xref: /linux/arch/mips/mm/tlbex.c (revision 60e13231561b3a4c5269bfa1ef6c0569ad6f28ec)
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  *
13  * ... and the days got worse and worse and now you see
14  * I've gone completly out of my mind.
15  *
16  * They're coming to take me a away haha
17  * they're coming to take me a away hoho hihi haha
18  * to the funny farm where code is beautiful all the time ...
19  *
20  * (Condolences to Napoleon XIV)
21  */
22 
23 #include <linux/bug.h>
24 #include <linux/kernel.h>
25 #include <linux/types.h>
26 #include <linux/smp.h>
27 #include <linux/string.h>
28 #include <linux/init.h>
29 #include <linux/cache.h>
30 
31 #include <asm/cacheflush.h>
32 #include <asm/pgtable.h>
33 #include <asm/war.h>
34 #include <asm/uasm.h>
35 
36 /*
37  * TLB load/store/modify handlers.
38  *
39  * Only the fastpath gets synthesized at runtime, the slowpath for
40  * do_page_fault remains normal asm.
41  */
42 extern void tlb_do_page_fault_0(void);
43 extern void tlb_do_page_fault_1(void);
44 
45 struct work_registers {
46 	int r1;
47 	int r2;
48 	int r3;
49 };
50 
51 struct tlb_reg_save {
52 	unsigned long a;
53 	unsigned long b;
54 } ____cacheline_aligned_in_smp;
55 
56 static struct tlb_reg_save handler_reg_save[NR_CPUS];
57 
58 static inline int r45k_bvahwbug(void)
59 {
60 	/* XXX: We should probe for the presence of this bug, but we don't. */
61 	return 0;
62 }
63 
64 static inline int r4k_250MHZhwbug(void)
65 {
66 	/* XXX: We should probe for the presence of this bug, but we don't. */
67 	return 0;
68 }
69 
70 static inline int __maybe_unused bcm1250_m3_war(void)
71 {
72 	return BCM1250_M3_WAR;
73 }
74 
75 static inline int __maybe_unused r10000_llsc_war(void)
76 {
77 	return R10000_LLSC_WAR;
78 }
79 
80 static int use_bbit_insns(void)
81 {
82 	switch (current_cpu_type()) {
83 	case CPU_CAVIUM_OCTEON:
84 	case CPU_CAVIUM_OCTEON_PLUS:
85 	case CPU_CAVIUM_OCTEON2:
86 		return 1;
87 	default:
88 		return 0;
89 	}
90 }
91 
92 static int use_lwx_insns(void)
93 {
94 	switch (current_cpu_type()) {
95 	case CPU_CAVIUM_OCTEON2:
96 		return 1;
97 	default:
98 		return 0;
99 	}
100 }
101 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
102     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
103 static bool scratchpad_available(void)
104 {
105 	return true;
106 }
107 static int scratchpad_offset(int i)
108 {
109 	/*
110 	 * CVMSEG starts at address -32768 and extends for
111 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
112 	 */
113 	i += 1; /* Kernel use starts at the top and works down. */
114 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
115 }
116 #else
117 static bool scratchpad_available(void)
118 {
119 	return false;
120 }
121 static int scratchpad_offset(int i)
122 {
123 	BUG();
124 	/* Really unreachable, but evidently some GCC want this. */
125 	return 0;
126 }
127 #endif
128 /*
129  * Found by experiment: At least some revisions of the 4kc throw under
130  * some circumstances a machine check exception, triggered by invalid
131  * values in the index register.  Delaying the tlbp instruction until
132  * after the next branch,  plus adding an additional nop in front of
133  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
134  * why; it's not an issue caused by the core RTL.
135  *
136  */
137 static int __cpuinit m4kc_tlbp_war(void)
138 {
139 	return (current_cpu_data.processor_id & 0xffff00) ==
140 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
141 }
142 
143 /* Handle labels (which must be positive integers). */
144 enum label_id {
145 	label_second_part = 1,
146 	label_leave,
147 	label_vmalloc,
148 	label_vmalloc_done,
149 	label_tlbw_hazard,
150 	label_split,
151 	label_tlbl_goaround1,
152 	label_tlbl_goaround2,
153 	label_nopage_tlbl,
154 	label_nopage_tlbs,
155 	label_nopage_tlbm,
156 	label_smp_pgtable_change,
157 	label_r3000_write_probe_fail,
158 	label_large_segbits_fault,
159 #ifdef CONFIG_HUGETLB_PAGE
160 	label_tlb_huge_update,
161 #endif
162 };
163 
164 UASM_L_LA(_second_part)
165 UASM_L_LA(_leave)
166 UASM_L_LA(_vmalloc)
167 UASM_L_LA(_vmalloc_done)
168 UASM_L_LA(_tlbw_hazard)
169 UASM_L_LA(_split)
170 UASM_L_LA(_tlbl_goaround1)
171 UASM_L_LA(_tlbl_goaround2)
172 UASM_L_LA(_nopage_tlbl)
173 UASM_L_LA(_nopage_tlbs)
174 UASM_L_LA(_nopage_tlbm)
175 UASM_L_LA(_smp_pgtable_change)
176 UASM_L_LA(_r3000_write_probe_fail)
177 UASM_L_LA(_large_segbits_fault)
178 #ifdef CONFIG_HUGETLB_PAGE
179 UASM_L_LA(_tlb_huge_update)
180 #endif
181 
182 /*
183  * For debug purposes.
184  */
185 static inline void dump_handler(const u32 *handler, int count)
186 {
187 	int i;
188 
189 	pr_debug("\t.set push\n");
190 	pr_debug("\t.set noreorder\n");
191 
192 	for (i = 0; i < count; i++)
193 		pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
194 
195 	pr_debug("\t.set pop\n");
196 }
197 
198 /* The only general purpose registers allowed in TLB handlers. */
199 #define K0		26
200 #define K1		27
201 
202 /* Some CP0 registers */
203 #define C0_INDEX	0, 0
204 #define C0_ENTRYLO0	2, 0
205 #define C0_TCBIND	2, 2
206 #define C0_ENTRYLO1	3, 0
207 #define C0_CONTEXT	4, 0
208 #define C0_PAGEMASK	5, 0
209 #define C0_BADVADDR	8, 0
210 #define C0_ENTRYHI	10, 0
211 #define C0_EPC		14, 0
212 #define C0_XCONTEXT	20, 0
213 
214 #ifdef CONFIG_64BIT
215 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
216 #else
217 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
218 #endif
219 
220 /* The worst case length of the handler is around 18 instructions for
221  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
222  * Maximum space available is 32 instructions for R3000 and 64
223  * instructions for R4000.
224  *
225  * We deliberately chose a buffer size of 128, so we won't scribble
226  * over anything important on overflow before we panic.
