xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision 63dc02bd)
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 #include <asm/setup.h>
36 
37 /*
38  * TLB load/store/modify handlers.
39  *
40  * Only the fastpath gets synthesized at runtime, the slowpath for
41  * do_page_fault remains normal asm.
42  */
43 extern void tlb_do_page_fault_0(void);
44 extern void tlb_do_page_fault_1(void);
45 
46 struct work_registers {
47 	int r1;
48 	int r2;
49 	int r3;
50 };
51 
52 struct tlb_reg_save {
53 	unsigned long a;
54 	unsigned long b;
55 } ____cacheline_aligned_in_smp;
56 
57 static struct tlb_reg_save handler_reg_save[NR_CPUS];
58 
59 static inline int r45k_bvahwbug(void)
60 {
61 	/* XXX: We should probe for the presence of this bug, but we don't. */
62 	return 0;
63 }
64 
65 static inline int r4k_250MHZhwbug(void)
66 {
67 	/* XXX: We should probe for the presence of this bug, but we don't. */
68 	return 0;
69 }
70 
71 static inline int __maybe_unused bcm1250_m3_war(void)
72 {
73 	return BCM1250_M3_WAR;
74 }
75 
76 static inline int __maybe_unused r10000_llsc_war(void)
77 {
78 	return R10000_LLSC_WAR;
79 }
80 
81 static int use_bbit_insns(void)
82 {
83 	switch (current_cpu_type()) {
84 	case CPU_CAVIUM_OCTEON:
85 	case CPU_CAVIUM_OCTEON_PLUS:
86 	case CPU_CAVIUM_OCTEON2:
87 		return 1;
88 	default:
89 		return 0;
90 	}
91 }
92 
93 static int use_lwx_insns(void)
94 {
95 	switch (current_cpu_type()) {
96 	case CPU_CAVIUM_OCTEON2:
97 		return 1;
98 	default:
99 		return 0;
100 	}
101 }
102 #if defined(CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE) && \
103     CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
104 static bool scratchpad_available(void)
105 {
106 	return true;
107 }
108 static int scratchpad_offset(int i)
109 {
110 	/*
111 	 * CVMSEG starts at address -32768 and extends for
112 	 * CAVIUM_OCTEON_CVMSEG_SIZE 128 byte cache lines.
113 	 */
114 	i += 1; /* Kernel use starts at the top and works down. */
115 	return CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE * 128 - (8 * i) - 32768;
116 }
117 #else
118 static bool scratchpad_available(void)
119 {
120 	return false;
121 }
122 static int scratchpad_offset(int i)
123 {
124 	BUG();
125 	/* Really unreachable, but evidently some GCC want this. */
126 	return 0;
127 }
128 #endif
129 /*
130  * Found by experiment: At least some revisions of the 4kc throw under
131  * some circumstances a machine check exception, triggered by invalid
132  * values in the index register.  Delaying the tlbp instruction until
133  * after the next branch,  plus adding an additional nop in front of
134  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
135  * why; it's not an issue caused by the core RTL.
136  *
137  */
138 static int __cpuinit m4kc_tlbp_war(void)
139 {
140 	return (current_cpu_data.processor_id & 0xffff00) ==
141 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
142 }
143 
144 /* Handle labels (which must be positive integers). */
145 enum label_id {
146 	label_second_part = 1,
147 	label_leave,
148 	label_vmalloc,
149 	label_vmalloc_done,
150 	label_tlbw_hazard,
151 	label_split,
152 	label_tlbl_goaround1,
153 	label_tlbl_goaround2,
154 	label_nopage_tlbl,
155 	label_nopage_tlbs,
156 	label_nopage_tlbm,
157 	label_smp_pgtable_change,
158 	label_r3000_write_probe_fail,
159 	label_large_segbits_fault,
160 #ifdef CONFIG_HUGETLB_PAGE
161 	label_tlb_huge_update,
162 #endif
163 };
164 
165 UASM_L_LA(_second_part)
166 UASM_L_LA(_leave)
167 UASM_L_LA(_vmalloc)
168 UASM_L_LA(_vmalloc_done)
169 UASM_L_LA(_tlbw_hazard)
170 UASM_L_LA(_split)
171 UASM_L_LA(_tlbl_goaround1)
172 UASM_L_LA(_tlbl_goaround2)
173 UASM_L_LA(_nopage_tlbl)
174 UASM_L_LA(_nopage_tlbs)
175 UASM_L_LA(_nopage_tlbm)
176 UASM_L_LA(_smp_pgtable_change)
177 UASM_L_LA(_r3000_write_probe_fail)
178 UASM_L_LA(_large_segbits_fault)
179 #ifdef CONFIG_HUGETLB_PAGE
180 UASM_L_LA(_tlb_huge_update)
181 #endif
182 
183 /*
184  * For debug purposes.
185  */
186 static inline void dump_handler(const u32 *handler, int count)
187 {
188 	int i;
189 
190 	pr_debug("\t.set push\n");
191 	pr_debug("\t.set noreorder\n");
192 
193 	for (i = 0; i < count; i++)
194 		pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
195 
196 	pr_debug("\t.set pop\n");
197 }
198 
199 /* The only general purpose registers allowed in TLB handlers. */
200 #define K0		26
201 #define K1		27
202 
203 /* Some CP0 registers */
204 #define C0_INDEX	0, 0
205 #define C0_ENTRYLO0	2, 0
206 #define C0_TCBIND	2, 2
207 #define C0_ENTRYLO1	3, 0
208 #define C0_CONTEXT	4, 0
209 #define C0_PAGEMASK	5, 0
210 #define C0_BADVADDR	8, 0
211 #define C0_ENTRYHI	10, 0
212 #define C0_EPC		14, 0
213 #define C0_XCONTEXT	20, 0
214 
215 #ifdef CONFIG_64BIT
216 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
217 #else
218 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
219 #endif
220 
221 /* The worst case length of the handler is around 18 instructions for
222  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
223  * Maximum space available is 32 instructions for R3000 and 64
224  * instructions for R4000.
225  *
226  * We deliberately chose a buffer size of 128, so we won't scribble
227  * over anything important on overflow before we panic.
228  */
229 static u32 tlb_handler[128] __cpuinitdata;
230 
231 /* simply assume worst case size for labels and relocs */
232 static struct uasm_label labels[128] __cpuinitdata;
233 static struct uasm_reloc relocs[128] __cpuinitdata;
234 
235 #ifdef CONFIG_64BIT
236 static int check_for_high_segbits __cpuinitdata;
237 #endif
238 
239 static int check_for_high_segbits __cpuinitdata;
240 
241 static unsigned int kscratch_used_mask __cpuinitdata;
242 
243 static int __cpuinit allocate_kscratch(void)
244 {
245 	int r;
246 	unsigned int a = cpu_data[0].kscratch_mask & ~kscratch_used_mask;
247 
248 	r = ffs(a);
249 
250 	if (r == 0)
251 		return -1;
252 
253 	r--; /* make it zero based */
254 
255 	kscratch_used_mask |= (1 << r);
256 
257 	return r;
258 }
259 
260 static int scratch_reg __cpuinitdata;
261 static int pgd_reg __cpuinitdata;
262 enum vmalloc64_mode {not_refill, refill_scratch, refill_noscratch};
263 
264 static struct work_registers __cpuinit build_get_work_registers(u32 **p)
265 {
266 	struct work_registers r;
267 
268 	int smp_processor_id_reg;
269 	int smp_processor_id_sel;
270 	int smp_processor_id_shift;
271 
272 	if (scratch_reg > 0) {
273 		/* Save in CPU local C0_KScratch? */
274 		UASM_i_MTC0(p, 1, 31, scratch_reg);
275 		r.r1 = K0;
276 		r.r2 = K1;
277 		r.r3 = 1;
278 		return r;
279 	}
280 
281 	if (num_possible_cpus() > 1) {
282 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
283 		smp_processor_id_shift = 51;
284 		smp_processor_id_reg = 20; /* XContext */
285 		smp_processor_id_sel = 0;
286 #else
287 # ifdef CONFIG_32BIT
288 		smp_processor_id_shift = 25;
289 		smp_processor_id_reg = 4; /* Context */
290 		smp_processor_id_sel = 0;
291 # endif
292 # ifdef CONFIG_64BIT
293 		smp_processor_id_shift = 26;
294 		smp_processor_id_reg = 4; /* Context */
295 		smp_processor_id_sel = 0;
296 # endif
297 #endif
298 		/* Get smp_processor_id */
299 		UASM_i_MFC0(p, K0, smp_processor_id_reg, smp_processor_id_sel);
300 		UASM_i_SRL_SAFE(p, K0, K0, smp_processor_id_shift);
301 
302 		/* handler_reg_save index in K0 */
303 		UASM_i_SLL(p, K0, K0, ilog2(sizeof(struct tlb_reg_save)));
304 
305 		UASM_i_LA(p, K1, (long)&handler_reg_save);
306 		UASM_i_ADDU(p, K0, K0, K1);
307 	} else {
308 		UASM_i_LA(p, K0, (long)&handler_reg_save);
309 	}
310 	/* K0 now points to save area, save $1 and $2  */
311 	UASM_i_SW(p, 1, offsetof(struct tlb_reg_save, a), K0);
312 	UASM_i_SW(p, 2, offsetof(struct tlb_reg_save, b), K0);
313 
314 	r.r1 = K1;
315 	r.r2 = 1;
316 	r.r3 = 2;
317 	return r;
318 }
319 
320 static void __cpuinit build_restore_work_registers(u32 **p)
321 {
322 	if (scratch_reg > 0) {
323 		UASM_i_MFC0(p, 1, 31, scratch_reg);
324 		return;
325 	}
326 	/* K0 already points to save area, restore $1 and $2  */
327 	UASM_i_LW(p, 1, offsetof(struct tlb_reg_save, a), K0);
328 	UASM_i_LW(p, 2, offsetof(struct tlb_reg_save, b), K0);
329 }
330 
331 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
332 
333 /*
334  * CONFIG_MIPS_PGD_C0_CONTEXT implies 64 bit and lack of pgd_current,
335  * we cannot do r3000 under these circumstances.
