uasm_resolve_relocs(relocs, labels);
1803 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1804 		 (unsigned int)(p - handle_tlbm));
1805 
1806 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1807 }
1808 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1809 
1810 /*
1811  * R4000 style TLB load/store/modify handlers.
1812  */
1813 static struct work_registers __cpuinit
1814 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1815 				   struct uasm_reloc **r)
1816 {
1817 	struct work_registers wr = build_get_work_registers(p);
1818 
1819 #ifdef CONFIG_64BIT
1820 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1821 #else
1822 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1823 #endif
1824 
1825 #ifdef CONFIG_HUGETLB_PAGE
1826 	/*
1827 	 * For huge tlb entries, pmd doesn't contain an address but
1828 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1829 	 * see if we need to jump to huge tlb processing.
1830 	 */
1831 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1832 #endif
1833 
1834 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1835 	UASM_i_LW(p, wr.r2, 0, wr.r2);
1836 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1837 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1838 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1839 
1840 #ifdef CONFIG_SMP
1841 	uasm_l_smp_pgtable_change(l, *p);
1842 #endif
1843 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1844 	if (!m4kc_tlbp_war())
1845 		build_tlb_probe_entry(p);
1846 	return wr;
1847 }
1848 
1849 static void __cpuinit
1850 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1851 				   struct uasm_reloc **r, unsigned int tmp,
1852 				   unsigned int ptr)
1853 {
1854 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1855 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1856 	build_update_entries(p, tmp, ptr);
1857 	build_tlb_write_entry(p, l, r, tlb_indexed);
1858 	uasm_l_leave(l, *p);
1859 	build_restore_work_registers(p);
1860 	uasm_i_eret(p); /* return from trap */
1861 
1862 #ifdef CONFIG_64BIT
1863 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1864 #endif
1865 }
1866 
1867 static void __cpuinit build_r4000_tlb_load_handler(void)
1868 {
1869 	u32 *p = handle_tlbl;
1870 	struct uasm_label *l = labels;
1871 	struct uasm_reloc *r = relocs;
1872 	struct work_registers wr;
1873 
1874 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1875 	memset(labels, 0, sizeof(labels));
1876 	memset(relocs, 0, sizeof(relocs));
1877 
1878 	if (bcm1250_m3_war()) {
1879 		unsigned int segbits = 44;
1880 
1881 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1882 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1883 		uasm_i_xor(&p, K0, K0, K1);
1884 		uasm_i_dsrl_safe(&p, K1, K0, 62);
1885 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1886 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1887 		uasm_i_or(&p, K0, K0, K1);
1888 		uasm_il_bnez(&p, &r, K0, label_leave);
1889 		/* No need for uasm_i_nop */
1890 	}
1891 
1892 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1893 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1894 	if (m4kc_tlbp_war())
1895 		build_tlb_probe_entry(&p);
1896 
1897 	if (cpu_has_rixi) {
1898 		/*
1899 		 * If the page is not _PAGE_VALID, RI or XI could not
1900 		 * have triggered it.  Skip the expensive test..
1901 		 */
1902 		if (use_bbit_insns()) {
1903 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1904 				      label_tlbl_goaround1);
1905 		} else {
1906 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1907 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1908 		}
1909 		uasm_i_nop(&p);
1910 
1911 		uasm_i_tlbr(&p);
1912 		/* Examine  entrylo 0 or 1 based on ptr. */
1913 		if (use_bbit_insns()) {
1914 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1915 		} else {
1916 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1917 			uasm_i_beqz(&p, wr.r3, 8);
1918 		}
1919 		/* load it in the delay slot*/
1920 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1921 		/* load it if ptr is odd */
1922 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1923 		/*
1924 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1925 		 * XI must have triggered it.
1926 		 */
1927 		if (use_bbit_insns()) {
1928 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1929 			uasm_i_nop(&p);
1930 			uasm_l_tlbl_goaround1(&l, p);
1931 		} else {
1932 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1933 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1934 			uasm_i_nop(&p);
1935 		}
1936 		uasm_l_tlbl_goaround1(&l, p);
1937 	}
1938 	build_make_valid(&p, &r, wr.r1, wr.r2);
1939 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1940 
1941 #ifdef CONFIG_HUGETLB_PAGE
1942 	/*
1943 	 * This is the entry point when build_r4000_tlbchange_handler_head
1944 	 * spots a huge page.