227  */
228 static u32 tlb_handler[128] __cpuinitdata;
229 
230 /* simply assume worst case size for labels and relocs */
231 static struct uasm_label labels[128] __cpuinitdata;
232 static struct uasm_reloc relocs[128] __cpuinitdata;
233 
234 #ifdef CONFIG_64BIT
235 static int check_for_high_segbits __cpuinitdata;
236 #endif
237 
238 static int check_for_high_segbits __cpuinitdata;
239 
240 static unsigned int kscratch_used_mask __cpuinitdata;
241 
242 static int __cpuinit allocate_kscratch(void)
243 {
244 	int r;
245 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
246 
247 	r = ffs(a);
248 
249 	if (r == 0)
250 		return -1;
251 
252 	r--; /* make it zero based */
253 
254 	kscratch_used_mask |= (1 << r);
255 
256 	return r;
257 }
258 
259 static int scratch_reg __cpuinitdata;
260 static int pgd_reg __cpuinitdata;
261 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
262 
263 static struct work_registers __cpuinit build_get_work_registers(u32 **p)
264 {
265 	struct work_registers r;
266 
267 	int smp_processor_id_reg;
268 	int smp_processor_id_sel;
269 	int smp_processor_id_shift;
270 
271 	if (scratch_reg > 0) {
272 		/* Save in CPU local C0_KScratch? */
273 		UASM_i_MTC0(p, 1, 31, scratch_reg);
274 		r.r1 = K0;
275 		r.r2 = K1;
276 		r.r3 = 1;
277 		return r;
278 	}
279 
280 	if (num_possible_cpus() > 1) {
281 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
282 		smp_processor_id_shift = 51;
283 		smp_processor_id_reg = 20; /* XContext */
284 		smp_processor_id_sel = 0;
285 #else
286 # ifdef CONFIG_32BIT
287 		smp_processor_id_shift = 25;
288 		smp_processor_id_reg = 4; /* Context */
289 		smp_processor_id_sel = 0;
290 # endif
291 # ifdef CONFIG_64BIT
292 		smp_processor_id_shift = 26;
293 		smp_processor_id_reg = 4; /* Context */
294 		smp_processor_id_sel = 0;
295 # endif
296 #endif
297 		/* Get smp_processor_id */
298 		UASM_i_MFC0(p, K0, smp_processor_id_reg, smp_processor_id_sel);
299 		UASM_i_SRL_SAFE(p, K0, K0, smp_processor_id_shift);
300 
301 		/* handler_reg_save index in K0 */
302 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
303 
304 		UASM_i_LA(p, K1, (long)&handler_reg_save);
305 		UASM_i_ADDU(p, K0, K0, K1);
306 	} else {
307 		UASM_i_LA(p, K0, (long)&handler_reg_save);
308 	}
309 	/* K0 now points to save area, save $1 and $2  */
310 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
311 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
312 
313 	r.r1 = K1;
314 	r.r2 = 1;
315 	r.r3 = 2;
316 	return r;
317 }
318 
319 static void __cpuinit build_restore_work_registers(u32 **p)
320 {
321 	if (scratch_reg > 0) {
322 		UASM_i_MFC0(p, 1, 31, scratch_reg);
323 		return;
324 	}
325 	/* K0 already points to save area, restore $1 and $2  */
326 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
327 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
328 }
329 
330 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
331 
332 /*
333  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
334  * we cannot do r3000 under these circumstances.
335  *
336  * Declare pgd_current here instead of including mmu_context.h to avoid type
337  * conflicts for tlbmiss_handler_setup_pgd
338  */
339 extern unsigned long pgd_current[];
340 
341 /*
342  * The R3000 TLB handler is simple.
343  */
344 static void __cpuinit build_r3000_tlb_refill_handler(void)
345 {
346 	long pgdc = (long)pgd_current;
347 	u32 *p;
348 
349 	memset(tlb_handler, 0, sizeof(tlb_handler));
350 	p = tlb_handler;
351 
352 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
353 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
354 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
355 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
356 	uasm_i_sll(&p, K0, K0, 2);
357 	uasm_i_addu(&p, K1, K1, K0);
358 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
359 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
360 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
361 	uasm_i_addu(&p, K1, K1, K0);
362 	uasm_i_lw(&p, K0, 0, K1);
363 	uasm_i_nop(&p); /* load delay */
364 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
365 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
366 	uasm_i_tlbwr(&p); /* cp0 delay */
367 	uasm_i_jr(&p, K1);
368 	uasm_i_rfe(&p); /* branch delay */
369 
370 	if (p > tlb_handler + 32)
371 		panic("TLB refill handler space exceeded");
372 
373 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
374 		 (unsigned int)(p - tlb_handler));
375 
376 	memcpy((void *)ebase, tlb_handler, 0x80);
377 
378 	dump_handler((u32 *)ebase, 32);
379 }
380 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
381 
382 /*
383  * The R4000 TLB handler is much more complicated. We have two
384  * consecutive handler areas with 32 instructions space each.
385  * Since they aren't used at the same time, we can overflow in the
386  * other one.To keep things simple, we first assume linear space,
387  * then we relocate it to the final handler layout as needed.
388  */
389 static u32 final_handler[64] __cpuinitdata;
390 
391 /*
392  * Hazards
393  *
394  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
395  * 2. A timing hazard exists for the TLBP instruction.
396  *
397  *      stalling_instruction
398  *      TLBP
399  *
400  * The JTLB is being read for the TLBP throughout the stall generated by the
401  * previous instruction. This is not really correct as the stalling instruction
402  * can modify the address used to access the JTLB.  The failure symptom is that
403  * the TLBP instruction will use an address created for the stalling instruction
404  * and not the address held in C0_ENHI and thus report the wrong results.
405  *
406  * The software work-around is to not allow the instruction preceding the TLBP
407  * to stall - make it an NOP or some other instruction guaranteed not to stall.
408  *
409  * Errata 2 will not be fixed.  This errata is also on the R5000.
410  *
411  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
412  */
413 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
414 {
415 	switch (current_cpu_type()) {
416 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
417 	case CPU_R4600:
418 	case CPU_R4700:
419 	case CPU_R5000:
420 	case CPU_R5000A:
421 	case CPU_NEVADA:
422 		uasm_i_nop(p);
423 		uasm_i_tlbp(p);
424 		break;
425 
426 	default:
427 		uasm_i_tlbp(p);
428 		break;
429 	}
430 }
431 
432 /*
433  * Write random or indexed TLB entry, and care about the hazards from
434  * the preceding mtc0 and for the following eret.
435  */
436 enum tlb_write_entry { tlb_random, tlb_indexed };
437 
438 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
439 					 struct uasm_reloc **r,
440 					 enum tlb_write_entry wmode)
441 {
442 	void(*tlbw)(u32 **) = NULL;
443 
444 	switch (wmode) {
445 	case tlb_random: tlbw = uasm_i_tlbwr; break;
446 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
447 	}
448 
449 	if (cpu_has_mips_r2) {
450 		if (cpu_has_mips_r2_exec_hazard)
451 			uasm_i_ehb(p);
452 		tlbw(p);
453 		return;
454 	}
455 
456 	switch (current_cpu_type()) {
457 	case CPU_R4000PC:
458 	case CPU_R4000SC:
459 	case CPU_R4000MC:
460 	case CPU_R4400PC:
461 	case CPU_R4400SC:
462 	case CPU_R4400MC:
463 		/*
464 		 * This branch uses up a mtc0 hazard nop slot and saves
465 		 * two nops after the tlbw instruction.
466 		 */
467 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
468 		tlbw(p);
469 		uasm_l_tlbw_hazard(l, *p);
470 		uasm_i_nop(p);
471 		break;
472 
473 	case CPU_R4600:
474 	case CPU_R4700:
475 	case CPU_R5000:
476 	case CPU_R5000A:
477 		uasm_i_nop(p);
478 		tlbw(p);
479 		uasm_i_nop(p);
480 		break;
481 
482 	case CPU_R4300:
483 	case CPU_5KC:
484 	case CPU_TX49XX:
485 	case CPU_PR4450:
486 	case CPU_XLR:
487 		uasm_i_nop(p);
488 		tlbw(p);
489 		break;
490 
491 	case CPU_R10000:
492 	case CPU_R12000:
493 	case CPU_R14000:
494 	case CPU_4KC:
495 	case CPU_4KEC:
496 	case CPU_SB1:
497 	case CPU_SB1A:
498 	case CPU_4KSC:
499 	case CPU_20KC:
500 	case CPU_25KF:
501 	case CPU_BMIPS32:
502 	case CPU_BMIPS3300:
503 	case CPU_BMIPS4350:
504 	case CPU_BMIPS4380:
505 	case CPU_BMIPS5000:
506 	case CPU_LOONGSON2:
507 	case CPU_R5500:
508 		if (m4kc_tlbp_war())
509 			uasm_i_nop(p);
510 	case CPU_ALCHEMY:
511 		tlbw(p);
512 		break;
513 
514 	case CPU_NEVADA:
515 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
516 		/*
517 		 * This branch uses up a mtc0 hazard nop slot and saves
518 		 * a nop after the tlbw instruction.
519 		 */
520 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
521 		tlbw(p);
522 		uasm_l_tlbw_hazard(l, *p);
523 		break;
524 
525 	case CPU_RM7000:
526 		uasm_i_nop(p);
527 		uasm_i_nop(p);
528 		uasm_i_nop(p);
529 		uasm_i_nop(p);
530 		tlbw(p);
531 		break;
532 
533 	case CPU_RM9000:
534 		/*
535 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
536 		 * use of the JTLB for instructions should not occur for 4
537 		 * cpu cycles and use for data translations should not occur
538 		 * for 3 cpu cycles.