336  *
337  * Declare pgd_current here instead of including mmu_context.h to avoid type
338  * conflicts for tlbmiss_handler_setup_pgd
339  */
340 extern unsigned long pgd_current[];
341 
342 /*
343  * The R3000 TLB handler is simple.
344  */
345 static void __cpuinit build_r3000_tlb_refill_handler(void)
346 {
347 	long pgdc = (long)pgd_current;
348 	u32 *p;
349 
350 	memset(tlb_handler, 0, sizeof(tlb_handler));
351 	p = tlb_handler;
352 
353 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
354 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
355 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
356 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
357 	uasm_i_sll(&p, K0, K0, 2);
358 	uasm_i_addu(&p, K1, K1, K0);
359 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
360 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
361 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
362 	uasm_i_addu(&p, K1, K1, K0);
363 	uasm_i_lw(&p, K0, 0, K1);
364 	uasm_i_nop(&p); /* load delay */
365 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
366 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
367 	uasm_i_tlbwr(&p); /* cp0 delay */
368 	uasm_i_jr(&p, K1);
369 	uasm_i_rfe(&p); /* branch delay */
370 
371 	if (p > tlb_handler + 32)
372 		panic("TLB refill handler space exceeded");
373 
374 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
375 		 (unsigned int)(p - tlb_handler));
376 
377 	memcpy((void *)ebase, tlb_handler, 0x80);
378 
379 	dump_handler((u32 *)ebase, 32);
380 }
381 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
382 
383 /*
384  * The R4000 TLB handler is much more complicated. We have two
385  * consecutive handler areas with 32 instructions space each.
386  * Since they aren't used at the same time, we can overflow in the
387  * other one.To keep things simple, we first assume linear space,
388  * then we relocate it to the final handler layout as needed.
389  */
390 static u32 final_handler[64] __cpuinitdata;
391 
392 /*
393  * Hazards
394  *
395  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
396  * 2. A timing hazard exists for the TLBP instruction.
397  *
398  *      stalling_instruction
399  *      TLBP
400  *
401  * The JTLB is being read for the TLBP throughout the stall generated by the
402  * previous instruction. This is not really correct as the stalling instruction
403  * can modify the address used to access the JTLB.  The failure symptom is that
404  * the TLBP instruction will use an address created for the stalling instruction
405  * and not the address held in C0_ENHI and thus report the wrong results.
406  *
407  * The software work-around is to not allow the instruction preceding the TLBP
408  * to stall - make it an NOP or some other instruction guaranteed not to stall.
409  *
410  * Errata 2 will not be fixed.  This errata is also on the R5000.
411  *
412  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
413  */
414 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
415 {
416 	switch (current_cpu_type()) {
417 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
418 	case CPU_R4600:
419 	case CPU_R4700:
420 	case CPU_R5000:
421 	case CPU_R5000A:
422 	case CPU_NEVADA:
423 		uasm_i_nop(p);
424 		uasm_i_tlbp(p);
425 		break;
426 
427 	default:
428 		uasm_i_tlbp(p);
429 		break;
430 	}
431 }
432 
433 /*
434  * Write random or indexed TLB entry, and care about the hazards from
435  * the preceding mtc0 and for the following eret.
436  */
437 enum tlb_write_entry { tlb_random, tlb_indexed };
438 
439 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
440 					 struct uasm_reloc **r,
441 					 enum tlb_write_entry wmode)
442 {
443 	void(*tlbw)(u32 **) = NULL;
444 
445 	switch (wmode) {
446 	case tlb_random: tlbw = uasm_i_tlbwr; break;
447 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
448 	}
449 
450 	if (cpu_has_mips_r2) {
451 		if (cpu_has_mips_r2_exec_hazard)
452 			uasm_i_ehb(p);
453 		tlbw(p);
454 		return;
455 	}
456 
457 	switch (current_cpu_type()) {
458 	case CPU_R4000PC:
459 	case CPU_R4000SC:
460 	case CPU_R4000MC:
461 	case CPU_R4400PC:
462 	case CPU_R4400SC:
463 	case CPU_R4400MC:
464 		/*
465 		 * This branch uses up a mtc0 hazard nop slot and saves
466 		 * two nops after the tlbw instruction.
467 		 */
468 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
469 		tlbw(p);
470 		uasm_l_tlbw_hazard(l, *p);
471 		uasm_i_nop(p);
472 		break;
473 
474 	case CPU_R4600:
475 	case CPU_R4700:
476 	case CPU_R5000:
477 	case CPU_R5000A:
478 		uasm_i_nop(p);
479 		tlbw(p);
480 		uasm_i_nop(p);
481 		break;
482 
483 	case CPU_R4300:
484 	case CPU_5KC:
485 	case CPU_TX49XX:
486 	case CPU_PR4450:
487 	case CPU_XLR:
488 		uasm_i_nop(p);
489 		tlbw(p);
490 		break;
491 
492 	case CPU_R10000:
493 	case CPU_R12000:
494 	case CPU_R14000:
495 	case CPU_4KC:
496 	case CPU_4KEC:
497 	case CPU_SB1:
498 	case CPU_SB1A:
499 	case CPU_4KSC:
500 	case CPU_20KC:
501 	case CPU_25KF:
502 	case CPU_BMIPS32:
503 	case CPU_BMIPS3300:
504 	case CPU_BMIPS4350:
505 	case CPU_BMIPS4380:
506 	case CPU_BMIPS5000:
507 	case CPU_LOONGSON2:
508 	case CPU_R5500:
509 		if (m4kc_tlbp_war())
510 			uasm_i_nop(p);
511 	case CPU_ALCHEMY:
512 		tlbw(p);
513 		break;
514 
515 	case CPU_NEVADA:
516 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
517 		/*
518 		 * This branch uses up a mtc0 hazard nop slot and saves
519 		 * a nop after the tlbw instruction.
520 		 */
521 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
522 		tlbw(p);
523 		uasm_l_tlbw_hazard(l, *p);
524 		break;
525 
526 	case CPU_RM7000:
527 		uasm_i_nop(p);
528 		uasm_i_nop(p);
529 		uasm_i_nop(p);
530 		uasm_i_nop(p);
531 		tlbw(p);
532 		break;
533 
534 	case CPU_RM9000:
535 		/*
536 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
537 		 * use of the JTLB for instructions should not occur for 4
538 		 * cpu cycles and use for data translations should not occur
539 		 * for 3 cpu cycles.