1945 	 */
1946 	uasm_l_tlb_huge_update(&l, p);
1947 	iPTE_LW(&p, wr.r1, wr.r2);
1948 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1949 	build_tlb_probe_entry(&p);
1950 
1951 	if (cpu_has_rixi) {
1952 		/*
1953 		 * If the page is not _PAGE_VALID, RI or XI could not
1954 		 * have triggered it.  Skip the expensive test..
1955 		 */
1956 		if (use_bbit_insns()) {
1957 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1958 				      label_tlbl_goaround2);
1959 		} else {
1960 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1961 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1962 		}
1963 		uasm_i_nop(&p);
1964 
1965 		uasm_i_tlbr(&p);
1966 		/* Examine  entrylo 0 or 1 based on ptr. */
1967 		if (use_bbit_insns()) {
1968 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1969 		} else {
1970 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1971 			uasm_i_beqz(&p, wr.r3, 8);
1972 		}
1973 		/* load it in the delay slot*/
1974 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1975 		/* load it if ptr is odd */
1976 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1977 		/*
1978 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1979 		 * XI must have triggered it.
1980 		 */
1981 		if (use_bbit_insns()) {
1982 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
1983 		} else {
1984 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1985 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1986 		}
1987 		if (PM_DEFAULT_MASK == 0)
1988 			uasm_i_nop(&p);
1989 		/*
1990 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
1991 		 * it is restored in build_huge_tlb_write_entry.
1992 		 */
1993 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
1994 
1995 		uasm_l_tlbl_goaround2(&l, p);
1996 	}
1997 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
1998 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
1999 #endif
2000 
2001 	uasm_l_nopage_tlbl(&l, p);
2002 	build_restore_work_registers(&p);
2003 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
2004 	uasm_i_nop(&p);
2005 
2006 	if ((p - handle_tlbl) > FASTPATH_SIZE)
2007 		panic("TLB load handler fastpath space exceeded");
2008 
2009 	uasm_resolve_relocs(relocs, labels);
2010 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
2011 		 (unsigned int)(p - handle_tlbl));
2012 
2013 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
2014 }
2015 
2016 static void __cpuinit build_r4000_tlb_store_handler(void)
2017 {
2018 	u32 *p = handle_tlbs;
2019 	struct uasm_label *l = labels;
2020 	struct uasm_reloc *r = relocs;
2021 	struct work_registers wr;
2022 
2023 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
2024 	memset(labels, 0, sizeof(labels));
2025 	memset(relocs, 0, sizeof(relocs));
2026 
2027 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2028 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2029 	if (m4kc_tlbp_war())
2030 		build_tlb_probe_entry(&p);
2031 	build_make_write(&p, &r, wr.r1, wr.r2);
2032 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2033 
2034 #ifdef CONFIG_HUGETLB_PAGE
2035 	/*
2036 	 * This is the entry point when
2037 	 * build_r4000_tlbchange_handler_head spots a huge page.