539 		 */
540 		uasm_i_ssnop(p);
541 		uasm_i_ssnop(p);
542 		uasm_i_ssnop(p);
543 		uasm_i_ssnop(p);
544 		tlbw(p);
545 		uasm_i_ssnop(p);
546 		uasm_i_ssnop(p);
547 		uasm_i_ssnop(p);
548 		uasm_i_ssnop(p);
549 		break;
550 
551 	case CPU_VR4111:
552 	case CPU_VR4121:
553 	case CPU_VR4122:
554 	case CPU_VR4181:
555 	case CPU_VR4181A:
556 		uasm_i_nop(p);
557 		uasm_i_nop(p);
558 		tlbw(p);
559 		uasm_i_nop(p);
560 		uasm_i_nop(p);
561 		break;
562 
563 	case CPU_VR4131:
564 	case CPU_VR4133:
565 	case CPU_R5432:
566 		uasm_i_nop(p);
567 		uasm_i_nop(p);
568 		tlbw(p);
569 		break;
570 
571 	case CPU_JZRISC:
572 		tlbw(p);
573 		uasm_i_nop(p);
574 		break;
575 
576 	default:
577 		panic("No TLB refill handler yet (CPU type: %d)",
578 		      current_cpu_data.cputype);
579 		break;
580 	}
581 }
582 
583 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
584 								  unsigned int reg)
585 {
586 	if (kernel_uses_smartmips_rixi) {
587 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
588 		UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
589 	} else {
590 #ifdef CONFIG_64BIT_PHYS_ADDR
591 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
592 #else
593 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
594 #endif
595 	}
596 }
597 
598 #ifdef CONFIG_HUGETLB_PAGE
599 
600 static __cpuinit void build_restore_pagemask(u32 **p,
601 					     struct uasm_reloc **r,
602 					     unsigned int tmp,
603 					     enum label_id lid,
604 					     int restore_scratch)
605 {
606 	if (restore_scratch) {
607 		/* Reset default page size */
608 		if (PM_DEFAULT_MASK >> 16) {
609 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
610 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
611 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
612 			uasm_il_b(p, r, lid);
613 		} else if (PM_DEFAULT_MASK) {
614 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
615 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
616 			uasm_il_b(p, r, lid);
617 		} else {
618 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
619 			uasm_il_b(p, r, lid);
620 		}
621 		if (scratch_reg > 0)
622 			UASM_i_MFC0(p, 1, 31, scratch_reg);
623 		else
624 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
625 	} else {
626 		/* Reset default page size */
627 		if (PM_DEFAULT_MASK >> 16) {
628 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
629 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
630 			uasm_il_b(p, r, lid);
631 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
632 		} else if (PM_DEFAULT_MASK) {
633 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
634 			uasm_il_b(p, r, lid);
635 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
636 		} else {
637 			uasm_il_b(p, r, lid);
638 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
639 		}
640 	}
641 }
642 
643 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
644 						 struct uasm_label **l,
645 						 struct uasm_reloc **r,
646 						 unsigned int tmp,
647 						 enum tlb_write_entry wmode,
648 						 int restore_scratch)
649 {
650 	/* Set huge page tlb entry size */
651 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
652 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
653 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
654 
655 	build_tlb_write_entry(p, l, r, wmode);
656 
657 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
658 }
659 
660 /*
661  * Check if Huge PTE is present, if so then jump to LABEL.
662  */
663 static void __cpuinit
664 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
665 		unsigned int pmd, int lid)
666 {
667 	UASM_i_LW(p, tmp, 0, pmd);
668 	if (use_bbit_insns()) {
669 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
670 	} else {
671 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
672 		uasm_il_bnez(p, r, tmp, lid);
673 	}
674 }
675 
676 static __cpuinit void build_huge_update_entries(u32 **p,
677 						unsigned int pte,
678 						unsigned int tmp)
679 {
680 	int small_sequence;
681 
682 	/*
683 	 * A huge PTE describes an area the size of the
684 	 * configured huge page size. This is twice the
685 	 * of the large TLB entry size we intend to use.
686 	 * A TLB entry half the size of the configured
687 	 * huge page size is configured into entrylo0
688 	 * and entrylo1 to cover the contiguous huge PTE
689 	 * address space.
690 	 */
691 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
692 
693 	/* We can clobber tmp.  It isn't used after this.*/
694 	if (!small_sequence)
695 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
696 
697 	build_convert_pte_to_entrylo(p, pte);
698 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
699 	/* convert to entrylo1 */
700 	if (small_sequence)
701 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
702 	else
703 		UASM_i_ADDU(p, pte, pte, tmp);
704 
705 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
706 }
707 
708 static __cpuinit void build_huge_handler_tail(u32 **p,
709 					      struct uasm_reloc **r,
710 					      struct uasm_label **l,
711 					      unsigned int pte,
712 					      unsigned int ptr)
713 {
714 #ifdef CONFIG_SMP
715 	UASM_i_SC(p, pte, 0, ptr);
716 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
717 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
718 #else
719 	UASM_i_SW(p, pte, 0, ptr);
720 #endif
721 	build_huge_update_entries(p, pte, ptr);
722 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
723 }
724 #endif /* CONFIG_HUGETLB_PAGE */
725 
726 #ifdef CONFIG_64BIT
727 /*
728  * TMP and PTR are scratch.
729  * TMP will be clobbered, PTR will hold the pmd entry.
730  */
731 static void __cpuinit
732 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
733 		 unsigned int tmp, unsigned int ptr)
734 {
735 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
736 	long pgdc = (long)pgd_current;
737 #endif
738 	/*
739 	 * The vmalloc handling is not in the hotpath.
740 	 */
741 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
742 
743 	if (check_for_high_segbits) {
744 		/*
745 		 * The kernel currently implicitely assumes that the
746 		 * MIPS SEGBITS parameter for the processor is
747 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
748 		 * allocate virtual addresses outside the maximum
749 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
750 		 * that doesn't prevent user code from accessing the
751 		 * higher xuseg addresses.  Here, we make sure that
752 		 * everything but the lower xuseg addresses goes down
753 		 * the module_alloc/vmalloc path.
754 		 */
755 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
756 		uasm_il_bnez(p, r, ptr, label_vmalloc);
757 	} else {
758 		uasm_il_bltz(p, r, tmp, label_vmalloc);
759 	}
760 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
761 
762 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
763 	if (pgd_reg != -1) {
764 		/* pgd is in pgd_reg */
765 		UASM_i_MFC0(p, ptr, 31, pgd_reg);
766 	} else {
767 		/*
768 		 * &pgd << 11 stored in CONTEXT [23..63].
769 		 */
770 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
771 
772 		/* Clear lower 23 bits of context. */
773 		uasm_i_dins(p, ptr, 0, 0, 23);
774 
775 		/* 1 0  1 0 1  << 6  xkphys cached */
776 		uasm_i_ori(p, ptr, ptr, 0x540);
777 		uasm_i_drotr(p, ptr, ptr, 11);
778 	}
779 #elif defined(CONFIG_SMP)
780 # ifdef  CONFIG_MIPS_MT_SMTC
781 	/*
782 	 * SMTC uses TCBind value as "CPU" index
783 	 */
784 	uasm_i_mfc0(p, ptr, C0_TCBIND);
785 	uasm_i_dsrl_safe(p, ptr, ptr, 19);
786 # else
787 	/*
788 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
789 	 * stored in CONTEXT.
790 	 */
791 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
792 	uasm_i_dsrl_safe(p, ptr, ptr, 23);
793 # endif
794 	UASM_i_LA_mostly(p, tmp, pgdc);
795 	uasm_i_daddu(p, ptr, ptr, tmp);
796 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
797 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
798 #else
799 	UASM_i_LA_mostly(p, ptr, pgdc);
800 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
801 #endif
802 
803 	uasm_l_vmalloc_done(l, *p);
804 
805 	/* get pgd offset in bytes */
806 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
807 
808 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
809 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
810 #ifndef __PAGETABLE_PMD_FOLDED
811 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
812 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
813 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
814 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
815 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
816 #endif
817 }
818 
819 /*
820  * BVADDR is the faulting address, PTR is scratch.
821  * PTR will hold the pgd for vmalloc.