540 		 */
541 		uasm_i_ssnop(p);
542 		uasm_i_ssnop(p);
543 		uasm_i_ssnop(p);
544 		uasm_i_ssnop(p);
545 		tlbw(p);
546 		uasm_i_ssnop(p);
547 		uasm_i_ssnop(p);
548 		uasm_i_ssnop(p);
549 		uasm_i_ssnop(p);
550 		break;
551 
552 	case CPU_VR4111:
553 	case CPU_VR4121:
554 	case CPU_VR4122:
555 	case CPU_VR4181:
556 	case CPU_VR4181A:
557 		uasm_i_nop(p);
558 		uasm_i_nop(p);
559 		tlbw(p);
560 		uasm_i_nop(p);
561 		uasm_i_nop(p);
562 		break;
563 
564 	case CPU_VR4131:
565 	case CPU_VR4133:
566 	case CPU_R5432:
567 		uasm_i_nop(p);
568 		uasm_i_nop(p);
569 		tlbw(p);
570 		break;
571 
572 	case CPU_JZRISC:
573 		tlbw(p);
574 		uasm_i_nop(p);
575 		break;
576 
577 	default:
578 		panic("No TLB refill handler yet (CPU type: %d)",
579 		      current_cpu_data.cputype);
580 		break;
581 	}
582 }
583 
584 static __cpuinit __maybe_unused void build_convert_pte_to_entrylo(u32 **p,
585 								  unsigned int reg)
586 {
587 	if (kernel_uses_smartmips_rixi) {
588 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_NO_EXEC));
589 		UASM_i_ROTR(p, reg, reg, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
590 	} else {
591 #ifdef CONFIG_64BIT_PHYS_ADDR
592 		uasm_i_dsrl_safe(p, reg, reg, ilog2(_PAGE_GLOBAL));
593 #else
594 		UASM_i_SRL(p, reg, reg, ilog2(_PAGE_GLOBAL));
595 #endif
596 	}
597 }
598 
599 #ifdef CONFIG_HUGETLB_PAGE
600 
601 static __cpuinit void build_restore_pagemask(u32 **p,
602 					     struct uasm_reloc **r,
603 					     unsigned int tmp,
604 					     enum label_id lid,
605 					     int restore_scratch)
606 {
607 	if (restore_scratch) {
608 		/* Reset default page size */
609 		if (PM_DEFAULT_MASK >> 16) {
610 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
611 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
612 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
613 			uasm_il_b(p, r, lid);
614 		} else if (PM_DEFAULT_MASK) {
615 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
616 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
617 			uasm_il_b(p, r, lid);
618 		} else {
619 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
620 			uasm_il_b(p, r, lid);
621 		}
622 		if (scratch_reg > 0)
623 			UASM_i_MFC0(p, 1, 31, scratch_reg);
624 		else
625 			UASM_i_LW(p, 1, scratchpad_offset(0), 0);
626 	} else {
627 		/* Reset default page size */
628 		if (PM_DEFAULT_MASK >> 16) {
629 			uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
630 			uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
631 			uasm_il_b(p, r, lid);
632 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
633 		} else if (PM_DEFAULT_MASK) {
634 			uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
635 			uasm_il_b(p, r, lid);
636 			uasm_i_mtc0(p, tmp, C0_PAGEMASK);
637 		} else {
638 			uasm_il_b(p, r, lid);
639 			uasm_i_mtc0(p, 0, C0_PAGEMASK);
640 		}
641 	}
642 }
643 
644 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
645 						 struct uasm_label **l,
646 						 struct uasm_reloc **r,
647 						 unsigned int tmp,
648 						 enum tlb_write_entry wmode,
649 						 int restore_scratch)
650 {
651 	/* Set huge page tlb entry size */
652 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
653 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
654 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
655 
656 	build_tlb_write_entry(p, l, r, wmode);
657 
658 	build_restore_pagemask(p, r, tmp, label_leave, restore_scratch);
659 }
660 
661 /*
662  * Check if Huge PTE is present, if so then jump to LABEL.
663  */
664 static void __cpuinit
665 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
666 		unsigned int pmd, int lid)
667 {
668 	UASM_i_LW(p, tmp, 0, pmd);
669 	if (use_bbit_insns()) {
670 		uasm_il_bbit1(p, r, tmp, ilog2(_PAGE_HUGE), lid);
671 	} else {
672 		uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
673 		uasm_il_bnez(p, r, tmp, lid);
674 	}
675 }
676 
677 static __cpuinit void build_huge_update_entries(u32 **p,
678 						unsigned int pte,
679 						unsigned int tmp)
680 {
681 	int small_sequence;
682 
683 	/*
684 	 * A huge PTE describes an area the size of the
685 	 * configured huge page size. This is twice the
686 	 * of the large TLB entry size we intend to use.
687 	 * A TLB entry half the size of the configured
688 	 * huge page size is configured into entrylo0
689 	 * and entrylo1 to cover the contiguous huge PTE
690 	 * address space.
691 	 */
692 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
693 
694 	/* We can clobber tmp.  It isn't used after this.*/
695 	if (!small_sequence)
696 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
697 
698 	build_convert_pte_to_entrylo(p, pte);
699 	UASM_i_MTC0(p, pte, C0_ENTRYLO0); /* load it */
700 	/* convert to entrylo1 */
701 	if (small_sequence)
702 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
703 	else
704 		UASM_i_ADDU(p, pte, pte, tmp);
705 
706 	UASM_i_MTC0(p, pte, C0_ENTRYLO1); /* load it */
707 }
708 
709 static __cpuinit void build_huge_handler_tail(u32 **p,
710 					      struct uasm_reloc **r,
711 					      struct uasm_label **l,
712 					      unsigned int pte,
713 					      unsigned int ptr)
714 {
715 #ifdef CONFIG_SMP
716 	UASM_i_SC(p, pte, 0, ptr);
717 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
718 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
719 #else
720 	UASM_i_SW(p, pte, 0, ptr);
721 #endif
722 	build_huge_update_entries(p, pte, ptr);
723 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed, 0);
724 }
725 #endif /* CONFIG_HUGETLB_PAGE */
726 
727 #ifdef CONFIG_64BIT
728 /*
729  * TMP and PTR are scratch.
730  * TMP will be clobbered, PTR will hold the pmd entry.
731  */
732 static void __cpuinit
733 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
734 		 unsigned int tmp, unsigned int ptr)
735 {
736 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
737 	long pgdc = (long)pgd_current;
738 #endif
739 	/*
740 	 * The vmalloc handling is not in the hotpath.
741 	 */
742 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
743 
744 	if (check_for_high_segbits) {
745 		/*
746 		 * The kernel currently implicitely assumes that the
747 		 * MIPS SEGBITS parameter for the processor is
748 		 * (PGDIR_SHIFT+PGDIR_BITS) or less, and will never
749 		 * allocate virtual addresses outside the maximum
750 		 * range for SEGBITS = (PGDIR_SHIFT+PGDIR_BITS). But
751 		 * that doesn't prevent user code from accessing the
752 		 * higher xuseg addresses.  Here, we make sure that
753 		 * everything but the lower xuseg addresses goes down
754 		 * the module_alloc/vmalloc path.
755 		 */
756 		uasm_i_dsrl_safe(p, ptr, tmp, PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
757 		uasm_il_bnez(p, r, ptr, label_vmalloc);
758 	} else {
759 		uasm_il_bltz(p, r, tmp, label_vmalloc);
760 	}
761 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
762 
763 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
764 	if (pgd_reg != -1) {
765 		/* pgd is in pgd_reg */
766 		UASM_i_MFC0(p, ptr, 31, pgd_reg);
767 	} else {
768 		/*
769 		 * &pgd << 11 stored in CONTEXT [23..63].
770 		 */
771 		UASM_i_MFC0(p, ptr, C0_CONTEXT);
772 
773 		/* Clear lower 23 bits of context. */
774 		uasm_i_dins(p, ptr, 0, 0, 23);
775 
776 		/* 1 0  1 0 1  << 6  xkphys cached */
777 		uasm_i_ori(p, ptr, ptr, 0x540);
778 		uasm_i_drotr(p, ptr, ptr, 11);
779 	}
780 #elif defined(CONFIG_SMP)
781 # ifdef  CONFIG_MIPS_MT_SMTC
782 	/*
783 	 * SMTC uses TCBind value as "CPU" index
784 	 */
785 	uasm_i_mfc0(p, ptr, C0_TCBIND);
786 	uasm_i_dsrl_safe(p, ptr, ptr, 19);
787 # else
788 	/*
789 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
790 	 * stored in CONTEXT.
791 	 */
792 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
793 	uasm_i_dsrl_safe(p, ptr, ptr, 23);
794 # endif
795 	UASM_i_LA_mostly(p, tmp, pgdc);
796 	uasm_i_daddu(p, ptr, ptr, tmp);
797 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
798 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
799 #else
800 	UASM_i_LA_mostly(p, ptr, pgdc);
801 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
802 #endif
803 
804 	uasm_l_vmalloc_done(l, *p);
805 
806 	/* get pgd offset in bytes */
807 	uasm_i_dsrl_safe(p, tmp, tmp, PGDIR_SHIFT - 3);
808 
809 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
810 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
811 #ifndef __PAGETABLE_PMD_FOLDED
812 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
813 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
814 	uasm_i_dsrl_safe(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
815 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
816 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
817 #endif
818 }
819 
820 /*
821  * BVADDR is the faulting address, PTR is scratch.