2038 	 */
2039 	uasm_l_tlb_huge_update(&l, p);
2040 	iPTE_LW(&p, wr.r1, wr.r2);
2041 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2042 	build_tlb_probe_entry(&p);
2043 	uasm_i_ori(&p, wr.r1, wr.r1,
2044 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2045 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2046 #endif
2047 
2048 	uasm_l_nopage_tlbs(&l, p);
2049 	build_restore_work_registers(&p);
2050 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2051 	uasm_i_nop(&p);
2052 
2053 	if ((p - handle_tlbs) > FASTPATH_SIZE)
2054 		panic("TLB store handler fastpath space exceeded");
2055 
2056 	uasm_resolve_relocs(relocs, labels);
2057 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2058 		 (unsigned int)(p - handle_tlbs));
2059 
2060 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
2061 }
2062 
2063 static void __cpuinit build_r4000_tlb_modify_handler(void)
2064 {
2065 	u32 *p = handle_tlbm;
2066 	struct uasm_label *l = labels;
2067 	struct uasm_reloc *r = relocs;
2068 	struct work_registers wr;
2069 
2070 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
2071 	memset(labels, 0, sizeof(labels));
2072 	memset(relocs, 0, sizeof(relocs));
2073 
2074 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2075 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2076 	if (m4kc_tlbp_war())
2077 		build_tlb_probe_entry(&p);
2078 	/* Present and writable bits set, set accessed and dirty bits. */
2079 	build_make_write(&p, &r, wr.r1, wr.r2);
2080 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2081 
2082 #ifdef CONFIG_HUGETLB_PAGE
2083 	/*
2084 	 * This is the entry point when
2085 	 * build_r4000_tlbchange_handler_head spots a huge page.
2086 	 */
2087 	uasm_l_tlb_huge_update(&l, p);
2088 	iPTE_LW(&p, wr.r1, wr.r2);
2089 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2090 	build_tlb_probe_entry(&p);
2091 	uasm_i_ori(&p, wr.r1, wr.r1,
2092 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2093 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2094 #endif
2095 
2096 	uasm_l_nopage_tlbm(&l, p);
2097 	build_restore_work_registers(&p);
2098 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2099 	uasm_i_nop(&p);
2100 
2101 	if ((p - handle_tlbm) > FASTPATH_SIZE)
2102 		panic("TLB modify handler fastpath space exceeded");
2103 
2104 	uasm_resolve_relocs(relocs, labels);
2105 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2106 		 (unsigned int)(p - handle_tlbm));
2107 
2108 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
2109 }
2110 
2111 void __cpuinit build_tlb_refill_handler(void)
2112 {
2113 	/*
2114 	 * The refill handler is generated per-CPU, multi-node systems
2115 	 * may have local storage for it. The other handlers are only
2116 	 * needed once.
2117 	 */
2118 	static int run_once = 0;
2119 
2120 #ifdef CONFIG_64BIT
2121 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2122 #endif
2123 
2124 	switch (current_cpu_type()) {
2125 	case CPU_R2000:
2126 	case CPU_R3000:
2127 	case CPU_R3000A:
2128 	case CPU_R3081E:
2129 	case CPU_TX3912:
2130 	case CPU_TX3922:
2131 	case CPU_TX3927:
2132 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2133 		build_r3000_tlb_refill_handler();
2134 		if (!run_once) {
2135 			build_r3000_tlb_load_handler();
2136 			build_r3000_tlb_store_handler();
2137 			build_r3000_tlb_modify_handler();
2138 			run_once++;
2139 		}
2140 #else
2141 		panic("No R3000 TLB refill handler");
2142 #endif
2143 		break;
2144 
2145 	case CPU_R6000:
2146 	case CPU_R6000A:
2147 		panic("No R6000 TLB refill handler yet");
2148 		break;
2149 
2150 	case CPU_R8000:
2151 		panic("No R8000 TLB refill handler yet");
2152 		break;
2153 
2154 	default:
2155 		if (!run_once) {
2156 			scratch_reg = allocate_kscratch();
2157 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2158 			build_r4000_setup_pgd();
2159 #endif
2160 			build_r4000_tlb_load_handler();
2161 			build_r4000_tlb_store_handler();
2162 			build_r4000_tlb_modify_handler();
2163 			run_once++;
2164 		}
2165 		build_r4000_tlb_refill_handler();
2166 	}
2167 }
2168 
2169 void __cpuinit flush_tlb_handlers(void)
2170 {
2171 	local_flush_icache_range((unsigned long)handle_tlbl,
2172 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2173 	local_flush_icache_range((unsigned long)handle_tlbs,
2174 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2175 	local_flush_icache_range((unsigned long)handle_tlbm,
2176 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2177 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2178 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2179 			   (unsigned long)tlbmiss_handler_setup_pgd + sizeof(handle_tlbm));
2180 #endif
2181 }
2182