822  */
823 static void __cpuinit
824 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
825 			unsigned int bvaddr, unsigned int ptr,
826 			enum vmalloc64_mode mode)
827 {
828 	long swpd = (long)swapper_pg_dir;
829 	int single_insn_swpd;
830 	int did_vmalloc_branch = 0;
831 
832 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
833 
834 	uasm_l_vmalloc(l, *p);
835 
836 	if (mode != not_refill && check_for_high_segbits) {
837 		if (single_insn_swpd) {
838 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
839 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
840 			did_vmalloc_branch = 1;
841 			/* fall through */
842 		} else {
843 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
844 		}
845 	}
846 	if (!did_vmalloc_branch) {
847 		if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
848 			uasm_il_b(p, r, label_vmalloc_done);
849 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
850 		} else {
851 			UASM_i_LA_mostly(p, ptr, swpd);
852 			uasm_il_b(p, r, label_vmalloc_done);
853 			if (uasm_in_compat_space_p(swpd))
854 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
855 			else
856 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
857 		}
858 	}
859 	if (mode != not_refill && check_for_high_segbits) {
860 		uasm_l_large_segbits_fault(l, *p);
861 		/*
862 		 * We get here if we are an xsseg address, or if we are
863 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
864 		 *
865 		 * Ignoring xsseg (assume disabled so would generate
866 		 * (address errors?), the only remaining possibility
867 		 * is the upper xuseg addresses.  On processors with
868 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
869 		 * addresses would have taken an address error. We try
870 		 * to mimic that here by taking a load/istream page
871 		 * fault.
872 		 */
873 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
874 		uasm_i_jr(p, ptr);
875 
876 		if (mode == refill_scratch) {
877 			if (scratch_reg > 0)
878 				UASM_i_MFC0(p, 1, 31, scratch_reg);
879 			else
880 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
881 		} else {
882 			uasm_i_nop(p);
883 		}
884 	}
885 }
886 
887 #else /* !CONFIG_64BIT */
888 
889 /*
890  * TMP and PTR are scratch.
891  * TMP will be clobbered, PTR will hold the pgd entry.
892  */
893 static void __cpuinit __maybe_unused
894 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
895 {
896 	long pgdc = (long)pgd_current;
897 
898 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
899 #ifdef CONFIG_SMP
900 #ifdef  CONFIG_MIPS_MT_SMTC
901 	/*
902 	 * SMTC uses TCBind value as "CPU" index
903 	 */
904 	uasm_i_mfc0(p, ptr, C0_TCBIND);
905 	UASM_i_LA_mostly(p, tmp, pgdc);
906 	uasm_i_srl(p, ptr, ptr, 19);
907 #else
908 	/*
909 	 * smp_processor_id() << 3 is stored in CONTEXT.
910          */
911 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
912 	UASM_i_LA_mostly(p, tmp, pgdc);
913 	uasm_i_srl(p, ptr, ptr, 23);
914 #endif
915 	uasm_i_addu(p, ptr, tmp, ptr);
916 #else
917 	UASM_i_LA_mostly(p, ptr, pgdc);
918 #endif
919 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
920 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
921 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
922 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
923 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
924 }
925 
926 #endif /* !CONFIG_64BIT */
927 
928 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
929 {
930 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
931 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
932 
933 	switch (current_cpu_type()) {
934 	case CPU_VR41XX:
935 	case CPU_VR4111:
936 	case CPU_VR4121:
937 	case CPU_VR4122:
938 	case CPU_VR4131:
939 	case CPU_VR4181:
940 	case CPU_VR4181A:
941 	case CPU_VR4133:
942 		shift += 2;
943 		break;
944 
945 	default:
946 		break;
947 	}
948 
949 	if (shift)
950 		UASM_i_SRL(p, ctx, ctx, shift);
951 	uasm_i_andi(p, ctx, ctx, mask);
952 }
953 
954 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
955 {
956 	/*
957 	 * Bug workaround for the Nevada. It seems as if under certain
958 	 * circumstances the move from cp0_context might produce a
959 	 * bogus result when the mfc0 instruction and its consumer are
960 	 * in a different cacheline or a load instruction, probably any
961 	 * memory reference, is between them.
962 	 */
963 	switch (current_cpu_type()) {
964 	case CPU_NEVADA:
965 		UASM_i_LW(p, ptr, 0, ptr);
966 		GET_CONTEXT(p, tmp); /* get context reg */
967 		break;
968 
969 	default:
970 		GET_CONTEXT(p, tmp); /* get context reg */
971 		UASM_i_LW(p, ptr, 0, ptr);
972 		break;
973 	}
974 
975 	build_adjust_context(p, tmp);
976 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
977 }
978 
979 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
980 					unsigned int ptep)
981 {
982 	/*
983 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
984 	 * Kernel is a special case. Only a few CPUs use it.
985 	 */
986 #ifdef CONFIG_64BIT_PHYS_ADDR
987 	if (cpu_has_64bits) {
988 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
989 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
990 		if (kernel_uses_smartmips_rixi) {
991 			UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
992 			UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
993 			UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
994 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
995 			UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
996 		} else {
997 			uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
998 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
999 			uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1000 		}
1001 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1002 	} else {
1003 		int pte_off_even = sizeof(pte_t) / 2;
1004 		int pte_off_odd = pte_off_even + sizeof(pte_t);
1005 
1006 		/* The pte entries are pre-shifted */
1007 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1008 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1009 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1010 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1011 	}
1012 #else
1013 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1014 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1015 	if (r45k_bvahwbug())
1016 		build_tlb_probe_entry(p);
1017 	if (kernel_uses_smartmips_rixi) {
1018 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
1019 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
1020 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1021 		if (r4k_250MHZhwbug())
1022 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1023 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1024 		UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1025 	} else {
1026 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1027 		if (r4k_250MHZhwbug())
1028 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1029 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1030 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1031 		if (r45k_bvahwbug())
1032 			uasm_i_mfc0(p, tmp, C0_INDEX);
1033 	}
1034 	if (r4k_250MHZhwbug())
1035 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1036 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1037 #endif
1038 }
1039 
1040 struct mips_huge_tlb_info {
1041 	int huge_pte;
1042 	int restore_scratch;
1043 };
1044 
1045 static struct mips_huge_tlb_info __cpuinit
1046 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1047 			       struct uasm_reloc **r, unsigned int tmp,
1048 			       unsigned int ptr, int c0_scratch)
1049 {
1050 	struct mips_huge_tlb_info rv;
1051 	unsigned int even, odd;
1052 	int vmalloc_branch_delay_filled = 0;
1053 	const int scratch = 1; /* Our extra working register */
1054 
1055 	rv.huge_pte = scratch;
1056 	rv.restore_scratch = 0;
1057 
1058 	if (check_for_high_segbits) {
1059 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1060 
1061 		if (pgd_reg != -1)
1062 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1063 		else
1064 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1065 
1066 		if (c0_scratch >= 0)
1067 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1068 		else
1069 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1070 
1071 		uasm_i_dsrl_safe(p, scratch, tmp,
1072 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1073 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1074 
1075 		if (pgd_reg == -1) {
1076 			vmalloc_branch_delay_filled = 1;
1077 			/* Clear lower 23 bits of context. */
1078 			uasm_i_dins(p, ptr, 0, 0, 23);
1079 		}
1080 	} else {
1081 		if (pgd_reg != -1)
1082 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1083 		else
1084 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1085 
1086 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1087 
1088 		if (c0_scratch >= 0)
1089 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1090 		else
1091 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1092 
1093 		if (pgd_reg == -1)
1094 			/* Clear lower 23 bits of context. */
1095 			uasm_i_dins(p, ptr, 0, 0, 23);
1096 
1097 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1098 	}
1099 
1100 	if (pgd_reg == -1) {
1101 		vmalloc_branch_delay_filled = 1;
1102 		/* 1 0  1 0 1  << 6  xkphys cached */
1103 		uasm_i_ori(p, ptr, ptr, 0x540);
1104 		uasm_i_drotr(p, ptr, ptr, 11);
1105 	}
1106 
1107 #ifdef __PAGETABLE_PMD_FOLDED
1108 #define LOC_PTEP scratch
1109 #else
1110 #define LOC_PTEP ptr
1111 #endif
1112 
1113 	if (!vmalloc_branch_delay_filled)
1114 		/* get pgd offset in bytes */
1115 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1116 
1117 	uasm_l_vmalloc_done(l, *p);
1118 
1119 	/*
1120 	 *                         tmp          ptr
1121 	 * fall-through case =   badvaddr  *pgd_current
1122 	 * vmalloc case      =   badvaddr  swapper_pg_dir
1123 	 */
1124 
1125 	if (vmalloc_branch_delay_filled)
1126 		/* get pgd offset in bytes */
1127 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1128 
1129 #ifdef __PAGETABLE_PMD_FOLDED
1130 	GET_CONTEXT(p, tmp); /* get context reg */
1131 #endif
1132 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1133 
1134 	if (use_lwx_insns()) {
1135 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1136 	} else {
1137 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1138 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1139 	}
1140 
1141 #ifndef __PAGETABLE_PMD_FOLDED
1142 	/* get pmd offset in bytes */
1143 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1144 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1145 	GET_CONTEXT(p, tmp); /* get context reg */
1146 
1147 	if (use_lwx_insns()) {
1148 		UASM_i_LWX(p, scratch, scratch, ptr);
1149 	} else {
1150 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1151 		UASM_i_LW(p, scratch, 0, ptr);
1152 	}
1153 #endif
1154 	/* Adjust the context during the load latency. */
1155 	build_adjust_context(p, tmp);
1156 
1157 #ifdef CONFIG_HUGETLB_PAGE
1158 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1159 	/*
1160 	 * The in the LWX case we don't want to do the load in the
1161 	 * delay slot.  It cannot issue in the same cycle and may be
1162 	 * speculative and unneeded.