822  * PTR will hold the pgd for vmalloc.
823  */
824 static void __cpuinit
825 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
826 			unsigned int bvaddr, unsigned int ptr,
827 			enum vmalloc64_mode mode)
828 {
829 	long swpd = (long)swapper_pg_dir;
830 	int single_insn_swpd;
831 	int did_vmalloc_branch = 0;
832 
833 	single_insn_swpd = uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd);
834 
835 	uasm_l_vmalloc(l, *p);
836 
837 	if (mode != not_refill && check_for_high_segbits) {
838 		if (single_insn_swpd) {
839 			uasm_il_bltz(p, r, bvaddr, label_vmalloc_done);
840 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
841 			did_vmalloc_branch = 1;
842 			/* fall through */
843 		} else {
844 			uasm_il_bgez(p, r, bvaddr, label_large_segbits_fault);
845 		}
846 	}
847 	if (!did_vmalloc_branch) {
848 		if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
849 			uasm_il_b(p, r, label_vmalloc_done);
850 			uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
851 		} else {
852 			UASM_i_LA_mostly(p, ptr, swpd);
853 			uasm_il_b(p, r, label_vmalloc_done);
854 			if (uasm_in_compat_space_p(swpd))
855 				uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
856 			else
857 				uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
858 		}
859 	}
860 	if (mode != not_refill && check_for_high_segbits) {
861 		uasm_l_large_segbits_fault(l, *p);
862 		/*
863 		 * We get here if we are an xsseg address, or if we are
864 		 * an xuseg address above (PGDIR_SHIFT+PGDIR_BITS) boundary.
865 		 *
866 		 * Ignoring xsseg (assume disabled so would generate
867 		 * (address errors?), the only remaining possibility
868 		 * is the upper xuseg addresses.  On processors with
869 		 * TLB_SEGBITS <= PGDIR_SHIFT+PGDIR_BITS, these
870 		 * addresses would have taken an address error. We try
871 		 * to mimic that here by taking a load/istream page
872 		 * fault.
873 		 */
874 		UASM_i_LA(p, ptr, (unsigned long)tlb_do_page_fault_0);
875 		uasm_i_jr(p, ptr);
876 
877 		if (mode == refill_scratch) {
878 			if (scratch_reg > 0)
879 				UASM_i_MFC0(p, 1, 31, scratch_reg);
880 			else
881 				UASM_i_LW(p, 1, scratchpad_offset(0), 0);
882 		} else {
883 			uasm_i_nop(p);
884 		}
885 	}
886 }
887 
888 #else /* !CONFIG_64BIT */
889 
890 /*
891  * TMP and PTR are scratch.
892  * TMP will be clobbered, PTR will hold the pgd entry.
893  */
894 static void __cpuinit __maybe_unused
895 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
896 {
897 	long pgdc = (long)pgd_current;
898 
899 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
900 #ifdef CONFIG_SMP
901 #ifdef  CONFIG_MIPS_MT_SMTC
902 	/*
903 	 * SMTC uses TCBind value as "CPU" index
904 	 */
905 	uasm_i_mfc0(p, ptr, C0_TCBIND);
906 	UASM_i_LA_mostly(p, tmp, pgdc);
907 	uasm_i_srl(p, ptr, ptr, 19);
908 #else
909 	/*
910 	 * smp_processor_id() << 3 is stored in CONTEXT.
911          */
912 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
913 	UASM_i_LA_mostly(p, tmp, pgdc);
914 	uasm_i_srl(p, ptr, ptr, 23);
915 #endif
916 	uasm_i_addu(p, ptr, tmp, ptr);
917 #else
918 	UASM_i_LA_mostly(p, ptr, pgdc);
919 #endif
920 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
921 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
922 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
923 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
924 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
925 }
926 
927 #endif /* !CONFIG_64BIT */
928 
929 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
930 {
931 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
932 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
933 
934 	switch (current_cpu_type()) {
935 	case CPU_VR41XX:
936 	case CPU_VR4111:
937 	case CPU_VR4121:
938 	case CPU_VR4122:
939 	case CPU_VR4131:
940 	case CPU_VR4181:
941 	case CPU_VR4181A:
942 	case CPU_VR4133:
943 		shift += 2;
944 		break;
945 
946 	default:
947 		break;
948 	}
949 
950 	if (shift)
951 		UASM_i_SRL(p, ctx, ctx, shift);
952 	uasm_i_andi(p, ctx, ctx, mask);
953 }
954 
955 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
956 {
957 	/*
958 	 * Bug workaround for the Nevada. It seems as if under certain
959 	 * circumstances the move from cp0_context might produce a
960 	 * bogus result when the mfc0 instruction and its consumer are
961 	 * in a different cacheline or a load instruction, probably any
962 	 * memory reference, is between them.
963 	 */
964 	switch (current_cpu_type()) {
965 	case CPU_NEVADA:
966 		UASM_i_LW(p, ptr, 0, ptr);
967 		GET_CONTEXT(p, tmp); /* get context reg */
968 		break;
969 
970 	default:
971 		GET_CONTEXT(p, tmp); /* get context reg */
972 		UASM_i_LW(p, ptr, 0, ptr);
973 		break;
974 	}
975 
976 	build_adjust_context(p, tmp);
977 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
978 }
979 
980 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
981 					unsigned int ptep)
982 {
983 	/*
984 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
985 	 * Kernel is a special case. Only a few CPUs use it.
986 	 */
987 #ifdef CONFIG_64BIT_PHYS_ADDR
988 	if (cpu_has_64bits) {
989 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
990 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
991 		if (kernel_uses_smartmips_rixi) {
992 			UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
993 			UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
994 			UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
995 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
996 			UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
997 		} else {
998 			uasm_i_dsrl_safe(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
999 			UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1000 			uasm_i_dsrl_safe(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1001 		}
1002 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1003 	} else {
1004 		int pte_off_even = sizeof(pte_t) / 2;
1005 		int pte_off_odd = pte_off_even + sizeof(pte_t);
1006 
1007 		/* The pte entries are pre-shifted */
1008 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
1009 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1010 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
1011 		UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1012 	}
1013 #else
1014 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
1015 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
1016 	if (r45k_bvahwbug())
1017 		build_tlb_probe_entry(p);
1018 	if (kernel_uses_smartmips_rixi) {
1019 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
1020 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
1021 		UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1022 		if (r4k_250MHZhwbug())
1023 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1024 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1025 		UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1026 	} else {
1027 		UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /* convert to entrylo0 */
1028 		if (r4k_250MHZhwbug())
1029 			UASM_i_MTC0(p, 0, C0_ENTRYLO0);
1030 		UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
1031 		UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /* convert to entrylo1 */
1032 		if (r45k_bvahwbug())
1033 			uasm_i_mfc0(p, tmp, C0_INDEX);
1034 	}
1035 	if (r4k_250MHZhwbug())
1036 		UASM_i_MTC0(p, 0, C0_ENTRYLO1);
1037 	UASM_i_MTC0(p, ptep, C0_ENTRYLO1); /* load it */
1038 #endif
1039 }
1040 
1041 struct mips_huge_tlb_info {
1042 	int huge_pte;
1043 	int restore_scratch;
1044 };
1045 
1046 static struct mips_huge_tlb_info __cpuinit
1047 build_fast_tlb_refill_handler (u32 **p, struct uasm_label **l,
1048 			       struct uasm_reloc **r, unsigned int tmp,
1049 			       unsigned int ptr, int c0_scratch)
1050 {
1051 	struct mips_huge_tlb_info rv;
1052 	unsigned int even, odd;
1053 	int vmalloc_branch_delay_filled = 0;
1054 	const int scratch = 1; /* Our extra working register */
1055 
1056 	rv.huge_pte = scratch;
1057 	rv.restore_scratch = 0;
1058 
1059 	if (check_for_high_segbits) {
1060 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1061 
1062 		if (pgd_reg != -1)
1063 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1064 		else
1065 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1066 
1067 		if (c0_scratch >= 0)
1068 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1069 		else
1070 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1071 
1072 		uasm_i_dsrl_safe(p, scratch, tmp,
1073 				 PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
1074 		uasm_il_bnez(p, r, scratch, label_vmalloc);
1075 
1076 		if (pgd_reg == -1) {
1077 			vmalloc_branch_delay_filled = 1;
1078 			/* Clear lower 23 bits of context. */
1079 			uasm_i_dins(p, ptr, 0, 0, 23);
1080 		}
1081 	} else {
1082 		if (pgd_reg != -1)
1083 			UASM_i_MFC0(p, ptr, 31, pgd_reg);
1084 		else
1085 			UASM_i_MFC0(p, ptr, C0_CONTEXT);
1086 
1087 		UASM_i_MFC0(p, tmp, C0_BADVADDR);
1088 
1089 		if (c0_scratch >= 0)
1090 			UASM_i_MTC0(p, scratch, 31, c0_scratch);
1091 		else
1092 			UASM_i_SW(p, scratch, scratchpad_offset(0), 0);
1093 
1094 		if (pgd_reg == -1)
1095 			/* Clear lower 23 bits of context. */
1096 			uasm_i_dins(p, ptr, 0, 0, 23);
1097 
1098 		uasm_il_bltz(p, r, tmp, label_vmalloc);
1099 	}
1100 
1101 	if (pgd_reg == -1) {
1102 		vmalloc_branch_delay_filled = 1;
1103 		/* 1 0  1 0 1  << 6  xkphys cached */
1104 		uasm_i_ori(p, ptr, ptr, 0x540);
1105 		uasm_i_drotr(p, ptr, ptr, 11);
1106 	}
1107 
1108 #ifdef __PAGETABLE_PMD_FOLDED
1109 #define LOC_PTEP scratch
1110 #else
1111 #define LOC_PTEP ptr
1112 #endif
1113 
1114 	if (!vmalloc_branch_delay_filled)
1115 		/* get pgd offset in bytes */
1116 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1117 
1118 	uasm_l_vmalloc_done(l, *p);
1119 
1120 	/*
1121 	 *                         tmp          ptr
1122 	 * fall-through case =   badvaddr  *pgd_current
1123 	 * vmalloc case      =   badvaddr  swapper_pg_dir
1124 	 */
1125 
1126 	if (vmalloc_branch_delay_filled)
1127 		/* get pgd offset in bytes */
1128 		uasm_i_dsrl_safe(p, scratch, tmp, PGDIR_SHIFT - 3);
1129 
1130 #ifdef __PAGETABLE_PMD_FOLDED
1131 	GET_CONTEXT(p, tmp); /* get context reg */
1132 #endif
1133 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PGD - 1) << 3);
1134 
1135 	if (use_lwx_insns()) {
1136 		UASM_i_LWX(p, LOC_PTEP, scratch, ptr);
1137 	} else {
1138 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pgd offset */
1139 		uasm_i_ld(p, LOC_PTEP, 0, ptr); /* get pmd pointer */
1140 	}
1141 
1142 #ifndef __PAGETABLE_PMD_FOLDED
1143 	/* get pmd offset in bytes */
1144 	uasm_i_dsrl_safe(p, scratch, tmp, PMD_SHIFT - 3);
1145 	uasm_i_andi(p, scratch, scratch, (PTRS_PER_PMD - 1) << 3);
1146 	GET_CONTEXT(p, tmp); /* get context reg */
1147 
1148 	if (use_lwx_insns()) {
1149 		UASM_i_LWX(p, scratch, scratch, ptr);
1150 	} else {
1151 		uasm_i_daddu(p, ptr, ptr, scratch); /* add in pmd offset */
1152 		UASM_i_LW(p, scratch, 0, ptr);
1153 	}
1154 #endif
1155 	/* Adjust the context during the load latency. */
1156 	build_adjust_context(p, tmp);
1157 
1158 #ifdef CONFIG_HUGETLB_PAGE
1159 	uasm_il_bbit1(p, r, scratch, ilog2(_PAGE_HUGE), label_tlb_huge_update);
1160 	/*
1161 	 * The in the LWX case we don't want to do the load in the
1162 	 * delay slot.  It cannot issue in the same cycle and may be
1163 	 * speculative and unneeded.
1164 	 */
1165 	if (use_lwx_insns())
1166 		uasm_i_nop(p);
1167 #endif /* CONFIG_HUGETLB_PAGE */
1168 
1169 
1170 	/* build_update_entries */
1171 	if (use_lwx_insns()) {
1172 		even = ptr;
1173 		odd = tmp;
1174 		UASM_i_LWX(p, even, scratch, tmp);
1175 		UASM_i_ADDIU(p, tmp, tmp, sizeof(pte_t));
1176 		UASM_i_LWX(p, odd, scratch, tmp);
1177 	} else {
1178 		UASM_i_ADDU(p, ptr, scratch, tmp); /* add in offset */
1179 		even = tmp;
1180 		odd = ptr;
1181 		UASM_i_LW(p, even, 0, ptr); /* get even pte */
1182 		UASM_i_LW(p, odd, sizeof(pte_t), ptr); /* get odd pte */
1183 	}
1184 	if (kernel_uses_smartmips_rixi) {
1185 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_NO_EXEC));
1186 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_NO_EXEC));
1187 		uasm_i_drotr(p, even, even,
1188 			     ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1189 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1190 		uasm_i_drotr(p, odd, odd,
1191 			     ilog2(_PAGE_GLOBAL) - ilog2(_PAGE_NO_EXEC));
1192 	} else {
1193 		uasm_i_dsrl_safe(p, even, even, ilog2(_PAGE_GLOBAL));
1194 		UASM_i_MTC0(p, even, C0_ENTRYLO0); /* load it */
1195 		uasm_i_dsrl_safe(p, odd, odd, ilog2(_PAGE_GLOBAL));
1196 	}
1197 	UASM_i_MTC0(p, odd, C0_ENTRYLO1); /* load it */
1198 
1199 	if (c0_scratch >= 0) {
1200 		UASM_i_MFC0(p, scratch, 31, c0_scratch);
1201 		build_tlb_write_entry(p, l, r, tlb_random);
1202 		uasm_l_leave(l, *p);
1203 		rv.restore_scratch = 1;
1204 	} else if (PAGE_SHIFT == 14 || PAGE_SHIFT == 13)  {
1205 		build_tlb_write_entry(p, l, r, tlb_random);
1206 		uasm_l_leave(l, *p);
1207 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1208 	} else {
1209 		UASM_i_LW(p, scratch, scratchpad_offset(0), 0);
1210 		build_tlb_write_entry(p, l, r, tlb_random);
1211 		uasm_l_leave(l, *p);
1212 		rv.restore_scratch = 1;
1213 	}
1214 
1215 	uasm_i_eret(p); /* return from trap */
1216 
1217 	return rv;
1218 }
1219 
1220 /*
1221  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
1222  * because EXL == 0.  If we wrap, we can also use the 32 instruction
1223  * slots before the XTLB refill exception handler which belong to the
1224  * unused TLB refill exception.
1225  */
1226 #define MIPS64_REFILL_INSNS 32
1227 
1228 static void __cpuinit build_r4000_tlb_refill_handler(void)
1229 {
1230 	u32 *p = tlb_handler;
1231 	struct uasm_label *l = labels;
1232 	struct uasm_reloc *r = relocs;
1233 	u32 *f;
1234 	unsigned int final_len;
1235 	struct mips_huge_tlb_info htlb_info __maybe_unused;
1236 	enum vmalloc64_mode vmalloc_mode __maybe_unused;
1237 
1238 	memset(tlb_handler, 0, sizeof(tlb_handler));
1239 	memset(labels, 0, sizeof(labels));
1240 	memset(relocs, 0, sizeof(relocs));
1241 	memset(final_handler, 0, sizeof(final_handler));
1242 
1243 	if ((scratch_reg > 0 || scratchpad_available()) && use_bbit_insns()) {
1244 		htlb_info = build_fast_tlb_refill_handler(&p, &l, &r, K0, K1,
1245 							  scratch_reg);
1246 		vmalloc_mode = refill_scratch;
1247 	} else {
1248 		htlb_info.huge_pte = K0;
1249 		htlb_info.restore_scratch = 0;
1250 		vmalloc_mode = refill_noscratch;
1251 		/*
1252 		 * create the plain linear handler
1253 		 */
1254 		if (bcm1250_m3_war()) {
1255 			unsigned int segbits = 44;
1256 
1257 			uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1258 			uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1259 			uasm_i_xor(&p, K0, K0, K1);
1260 			uasm_i_dsrl_safe(&p, K1, K0, 62);
1261 			uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1262 			uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1263 			uasm_i_or(&p, K0, K0, K1);
1264 			uasm_il_bnez(&p, &r, K0, label_leave);
1265 			/* No need for uasm_i_nop */
1266 		}
1267 
1268 #ifdef CONFIG_64BIT
1269 		build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
1270 #else
1271 		build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
1272 #endif
1273 
1274 #ifdef CONFIG_HUGETLB_PAGE
1275 		build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
1276 #endif
1277 
1278 		build_get_ptep(&p, K0, K1);
1279 		build_update_entries(&p, K0, K1);
1280 		build_tlb_write_entry(&p, &l, &r, tlb_random);
1281 		uasm_l_leave(&l, p);
1282 		uasm_i_eret(&p); /* return from trap */
1283 	}
1284 #ifdef CONFIG_HUGETLB_PAGE
1285 	uasm_l_tlb_huge_update(&l, p);
1286 	build_huge_update_entries(&p, htlb_info.huge_pte, K1);
1287 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random,
1288 				   htlb_info.restore_scratch);
1289 #endif
1290 
1291 #ifdef CONFIG_64BIT
1292 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1, vmalloc_mode);
1293 #endif
1294 
1295 	/*
1296 	 * Overflow check: For the 64bit handler, we need at least one
1297 	 * free instruction slot for the wrap-around branch. In worst
1298 	 * case, if the intended insertion point is a delay slot, we
1299 	 * need three, with the second nop'ed and the third being
1300 	 * unused.