1163 	 */
1164 	if (use_lwx_insns())
1165 		uasm_i_nop(p);
1166 #endif /* CONFIG_HUGETLB_PAGE */
1167 
1168 
1169 	/* build_update_entries */
1170 	if (use_lwx_insns()) {
1171 		even = ptr;
1172 		odd = tmp;
1173 		UASM_i_LWX(p, even, scratch, tmp);
1174 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1175 		UASM_i_LWX(p, odd, scratch, tmp);
1176 	} else {
1177 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1178 		even = tmp;
1179 		odd = ptr;
1180 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1181 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1182 	}
1183 	if (kernel_uses_smartmips_rixi) {
1184 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_NO_EXEC));
1185 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_NO_EXEC));
1186 		uasm_i_drotr(p, even, even,
1187 			     ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1188 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1189 		uasm_i_drotr(p, odd, odd,
1190 			     ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1191 	} else {
1192 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1193 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1194 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1195 	}
1196 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1197 
1198 	if (c0_scratch >= 0) {
1199 		UASM_i_MFC0(p, scratch, 31, c0_scratch);
1200 		build_tlb_write_entry(p, l, r, tlb_random);
1201 		uasm_l_leave(l, *p);
1202 		rv.restore_scratch = 1;
1203 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1204 		build_tlb_write_entry(p, l, r, tlb_random);
1205 		uasm_l_leave(l, *p);
1206 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1207 	} else {
1208 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1209 		build_tlb_write_entry(p, l, r, tlb_random);
1210 		uasm_l_leave(l, *p);
1211 		rv.restore_scratch = 1;
1212 	}
1213 
1214 	uasm_i_eret(p); /* return from trap */
1215 
1216 	return rv;
1217 }
1218 
1219 /*
1220  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1221  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1222  * slots before the XTLB refill exception handler which belong to the
1223  * unused TLB refill exception.
1224  */
1225 #define MIPS64_REFILL_INSNS 32
1226 
1227 static void __cpuinit build_r4000_tlb_refill_handler(void)
1228 {
1229 	u32 *p = tlb_handler;
1230 	struct uasm_label *l = labels;
1231 	struct uasm_reloc *r = relocs;
1232 	u32 *f;
1233 	unsigned int final_len;
1234 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1235 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1236 
1237 	memset(tlb_handler, 0, sizeof(tlb_handler));
1238 	memset(labels, 0, sizeof(labels));
1239 	memset(relocs, 0, sizeof(relocs));
1240 	memset(final_handler, 0, sizeof(final_handler));
1241 
1242 	if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1243 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1244 							  scratch_reg);
1245 		vmalloc_mode = refill_scratch;
1246 	} else {
1247 		htlb_info.huge_pte = K0;
1248 		htlb_info.restore_scratch = 0;
1249 		vmalloc_mode = refill_noscratch;
1250 		/*
1251 		 * create the plain linear handler
1252 		 */
1253 		if (bcm1250_m3_war()) {
1254 			unsigned int segbits = 44;
1255 
1256 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1257 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1258 			uasm_i_xor(&p, K0, K0, K1);
1259 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1260 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1261 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1262 			uasm_i_or(&p, K0, K0, K1);
1263 			uasm_il_bnez(&p, &r, K0, label_leave);
1264 			/* No need for uasm_i_nop */
1265 		}
1266 
1267 #ifdef CONFIG_64BIT
1268 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1269 #else
1270 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1271 #endif
1272 
1273 #ifdef CONFIG_HUGETLB_PAGE
1274 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1275 #endif
1276 
1277 		build_get_ptep(&p, K0, K1);
1278 		build_update_entries(&p, K0, K1);
1279 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1280 		uasm_l_leave(&l, p);
1281 		uasm_i_eret(&p); /* return from trap */
1282 	}
1283 #ifdef CONFIG_HUGETLB_PAGE
1284 	uasm_l_tlb_huge_update(&l, p);
1285 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1286 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1287 				   htlb_info.restore_scratch);
1288 #endif
1289 
1290 #ifdef CONFIG_64BIT
1291 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1292 #endif
1293 
1294 	/*
1295 	 * Overflow check: For the 64bit handler, we need at least one
1296 	 * free instruction slot for the wrap-around branch. In worst
1297 	 * case, if the intended insertion point is a delay slot, we
1298 	 * need three, with the second nop'ed and the third being
1299 	 * unused.
1300 	 */
1301 	/* Loongson2 ebase is different than r4k, we have more space */
1302 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1303 	if ((p - tlb_handler) > 64)
1304 		panic("TLB refill handler space exceeded");
1305 #else
1306 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1307 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1308 		&& uasm_insn_has_bdelay(relocs,
1309 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
1310 		panic("TLB refill handler space exceeded");
1311 #endif
1312 
1313 	/*
1314 	 * Now fold the handler in the TLB refill handler space.
1315 	 */
1316 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1317 	f = final_handler;
1318 	/* Simplest case, just copy the handler. */
1319 	uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1320 	final_len = p - tlb_handler;
1321 #else /* CONFIG_64BIT */
1322 	f = final_handler + MIPS64_REFILL_INSNS;
1323 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1324 		/* Just copy the handler. */
1325 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1326 		final_len = p - tlb_handler;
1327 	} else {
1328 #if defined(CONFIG_HUGETLB_PAGE)
1329 		const enum label_id ls = label_tlb_huge_update;
1330 #else
1331 		const enum label_id ls = label_vmalloc;
1332 #endif
1333 		u32 *split;
1334 		int ov = 0;
1335 		int i;
1336 
1337 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1338 			;
1339 		BUG_ON(i == ARRAY_SIZE(labels));
1340 		split = labels[i].addr;
1341 
1342 		/*
1343 		 * See if we have overflown one way or the other.
1344 		 */
1345 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1346 		    split < p - MIPS64_REFILL_INSNS)
1347 			ov = 1;
1348 
1349 		if (ov) {
1350 			/*
1351 			 * Split two instructions before the end.  One
1352 			 * for the branch and one for the instruction
1353 			 * in the delay slot.
1354 			 */
1355 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1356 
1357 			/*
1358 			 * If the branch would fall in a delay slot,
1359 			 * we must back up an additional instruction
1360 			 * so that it is no longer in a delay slot.
1361 			 */
1362 			if (uasm_insn_has_bdelay(relocs, split - 1))
1363 				split--;
1364 		}
1365 		/* Copy first part of the handler. */
1366 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1367 		f += split - tlb_handler;
1368 
1369 		if (ov) {
1370 			/* Insert branch. */
1371 			uasm_l_split(&l, final_handler);
1372 			uasm_il_b(&f, &r, label_split);
1373 			if (uasm_insn_has_bdelay(relocs, split))
1374 				uasm_i_nop(&f);
1375 			else {
1376 				uasm_copy_handler(relocs, labels,
1377 						  split, split + 1, f);
1378 				uasm_move_labels(labels, f, f + 1, -1);
1379 				f++;
1380 				split++;
1381 			}
1382 		}
1383 
1384 		/* Copy the rest of the handler. */
1385 		uasm_copy_handler(relocs, labels, split, p, final_handler);
1386 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1387 			    (p - split);
1388 	}
1389 #endif /* CONFIG_64BIT */
1390 
1391 	uasm_resolve_relocs(relocs, labels);
1392 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1393 		 final_len);
1394 
1395 	memcpy((void *)ebase, final_handler, 0x100);
1396 
1397 	dump_handler((u32 *)ebase, 64);
1398 }
1399 
1400 /*
1401  * 128 instructions for the fastpath handler is generous and should
1402  * never be exceeded.
1403  */
1404 #define FASTPATH_SIZE 128
1405 
1406 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1407 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1408 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1409 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1410 u32 tlbmiss_handler_setup_pgd[16] __cacheline_aligned;
1411 
1412 static void __cpuinit build_r4000_setup_pgd(void)
1413 {
1414 	const int a0 = 4;
1415 	const int a1 = 5;
1416 	u32 *p = tlbmiss_handler_setup_pgd;
1417 	struct uasm_label *l = labels;
1418 	struct uasm_reloc *r = relocs;
1419 
1420 	memset(tlbmiss_handler_setup_pgd, 0, sizeof(tlbmiss_handler_setup_pgd));
1421 	memset(labels, 0, sizeof(labels));
1422 	memset(relocs, 0, sizeof(relocs));
1423 
1424 	pgd_reg = allocate_kscratch();
1425 
1426 	if (pgd_reg == -1) {
1427 		/* PGD << 11 in c0_Context */
1428 		/*
1429 		 * If it is a ckseg0 address, convert to a physical
1430 		 * address.  Shifting right by 29 and adding 4 will
1431 		 * result in zero for these addresses.
1432 		 *
1433 		 */
1434 		UASM_i_SRA(&p, a1, a0, 29);
1435 		UASM_i_ADDIU(&p, a1, a1, 4);
1436 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1437 		uasm_i_nop(&p);
1438 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1439 		uasm_l_tlbl_goaround1(&l, p);
1440 		UASM_i_SLL(&p, a0, a0, 11);
1441 		uasm_i_jr(&p, 31);
1442 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1443 	} else {
1444 		/* PGD in c0_KScratch */
1445 		uasm_i_jr(&p, 31);
1446 		UASM_i_MTC0(&p, a0, 31, pgd_reg);
1447 	}
1448 	if (p - tlbmiss_handler_setup_pgd > ARRAY_SIZE(tlbmiss_handler_setup_pgd))
1449 		panic("tlbmiss_handler_setup_pgd space exceeded");
1450 	uasm_resolve_relocs(relocs, labels);
1451 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1452 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1453 
1454 	dump_handler(tlbmiss_handler_setup_pgd,
1455 		     ARRAY_SIZE(tlbmiss_handler_setup_pgd));
1456 }
1457 #endif
1458 
1459 static void __cpuinit
1460 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1461 {
1462 #ifdef CONFIG_SMP
1463 # ifdef CONFIG_64BIT_PHYS_ADDR
1464 	if (cpu_has_64bits)
1465 		uasm_i_lld(p, pte, 0, ptr);
1466 	else
1467 # endif
1468 		UASM_i_LL(p, pte, 0, ptr);
1469 #else
1470 # ifdef CONFIG_64BIT_PHYS_ADDR
1471 	if (cpu_has_64bits)
1472 		uasm_i_ld(p, pte, 0, ptr);
1473 	else
1474 # endif
1475 		UASM_i_LW(p, pte, 0, ptr);
1476 #endif
1477 }
1478 
1479 static void __cpuinit
1480 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1481 	unsigned int mode)
1482 {
1483 #ifdef CONFIG_64BIT_PHYS_ADDR
1484 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1485 #endif
1486 
1487 	uasm_i_ori(p, pte, pte, mode);
1488 #ifdef CONFIG_SMP
1489 # ifdef CONFIG_64BIT_PHYS_ADDR
1490 	if (cpu_has_64bits)
1491 		uasm_i_scd(p, pte, 0, ptr);
1492 	else
1493 # endif
1494 		UASM_i_SC(p, pte, 0, ptr);
1495 
1496 	if (r10000_llsc_war())
1497 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1498 	else
1499 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1500 
1501 # ifdef CONFIG_64BIT_PHYS_ADDR
1502 	if (!cpu_has_64bits) {
1503 		/* no uasm_i_nop needed */
1504 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1505 		uasm_i_ori(p, pte, pte, hwmode);
1506 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1507 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1508 		/* no uasm_i_nop needed */
1509 		uasm_i_lw(p, pte, 0, ptr);
1510 	} else
1511 		uasm_i_nop(p);
1512 # else
1513 	uasm_i_nop(p);
1514 # endif
1515 #else
1516 # ifdef CONFIG_64BIT_PHYS_ADDR
1517 	if (cpu_has_64bits)
1518 		uasm_i_sd(p, pte, 0, ptr);
1519 	else
1520 # endif
1521 		UASM_i_SW(p, pte, 0, ptr);
1522 
1523 # ifdef CONFIG_64BIT_PHYS_ADDR
1524 	if (!cpu_has_64bits) {
1525 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1526 		uasm_i_ori(p, pte, pte, hwmode);
1527 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1528 		uasm_i_lw(p, pte, 0, ptr);
1529 	}
1530 # endif
1531 #endif
1532 }
1533 
1534 /*
1535  * Check if PTE is present, if not then jump to LABEL. PTR points to
1536  * the page table where this PTE is located, PTE will be re-loaded
1537  * with it's original value.
1538  */
1539 static void __cpuinit
1540 build_pte_present(u32 **p, struct uasm_reloc **r,
1541 		  int pte, int ptr, int scratch, enum label_id lid)
1542 {
1543 	int t = scratch >= 0 ? scratch : pte;
1544 
1545 	if (kernel_uses_smartmips_rixi) {
1546 		if (use_bbit_insns()) {
1547 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1548 			uasm_i_nop(p);
1549 		} else {
1550 			uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1551 			uasm_il_beqz(p, r, t, lid);
1552 			if (pte == t)
1553 				/* You lose the SMP race :-(*/
1554 				iPTE_LW(p, pte, ptr);
1555 		}
1556 	} else {
1557 		uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1558 		uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1559 		uasm_il_bnez(p, r, t, lid);
1560 		if (pte == t)
1561 			/* You lose the SMP race :-(*/
1562 			iPTE_LW(p, pte, ptr);
1563 	}
1564 }
1565 
1566 /* Make PTE valid, store result in PTR. */
1567 static void __cpuinit
1568 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1569 		 unsigned int ptr)
1570 {
1571 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1572 
1573 	iPTE_SW(p, r, pte, ptr, mode);
1574 }
1575 
1576 /*
1577  * Check if PTE can be written to, if not branch to LABEL. Regardless
1578  * restore PTE with value from PTR when done.
1579  */
1580 static void __cpuinit
1581 build_pte_writable(u32 **p, struct uasm_reloc **r,
1582 		   unsigned int pte, unsigned int ptr, int scratch,
1583 		   enum label_id lid)
1584 {
1585 	int t = scratch >= 0 ? scratch : pte;
1586 
1587 	uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1588 	uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1589 	uasm_il_bnez(p, r, t, lid);
1590 	if (pte == t)
1591 		/* You lose the SMP race :-(*/
1592 		iPTE_LW(p, pte, ptr);
1593 	else
1594 		uasm_i_nop(p);
1595 }
1596 
1597 /* Make PTE writable, update software status bits as well, then store
1598  * at PTR.
1599  */
1600 static void __cpuinit
1601 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1602 		 unsigned int ptr)
1603 {
1604 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1605 			     | _PAGE_DIRTY);
1606 
1607 	iPTE_SW(p, r, pte, ptr, mode);
1608 }
1609 
1610 /*
1611  * Check if PTE can be modified, if not branch to LABEL. Regardless
1612  * restore PTE with value from PTR when done.
1613  */
1614 static void __cpuinit
1615 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1616 		     unsigned int pte, unsigned int ptr, int scratch,
1617 		     enum label_id lid)
1618 {
1619 	if (use_bbit_insns()) {
1620 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1621 		uasm_i_nop(p);
1622 	} else {
1623 		int t = scratch >= 0 ? scratch : pte;
1624 		uasm_i_andi(p, t, pte, _PAGE_WRITE);
1625 		uasm_il_beqz(p, r, t, lid);
1626 		if (pte == t)
1627 			/* You lose the SMP race :-(*/
1628 			iPTE_LW(p, pte, ptr);
1629 	}
1630 }
1631 
1632 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1633 
1634 
1635 /*
1636  * R3000 style TLB load/store/modify handlers.