1301 	 */
1302 	/* Loongson2 ebase is different than r4k, we have more space */
1303 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1304 	if ((p - tlb_handler) > 64)
1305 		panic("TLB refill handler space exceeded");
1306 #else
1307 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
1308 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
1309 		&& uasm_insn_has_bdelay(relocs,
1310 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
1311 		panic("TLB refill handler space exceeded");
1312 #endif
1313 
1314 	/*
1315 	 * Now fold the handler in the TLB refill handler space.
1316 	 */
1317 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
1318 	f = final_handler;
1319 	/* Simplest case, just copy the handler. */
1320 	uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1321 	final_len = p - tlb_handler;
1322 #else /* CONFIG_64BIT */
1323 	f = final_handler + MIPS64_REFILL_INSNS;
1324 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
1325 		/* Just copy the handler. */
1326 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
1327 		final_len = p - tlb_handler;
1328 	} else {
1329 #if defined(CONFIG_HUGETLB_PAGE)
1330 		const enum label_id ls = label_tlb_huge_update;
1331 #else
1332 		const enum label_id ls = label_vmalloc;
1333 #endif
1334 		u32 *split;
1335 		int ov = 0;
1336 		int i;
1337 
1338 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
1339 			;
1340 		BUG_ON(i == ARRAY_SIZE(labels));
1341 		split = labels[i].addr;
1342 
1343 		/*
1344 		 * See if we have overflown one way or the other.
1345 		 */
1346 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
1347 		    split < p - MIPS64_REFILL_INSNS)
1348 			ov = 1;
1349 
1350 		if (ov) {
1351 			/*
1352 			 * Split two instructions before the end.  One
1353 			 * for the branch and one for the instruction
1354 			 * in the delay slot.
1355 			 */
1356 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
1357 
1358 			/*
1359 			 * If the branch would fall in a delay slot,
1360 			 * we must back up an additional instruction
1361 			 * so that it is no longer in a delay slot.
1362 			 */
1363 			if (uasm_insn_has_bdelay(relocs, split - 1))
1364 				split--;
1365 		}
1366 		/* Copy first part of the handler. */
1367 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
1368 		f += split - tlb_handler;
1369 
1370 		if (ov) {
1371 			/* Insert branch. */
1372 			uasm_l_split(&l, final_handler);
1373 			uasm_il_b(&f, &r, label_split);
1374 			if (uasm_insn_has_bdelay(relocs, split))
1375 				uasm_i_nop(&f);
1376 			else {
1377 				uasm_copy_handler(relocs, labels,
1378 						  split, split + 1, f);
1379 				uasm_move_labels(labels, f, f + 1, -1);
1380 				f++;
1381 				split++;
1382 			}
1383 		}
1384 
1385 		/* Copy the rest of the handler. */
1386 		uasm_copy_handler(relocs, labels, split, p, final_handler);
1387 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
1388 			    (p - split);
1389 	}
1390 #endif /* CONFIG_64BIT */
1391 
1392 	uasm_resolve_relocs(relocs, labels);
1393 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
1394 		 final_len);
1395 
1396 	memcpy((void *)ebase, final_handler, 0x100);
1397 
1398 	dump_handler((u32 *)ebase, 64);
1399 }
1400 
1401 /*
1402  * 128 instructions for the fastpath handler is generous and should
1403  * never be exceeded.
1404  */
1405 #define FASTPATH_SIZE 128
1406 
1407 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
1408 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
1409 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
1410 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
1411 u32 tlbmiss_handler_setup_pgd[16] __cacheline_aligned;
1412 
1413 static void __cpuinit build_r4000_setup_pgd(void)
1414 {
1415 	const int a0 = 4;
1416 	const int a1 = 5;
1417 	u32 *p = tlbmiss_handler_setup_pgd;
1418 	struct uasm_label *l = labels;
1419 	struct uasm_reloc *r = relocs;
1420 
1421 	memset(tlbmiss_handler_setup_pgd, 0, sizeof(tlbmiss_handler_setup_pgd));
1422 	memset(labels, 0, sizeof(labels));
1423 	memset(relocs, 0, sizeof(relocs));
1424 
1425 	pgd_reg = allocate_kscratch();
1426 
1427 	if (pgd_reg == -1) {
1428 		/* PGD << 11 in c0_Context */
1429 		/*
1430 		 * If it is a ckseg0 address, convert to a physical
1431 		 * address.  Shifting right by 29 and adding 4 will
1432 		 * result in zero for these addresses.
1433 		 *
1434 		 */
1435 		UASM_i_SRA(&p, a1, a0, 29);
1436 		UASM_i_ADDIU(&p, a1, a1, 4);
1437 		uasm_il_bnez(&p, &r, a1, label_tlbl_goaround1);
1438 		uasm_i_nop(&p);
1439 		uasm_i_dinsm(&p, a0, 0, 29, 64 - 29);
1440 		uasm_l_tlbl_goaround1(&l, p);
1441 		UASM_i_SLL(&p, a0, a0, 11);
1442 		uasm_i_jr(&p, 31);
1443 		UASM_i_MTC0(&p, a0, C0_CONTEXT);
1444 	} else {
1445 		/* PGD in c0_KScratch */
1446 		uasm_i_jr(&p, 31);
1447 		UASM_i_MTC0(&p, a0, 31, pgd_reg);
1448 	}
1449 	if (p - tlbmiss_handler_setup_pgd > ARRAY_SIZE(tlbmiss_handler_setup_pgd))
1450 		panic("tlbmiss_handler_setup_pgd space exceeded");
1451 	uasm_resolve_relocs(relocs, labels);
1452 	pr_debug("Wrote tlbmiss_handler_setup_pgd (%u instructions).\n",
1453 		 (unsigned int)(p - tlbmiss_handler_setup_pgd));
1454 
1455 	dump_handler(tlbmiss_handler_setup_pgd,
1456 		     ARRAY_SIZE(tlbmiss_handler_setup_pgd));
1457 }
1458 #endif
1459 
1460 static void __cpuinit
1461 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
1462 {
1463 #ifdef CONFIG_SMP
1464 # ifdef CONFIG_64BIT_PHYS_ADDR
1465 	if (cpu_has_64bits)
1466 		uasm_i_lld(p, pte, 0, ptr);
1467 	else
1468 # endif
1469 		UASM_i_LL(p, pte, 0, ptr);
1470 #else
1471 # ifdef CONFIG_64BIT_PHYS_ADDR
1472 	if (cpu_has_64bits)
1473 		uasm_i_ld(p, pte, 0, ptr);
1474 	else
1475 # endif
1476 		UASM_i_LW(p, pte, 0, ptr);
1477 #endif
1478 }
1479 
1480 static void __cpuinit
1481 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
1482 	unsigned int mode)
1483 {
1484 #ifdef CONFIG_64BIT_PHYS_ADDR
1485 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
1486 #endif
1487 
1488 	uasm_i_ori(p, pte, pte, mode);
1489 #ifdef CONFIG_SMP
1490 # ifdef CONFIG_64BIT_PHYS_ADDR
1491 	if (cpu_has_64bits)
1492 		uasm_i_scd(p, pte, 0, ptr);
1493 	else
1494 # endif
1495 		UASM_i_SC(p, pte, 0, ptr);
1496 
1497 	if (r10000_llsc_war())
1498 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
1499 	else
1500 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1501 
1502 # ifdef CONFIG_64BIT_PHYS_ADDR
1503 	if (!cpu_has_64bits) {
1504 		/* no uasm_i_nop needed */
1505 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
1506 		uasm_i_ori(p, pte, pte, hwmode);
1507 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
1508 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
1509 		/* no uasm_i_nop needed */
1510 		uasm_i_lw(p, pte, 0, ptr);
1511 	} else
1512 		uasm_i_nop(p);
1513 # else
1514 	uasm_i_nop(p);
1515 # endif
1516 #else
1517 # ifdef CONFIG_64BIT_PHYS_ADDR
1518 	if (cpu_has_64bits)
1519 		uasm_i_sd(p, pte, 0, ptr);
1520 	else
1521 # endif
1522 		UASM_i_SW(p, pte, 0, ptr);
1523 
1524 # ifdef CONFIG_64BIT_PHYS_ADDR
1525 	if (!cpu_has_64bits) {
1526 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
1527 		uasm_i_ori(p, pte, pte, hwmode);
1528 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
1529 		uasm_i_lw(p, pte, 0, ptr);
1530 	}
1531 # endif
1532 #endif
1533 }
1534 
1535 /*
1536  * Check if PTE is present, if not then jump to LABEL. PTR points to
1537  * the page table where this PTE is located, PTE will be re-loaded
1538  * with it's original value.