1637  */
1638 
1639 /*
1640  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1641  * Then it returns.
1642  */
1643 static void __cpuinit
1644 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1645 {
1646 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1647 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1648 	uasm_i_tlbwi(p);
1649 	uasm_i_jr(p, tmp);
1650 	uasm_i_rfe(p); /* branch delay */
1651 }
1652 
1653 /*
1654  * This places the pte into ENTRYLO0 and writes it with tlbwi
1655  * or tlbwr as appropriate.  This is because the index register
1656  * may have the probe fail bit set as a result of a trap on a
1657  * kseg2 access, i.e. without refill.  Then it returns.
1658  */
1659 static void __cpuinit
1660 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1661 			     struct uasm_reloc **r, unsigned int pte,
1662 			     unsigned int tmp)
1663 {
1664 	uasm_i_mfc0(p, tmp, C0_INDEX);
1665 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1666 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1667 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1668 	uasm_i_tlbwi(p); /* cp0 delay */
1669 	uasm_i_jr(p, tmp);
1670 	uasm_i_rfe(p); /* branch delay */
1671 	uasm_l_r3000_write_probe_fail(l, *p);
1672 	uasm_i_tlbwr(p); /* cp0 delay */
1673 	uasm_i_jr(p, tmp);
1674 	uasm_i_rfe(p); /* branch delay */
1675 }
1676 
1677 static void __cpuinit
1678 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1679 				   unsigned int ptr)
1680 {
1681 	long pgdc = (long)pgd_current;
1682 
1683 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1684 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1685 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1686 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1687 	uasm_i_sll(p, pte, pte, 2);
1688 	uasm_i_addu(p, ptr, ptr, pte);
1689 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1690 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1691 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1692 	uasm_i_addu(p, ptr, ptr, pte);
1693 	uasm_i_lw(p, pte, 0, ptr);
1694 	uasm_i_tlbp(p); /* load delay */
1695 }
1696 
1697 static void __cpuinit build_r3000_tlb_load_handler(void)
1698 {
1699 	u32 *p = handle_tlbl;
1700 	struct uasm_label *l = labels;
1701 	struct uasm_reloc *r = relocs;
1702 
1703 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1704 	memset(labels, 0, sizeof(labels));
1705 	memset(relocs, 0, sizeof(relocs));
1706 
1707 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1708 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1709 	uasm_i_nop(&p); /* load delay */
1710 	build_make_valid(&p, &r, K0, K1);
1711 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1712 
1713 	uasm_l_nopage_tlbl(&l, p);
1714 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1715 	uasm_i_nop(&p);
1716 
1717 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1718 		panic("TLB load handler fastpath space exceeded");
1719 
1720 	uasm_resolve_relocs(relocs, labels);
1721 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1722 		 (unsigned int)(p - handle_tlbl));
1723 
1724 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1725 }
1726 
1727 static void __cpuinit build_r3000_tlb_store_handler(void)
1728 {
1729 	u32 *p = handle_tlbs;
1730 	struct uasm_label *l = labels;
1731 	struct uasm_reloc *r = relocs;
1732 
1733 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1734 	memset(labels, 0, sizeof(labels));
1735 	memset(relocs, 0, sizeof(relocs));
1736 
1737 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1738 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1739 	uasm_i_nop(&p); /* load delay */
1740 	build_make_write(&p, &r, K0, K1);
1741 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1742 
1743 	uasm_l_nopage_tlbs(&l, p);
1744 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1745 	uasm_i_nop(&p);
1746 
1747 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1748 		panic("TLB store handler fastpath space exceeded");
1749 
1750 	uasm_resolve_relocs(relocs, labels);
1751 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1752 		 (unsigned int)(p - handle_tlbs));
1753 
1754 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1755 }
1756 
1757 static void __cpuinit build_r3000_tlb_modify_handler(void)
1758 {
1759 	u32 *p = handle_tlbm;
1760 	struct uasm_label *l = labels;
1761 	struct uasm_reloc *r = relocs;
1762 
1763 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1764 	memset(labels, 0, sizeof(labels));
1765 	memset(relocs, 0, sizeof(relocs));
1766 
1767 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1768 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1769 	uasm_i_nop(&p); /* load delay */
1770 	build_make_write(&p, &r, K0, K1);
1771 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1772 
1773 	uasm_l_nopage_tlbm(&l, p);
1774 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1775 	uasm_i_nop(&p);
1776 
1777 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1778 		panic("TLB modify handler fastpath space exceeded");
1779 
1780 	uasm_resolve_relocs(relocs, labels);
1781 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1782 		 (unsigned int)(p - handle_tlbm));
1783 
1784 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1785 }
1786 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1787 
1788 /*
1789  * R4000 style TLB load/store/modify handlers.
1790  */
1791 static struct work_registers __cpuinit
1792 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1793 				   struct uasm_reloc **r)
1794 {
1795 	struct work_registers wr = build_get_work_registers(p);
1796 
1797 #ifdef CONFIG_64BIT
1798 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1799 #else
1800 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1801 #endif
1802 
1803 #ifdef CONFIG_HUGETLB_PAGE
1804 	/*
1805 	 * For huge tlb entries, pmd doesn't contain an address but
1806 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1807 	 * see if we need to jump to huge tlb processing.
1808 	 */
1809 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1810 #endif
1811 
1812 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1813 	UASM_i_LW(p, wr.r2, 0, wr.r2);
1814 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1815 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1816 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1817 
1818 #ifdef CONFIG_SMP
1819 	uasm_l_smp_pgtable_change(l, *p);
1820 #endif
1821 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1822 	if (!m4kc_tlbp_war())
1823 		build_tlb_probe_entry(p);
1824 	return wr;
1825 }
1826 
1827 static void __cpuinit
1828 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1829 				   struct uasm_reloc **r, unsigned int tmp,
1830 				   unsigned int ptr)
1831 {
1832 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1833 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1834 	build_update_entries(p, tmp, ptr);
1835 	build_tlb_write_entry(p, l, r, tlb_indexed);
1836 	uasm_l_leave(l, *p);
1837 	build_restore_work_registers(p);
1838 	uasm_i_eret(p); /* return from trap */
1839 
1840 #ifdef CONFIG_64BIT
1841 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1842 #endif
1843 }
1844 
1845 static void __cpuinit build_r4000_tlb_load_handler(void)
1846 {
1847 	u32 *p = handle_tlbl;
1848 	struct uasm_label *l = labels;
1849 	struct uasm_reloc *r = relocs;
1850 	struct work_registers wr;
1851 
1852 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1853 	memset(labels, 0, sizeof(labels));
1854 	memset(relocs, 0, sizeof(relocs));
1855 
1856 	if (bcm1250_m3_war()) {
1857 		unsigned int segbits = 44;
1858 
1859 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1860 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1861 		uasm_i_xor(&p, K0, K0, K1);
1862 		uasm_i_dsrl_safe(&p, K1, K0, 62);
1863 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1864 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1865 		uasm_i_or(&p, K0, K0, K1);
1866 		uasm_il_bnez(&p, &r, K0, label_leave);
1867 		/* No need for uasm_i_nop */
1868 	}
1869 
1870 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1871 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1872 	if (m4kc_tlbp_war())
1873 		build_tlb_probe_entry(&p);
1874 
1875 	if (kernel_uses_smartmips_rixi) {
1876 		/*
1877 		 * If the page is not _PAGE_VALID, RI or XI could not
1878 		 * have triggered it.  Skip the expensive test..
1879 		 */
1880 		if (use_bbit_insns()) {
1881 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1882 				      label_tlbl_goaround1);
1883 		} else {
1884 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1885 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1886 		}
1887 		uasm_i_nop(&p);
1888 
1889 		uasm_i_tlbr(&p);
1890 		/* Examine  entrylo 0 or 1 based on ptr. */
1891 		if (use_bbit_insns()) {
1892 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1893 		} else {
1894 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1895 			uasm_i_beqz(&p, wr.r3, 8);
1896 		}
1897 		/* load it in the delay slot*/
1898 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1899 		/* load it if ptr is odd */
1900 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1901 		/*
1902 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1903 		 * XI must have triggered it.
1904 		 */
1905 		if (use_bbit_insns()) {
1906 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1907 			uasm_i_nop(&p);
1908 			uasm_l_tlbl_goaround1(&l, p);
1909 		} else {
1910 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1911 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1912 			uasm_i_nop(&p);
1913 		}
1914 		uasm_l_tlbl_goaround1(&l, p);
1915 	}
1916 	build_make_valid(&p, &r, wr.r1, wr.r2);
1917 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1918 
1919 #ifdef CONFIG_HUGETLB_PAGE
1920 	/*
1921 	 * This is the entry point when build_r4000_tlbchange_handler_head
1922 	 * spots a huge page.
1923 	 */
1924 	uasm_l_tlb_huge_update(&l, p);
1925 	iPTE_LW(&p, wr.r1, wr.r2);
1926 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1927 	build_tlb_probe_entry(&p);
1928 
1929 	if (kernel_uses_smartmips_rixi) {
1930 		/*
1931 		 * If the page is not _PAGE_VALID, RI or XI could not
1932 		 * have triggered it.  Skip the expensive test..
1933 		 */
1934 		if (use_bbit_insns()) {
1935 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1936 				      label_tlbl_goaround2);
1937 		} else {
1938 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1939 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1940 		}
1941 		uasm_i_nop(&p);
1942 
1943 		uasm_i_tlbr(&p);
1944 		/* Examine  entrylo 0 or 1 based on ptr. */
1945 		if (use_bbit_insns()) {
1946 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1947 		} else {
1948 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1949 			uasm_i_beqz(&p, wr.r3, 8);
1950 		}
1951 		/* load it in the delay slot*/
1952 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1953 		/* load it if ptr is odd */
1954 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1955 		/*
1956 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1957 		 * XI must have triggered it.
1958 		 */
1959 		if (use_bbit_insns()) {
1960 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
1961 		} else {
1962 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1963 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1964 		}
1965 		if (PM_DEFAULT_MASK == 0)
1966 			uasm_i_nop(&p);
1967 		/*
1968 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
1969 		 * it is restored in build_huge_tlb_write_entry.
1970 		 */
1971 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
1972 
1973 		uasm_l_tlbl_goaround2(&l, p);
1974 	}
1975 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
1976 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
1977 #endif
1978 
1979 	uasm_l_nopage_tlbl(&l, p);
1980 	build_restore_work_registers(&p);
1981 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1982 	uasm_i_nop(&p);
1983 
1984 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1985 		panic("TLB load handler fastpath space exceeded");
1986 
1987 	uasm_resolve_relocs(relocs, labels);
1988 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1989 		 (unsigned int)(p - handle_tlbl));
1990 
1991 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1992 }
1993 
1994 static void __cpuinit build_r4000_tlb_store_handler(void)
1995 {
1996 	u32 *p = handle_tlbs;
1997 	struct uasm_label *l = labels;
1998 	struct uasm_reloc *r = relocs;
1999 	struct work_registers wr;
2000 
2001 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
2002 	memset(labels, 0, sizeof(labels));
2003 	memset(relocs, 0, sizeof(relocs));
2004 
2005 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2006 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2007 	if (m4kc_tlbp_war())
2008 		build_tlb_probe_entry(&p);
2009 	build_make_write(&p, &r, wr.r1, wr.r2);
2010 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2011 
2012 #ifdef CONFIG_HUGETLB_PAGE
2013 	/*
2014 	 * This is the entry point when
2015 	 * build_r4000_tlbchange_handler_head spots a huge page.
2016 	 */
2017 	uasm_l_tlb_huge_update(&l, p);
2018 	iPTE_LW(&p, wr.r1, wr.r2);
2019 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2020 	build_tlb_probe_entry(&p);
2021 	uasm_i_ori(&p, wr.r1, wr.r1,
2022 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2023 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2024 #endif
2025 
2026 	uasm_l_nopage_tlbs(&l, p);
2027 	build_restore_work_registers(&p);
2028 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2029 	uasm_i_nop(&p);
2030 
2031 	if ((p - handle_tlbs) > FASTPATH_SIZE)
2032 		panic("TLB store handler fastpath space exceeded");
2033 
2034 	uasm_resolve_relocs(relocs, labels);
2035 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2036 		 (unsigned int)(p - handle_tlbs));
2037 
2038 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
2039 }
2040 
2041 static void __cpuinit build_r4000_tlb_modify_handler(void)
2042 {
2043 	u32 *p = handle_tlbm;
2044 	struct uasm_label *l = labels;
2045 	struct uasm_reloc *r = relocs;
2046 	struct work_registers wr;
2047 
2048 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
2049 	memset(labels, 0, sizeof(labels));
2050 	memset(relocs, 0, sizeof(relocs));
2051 
2052 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2053 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2054 	if (m4kc_tlbp_war())
2055 		build_tlb_probe_entry(&p);
2056 	/* Present and writable bits set, set accessed and dirty bits. */
2057 	build_make_write(&p, &r, wr.r1, wr.r2);
2058 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2059 
2060 #ifdef CONFIG_HUGETLB_PAGE
2061 	/*
2062 	 * This is the entry point when
2063 	 * build_r4000_tlbchange_handler_head spots a huge page.
2064 	 */
2065 	uasm_l_tlb_huge_update(&l, p);
2066 	iPTE_LW(&p, wr.r1, wr.r2);
2067 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2068 	build_tlb_probe_entry(&p);
2069 	uasm_i_ori(&p, wr.r1, wr.r1,
2070 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2071 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2072 #endif
2073 
2074 	uasm_l_nopage_tlbm(&l, p);
2075 	build_restore_work_registers(&p);
2076 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2077 	uasm_i_nop(&p);
2078 
2079 	if ((p - handle_tlbm) > FASTPATH_SIZE)
2080 		panic("TLB modify handler fastpath space exceeded");
2081 
2082 	uasm_resolve_relocs(relocs, labels);
2083 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2084 		 (unsigned int)(p - handle_tlbm));
2085 
2086 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
2087 }
2088 
2089 void __cpuinit build_tlb_refill_handler(void)
2090 {
2091 	/*
2092 	 * The refill handler is generated per-CPU, multi-node systems
2093 	 * may have local storage for it. The other handlers are only
2094 	 * needed once.
2095 	 */
2096 	static int run_once = 0;
2097 
2098 #ifdef CONFIG_64BIT
2099 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2100 #endif
2101 
2102 	switch (current_cpu_type()) {
2103 	case CPU_R2000:
2104 	case CPU_R3000:
2105 	case CPU_R3000A:
2106 	case CPU_R3081E:
2107 	case CPU_TX3912:
2108 	case CPU_TX3922:
2109 	case CPU_TX3927:
2110 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2111 		build_r3000_tlb_refill_handler();
2112 		if (!run_once) {
2113 			build_r3000_tlb_load_handler();
2114 			build_r3000_tlb_store_handler();
2115 			build_r3000_tlb_modify_handler();
2116 			run_once++;
2117 		}
2118 #else
2119 		panic("No R3000 TLB refill handler");
2120 #endif
2121 		break;
2122 
2123 	case CPU_R6000:
2124 	case CPU_R6000A:
2125 		panic("No R6000 TLB refill handler yet");
2126 		break;
2127 
2128 	case CPU_R8000:
2129 		panic("No R8000 TLB refill handler yet");
2130 		break;
2131 
2132 	default:
2133 		if (!run_once) {
2134 			scratch_reg = allocate_kscratch();
2135 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2136 			build_r4000_setup_pgd();
2137 #endif
2138 			build_r4000_tlb_load_handler();
2139 			build_r4000_tlb_store_handler();
2140 			build_r4000_tlb_modify_handler();
2141 			run_once++;
2142 		}
2143 		build_r4000_tlb_refill_handler();
2144 	}
2145 }
2146 
2147 void __cpuinit flush_tlb_handlers(void)
2148 {
2149 	local_flush_icache_range((unsigned long)handle_tlbl,
2150 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2151 	local_flush_icache_range((unsigned long)handle_tlbs,
2152 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2153 	local_flush_icache_range((unsigned long)handle_tlbm,
2154 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2155 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2156 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2157 			   (unsigned long)tlbmiss_handler_setup_pgd + sizeof(handle_tlbm));
2158 #endif
2159 }
2160