1539  */
1540 static void __cpuinit
1541 build_pte_present(u32 **p, struct uasm_reloc **r,
1542 		  int pte, int ptr, int scratch, enum label_id lid)
1543 {
1544 	int t = scratch >= 0 ? scratch : pte;
1545 
1546 	if (kernel_uses_smartmips_rixi) {
1547 		if (use_bbit_insns()) {
1548 			uasm_il_bbit0(p, r, pte, ilog2(_PAGE_PRESENT), lid);
1549 			uasm_i_nop(p);
1550 		} else {
1551 			uasm_i_andi(p, t, pte, _PAGE_PRESENT);
1552 			uasm_il_beqz(p, r, t, lid);
1553 			if (pte == t)
1554 				/* You lose the SMP race :-(*/
1555 				iPTE_LW(p, pte, ptr);
1556 		}
1557 	} else {
1558 		uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_READ);
1559 		uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_READ);
1560 		uasm_il_bnez(p, r, t, lid);
1561 		if (pte == t)
1562 			/* You lose the SMP race :-(*/
1563 			iPTE_LW(p, pte, ptr);
1564 	}
1565 }
1566 
1567 /* Make PTE valid, store result in PTR. */
1568 static void __cpuinit
1569 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
1570 		 unsigned int ptr)
1571 {
1572 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
1573 
1574 	iPTE_SW(p, r, pte, ptr, mode);
1575 }
1576 
1577 /*
1578  * Check if PTE can be written to, if not branch to LABEL. Regardless
1579  * restore PTE with value from PTR when done.
1580  */
1581 static void __cpuinit
1582 build_pte_writable(u32 **p, struct uasm_reloc **r,
1583 		   unsigned int pte, unsigned int ptr, int scratch,
1584 		   enum label_id lid)
1585 {
1586 	int t = scratch >= 0 ? scratch : pte;
1587 
1588 	uasm_i_andi(p, t, pte, _PAGE_PRESENT | _PAGE_WRITE);
1589 	uasm_i_xori(p, t, t, _PAGE_PRESENT | _PAGE_WRITE);
1590 	uasm_il_bnez(p, r, t, lid);
1591 	if (pte == t)
1592 		/* You lose the SMP race :-(*/
1593 		iPTE_LW(p, pte, ptr);
1594 	else
1595 		uasm_i_nop(p);
1596 }
1597 
1598 /* Make PTE writable, update software status bits as well, then store
1599  * at PTR.
1600  */
1601 static void __cpuinit
1602 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1603 		 unsigned int ptr)
1604 {
1605 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1606 			     | _PAGE_DIRTY);
1607 
1608 	iPTE_SW(p, r, pte, ptr, mode);
1609 }
1610 
1611 /*
1612  * Check if PTE can be modified, if not branch to LABEL. Regardless
1613  * restore PTE with value from PTR when done.
1614  */
1615 static void __cpuinit
1616 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1617 		     unsigned int pte, unsigned int ptr, int scratch,
1618 		     enum label_id lid)
1619 {
1620 	if (use_bbit_insns()) {
1621 		uasm_il_bbit0(p, r, pte, ilog2(_PAGE_WRITE), lid);
1622 		uasm_i_nop(p);
1623 	} else {
1624 		int t = scratch >= 0 ? scratch : pte;
1625 		uasm_i_andi(p, t, pte, _PAGE_WRITE);
1626 		uasm_il_beqz(p, r, t, lid);
1627 		if (pte == t)
1628 			/* You lose the SMP race :-(*/
1629 			iPTE_LW(p, pte, ptr);
1630 	}
1631 }
1632 
1633 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
1634 
1635 
1636 /*
1637  * R3000 style TLB load/store/modify handlers.
1638  */
1639 
1640 /*
1641  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1642  * Then it returns.
1643  */
1644 static void __cpuinit
1645 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1646 {
1647 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1648 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1649 	uasm_i_tlbwi(p);
1650 	uasm_i_jr(p, tmp);
1651 	uasm_i_rfe(p); /* branch delay */
1652 }
1653 
1654 /*
1655  * This places the pte into ENTRYLO0 and writes it with tlbwi
1656  * or tlbwr as appropriate.  This is because the index register
1657  * may have the probe fail bit set as a result of a trap on a
1658  * kseg2 access, i.e. without refill.  Then it returns.
1659  */
1660 static void __cpuinit
1661 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1662 			     struct uasm_reloc **r, unsigned int pte,
1663 			     unsigned int tmp)
1664 {
1665 	uasm_i_mfc0(p, tmp, C0_INDEX);
1666 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1667 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1668 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1669 	uasm_i_tlbwi(p); /* cp0 delay */
1670 	uasm_i_jr(p, tmp);
1671 	uasm_i_rfe(p); /* branch delay */
1672 	uasm_l_r3000_write_probe_fail(l, *p);
1673 	uasm_i_tlbwr(p); /* cp0 delay */
1674 	uasm_i_jr(p, tmp);
1675 	uasm_i_rfe(p); /* branch delay */
1676 }
1677 
1678 static void __cpuinit
1679 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1680 				   unsigned int ptr)
1681 {
1682 	long pgdc = (long)pgd_current;
1683 
1684 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1685 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1686 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1687 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1688 	uasm_i_sll(p, pte, pte, 2);
1689 	uasm_i_addu(p, ptr, ptr, pte);
1690 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1691 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1692 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1693 	uasm_i_addu(p, ptr, ptr, pte);
1694 	uasm_i_lw(p, pte, 0, ptr);
1695 	uasm_i_tlbp(p); /* load delay */
1696 }
1697 
1698 static void __cpuinit build_r3000_tlb_load_handler(void)
1699 {
1700 	u32 *p = handle_tlbl;
1701 	struct uasm_label *l = labels;
1702 	struct uasm_reloc *r = relocs;
1703 
1704 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1705 	memset(labels, 0, sizeof(labels));
1706 	memset(relocs, 0, sizeof(relocs));
1707 
1708 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1709 	build_pte_present(&p, &r, K0, K1, -1, label_nopage_tlbl);
1710 	uasm_i_nop(&p); /* load delay */
1711 	build_make_valid(&p, &r, K0, K1);
1712 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1713 
1714 	uasm_l_nopage_tlbl(&l, p);
1715 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1716 	uasm_i_nop(&p);
1717 
1718 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1719 		panic("TLB load handler fastpath space exceeded");
1720 
1721 	uasm_resolve_relocs(relocs, labels);
1722 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1723 		 (unsigned int)(p - handle_tlbl));
1724 
1725 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1726 }
1727 
1728 static void __cpuinit build_r3000_tlb_store_handler(void)
1729 {
1730 	u32 *p = handle_tlbs;
1731 	struct uasm_label *l = labels;
1732 	struct uasm_reloc *r = relocs;
1733 
1734 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1735 	memset(labels, 0, sizeof(labels));
1736 	memset(relocs, 0, sizeof(relocs));
1737 
1738 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1739 	build_pte_writable(&p, &r, K0, K1, -1, label_nopage_tlbs);
1740 	uasm_i_nop(&p); /* load delay */
1741 	build_make_write(&p, &r, K0, K1);
1742 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1743 
1744 	uasm_l_nopage_tlbs(&l, p);
1745 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1746 	uasm_i_nop(&p);
1747 
1748 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1749 		panic("TLB store handler fastpath space exceeded");
1750 
1751 	uasm_resolve_relocs(relocs, labels);
1752 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1753 		 (unsigned int)(p - handle_tlbs));
1754 
1755 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1756 }
1757 
1758 static void __cpuinit build_r3000_tlb_modify_handler(void)
1759 {
1760 	u32 *p = handle_tlbm;
1761 	struct uasm_label *l = labels;
1762 	struct uasm_reloc *r = relocs;
1763 
1764 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1765 	memset(labels, 0, sizeof(labels));
1766 	memset(relocs, 0, sizeof(relocs));
1767 
1768 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1769 	build_pte_modifiable(&p, &r, K0, K1,  -1, label_nopage_tlbm);
1770 	uasm_i_nop(&p); /* load delay */
1771 	build_make_write(&p, &r, K0, K1);
1772 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1773 
1774 	uasm_l_nopage_tlbm(&l, p);
1775 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1776 	uasm_i_nop(&p);
1777 
1778 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1779 		panic("TLB modify handler fastpath space exceeded");
1780 
1781 	uasm_resolve_relocs(relocs, labels);
1782 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1783 		 (unsigned int)(p - handle_tlbm));
1784 
1785 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1786 }
1787 #endif /* CONFIG_MIPS_PGD_C0_CONTEXT */
1788 
1789 /*
1790  * R4000 style TLB load/store/modify handlers.
1791  */
1792 static struct work_registers __cpuinit
1793 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1794 				   struct uasm_reloc **r)
1795 {
1796 	struct work_registers wr = build_get_work_registers(p);
1797 
1798 #ifdef CONFIG_64BIT
1799 	build_get_pmde64(p, l, r, wr.r1, wr.r2); /* get pmd in ptr */
1800 #else
1801 	build_get_pgde32(p, wr.r1, wr.r2); /* get pgd in ptr */
1802 #endif
1803 
1804 #ifdef CONFIG_HUGETLB_PAGE
1805 	/*
1806 	 * For huge tlb entries, pmd doesn't contain an address but
1807 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1808 	 * see if we need to jump to huge tlb processing.
1809 	 */
1810 	build_is_huge_pte(p, r, wr.r1, wr.r2, label_tlb_huge_update);
1811 #endif
1812 
1813 	UASM_i_MFC0(p, wr.r1, C0_BADVADDR);
1814 	UASM_i_LW(p, wr.r2, 0, wr.r2);
1815 	UASM_i_SRL(p, wr.r1, wr.r1, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1816 	uasm_i_andi(p, wr.r1, wr.r1, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1817 	UASM_i_ADDU(p, wr.r2, wr.r2, wr.r1);
1818 
1819 #ifdef CONFIG_SMP
1820 	uasm_l_smp_pgtable_change(l, *p);
1821 #endif
1822 	iPTE_LW(p, wr.r1, wr.r2); /* get even pte */
1823 	if (!m4kc_tlbp_war())
1824 		build_tlb_probe_entry(p);
1825 	return wr;
1826 }
1827 
1828 static void __cpuinit
1829 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1830 				   struct uasm_reloc **r, unsigned int tmp,
1831 				   unsigned int ptr)
1832 {
1833 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1834 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1835 	build_update_entries(p, tmp, ptr);
1836 	build_tlb_write_entry(p, l, r, tlb_indexed);
1837 	uasm_l_leave(l, *p);
1838 	build_restore_work_registers(p);
1839 	uasm_i_eret(p); /* return from trap */
1840 
1841 #ifdef CONFIG_64BIT
1842 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr, not_refill);
1843 #endif
1844 }
1845 
1846 static void __cpuinit build_r4000_tlb_load_handler(void)
1847 {
1848 	u32 *p = handle_tlbl;
1849 	struct uasm_label *l = labels;
1850 	struct uasm_reloc *r = relocs;
1851 	struct work_registers wr;
1852 
1853 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1854 	memset(labels, 0, sizeof(labels));
1855 	memset(relocs, 0, sizeof(relocs));
1856 
1857 	if (bcm1250_m3_war()) {
1858 		unsigned int segbits = 44;
1859 
1860 		uasm_i_dmfc0(&p, K0, C0_BADVADDR);
1861 		uasm_i_dmfc0(&p, K1, C0_ENTRYHI);
1862 		uasm_i_xor(&p, K0, K0, K1);
1863 		uasm_i_dsrl_safe(&p, K1, K0, 62);
1864 		uasm_i_dsrl_safe(&p, K0, K0, 12 + 1);
1865 		uasm_i_dsll_safe(&p, K0, K0, 64 + 12 + 1 - segbits);
1866 		uasm_i_or(&p, K0, K0, K1);
1867 		uasm_il_bnez(&p, &r, K0, label_leave);
1868 		/* No need for uasm_i_nop */
1869 	}
1870 
1871 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
1872 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1873 	if (m4kc_tlbp_war())
1874 		build_tlb_probe_entry(&p);
1875 
1876 	if (kernel_uses_smartmips_rixi) {
1877 		/*
1878 		 * If the page is not _PAGE_VALID, RI or XI could not
1879 		 * have triggered it.  Skip the expensive test..
1880 		 */
1881 		if (use_bbit_insns()) {
1882 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1883 				      label_tlbl_goaround1);
1884 		} else {
1885 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1886 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround1);
1887 		}
1888 		uasm_i_nop(&p);
1889 
1890 		uasm_i_tlbr(&p);
1891 		/* Examine  entrylo 0 or 1 based on ptr. */
1892 		if (use_bbit_insns()) {
1893 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1894 		} else {
1895 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1896 			uasm_i_beqz(&p, wr.r3, 8);
1897 		}
1898 		/* load it in the delay slot*/
1899 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1900 		/* load it if ptr is odd */
1901 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1902 		/*
1903 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1904 		 * XI must have triggered it.
1905 		 */
1906 		if (use_bbit_insns()) {
1907 			uasm_il_bbit1(&p, &r, wr.r3, 1, label_nopage_tlbl);
1908 			uasm_i_nop(&p);
1909 			uasm_l_tlbl_goaround1(&l, p);
1910 		} else {
1911 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1912 			uasm_il_bnez(&p, &r, wr.r3, label_nopage_tlbl);
1913 			uasm_i_nop(&p);
1914 		}
1915 		uasm_l_tlbl_goaround1(&l, p);
1916 	}
1917 	build_make_valid(&p, &r, wr.r1, wr.r2);
1918 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
1919 
1920 #ifdef CONFIG_HUGETLB_PAGE
1921 	/*
1922 	 * This is the entry point when build_r4000_tlbchange_handler_head
1923 	 * spots a huge page.
1924 	 */
1925 	uasm_l_tlb_huge_update(&l, p);
1926 	iPTE_LW(&p, wr.r1, wr.r2);
1927 	build_pte_present(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbl);
1928 	build_tlb_probe_entry(&p);
1929 
1930 	if (kernel_uses_smartmips_rixi) {
1931 		/*
1932 		 * If the page is not _PAGE_VALID, RI or XI could not
1933 		 * have triggered it.  Skip the expensive test..
1934 		 */
1935 		if (use_bbit_insns()) {
1936 			uasm_il_bbit0(&p, &r, wr.r1, ilog2(_PAGE_VALID),
1937 				      label_tlbl_goaround2);
1938 		} else {
1939 			uasm_i_andi(&p, wr.r3, wr.r1, _PAGE_VALID);
1940 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1941 		}
1942 		uasm_i_nop(&p);
1943 
1944 		uasm_i_tlbr(&p);
1945 		/* Examine  entrylo 0 or 1 based on ptr. */
1946 		if (use_bbit_insns()) {
1947 			uasm_i_bbit0(&p, wr.r2, ilog2(sizeof(pte_t)), 8);
1948 		} else {
1949 			uasm_i_andi(&p, wr.r3, wr.r2, sizeof(pte_t));
1950 			uasm_i_beqz(&p, wr.r3, 8);
1951 		}
1952 		/* load it in the delay slot*/
1953 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO0);
1954 		/* load it if ptr is odd */
1955 		UASM_i_MFC0(&p, wr.r3, C0_ENTRYLO1);
1956 		/*
1957 		 * If the entryLo (now in wr.r3) is valid (bit 1), RI or
1958 		 * XI must have triggered it.
1959 		 */
1960 		if (use_bbit_insns()) {
1961 			uasm_il_bbit0(&p, &r, wr.r3, 1, label_tlbl_goaround2);
1962 		} else {
1963 			uasm_i_andi(&p, wr.r3, wr.r3, 2);
1964 			uasm_il_beqz(&p, &r, wr.r3, label_tlbl_goaround2);
1965 		}
1966 		if (PM_DEFAULT_MASK == 0)
1967 			uasm_i_nop(&p);
1968 		/*
1969 		 * We clobbered C0_PAGEMASK, restore it.  On the other branch
1970 		 * it is restored in build_huge_tlb_write_entry.
1971 		 */
1972 		build_restore_pagemask(&p, &r, wr.r3, label_nopage_tlbl, 0);
1973 
1974 		uasm_l_tlbl_goaround2(&l, p);
1975 	}
1976 	uasm_i_ori(&p, wr.r1, wr.r1, (_PAGE_ACCESSED | _PAGE_VALID));
1977 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
1978 #endif
1979 
1980 	uasm_l_nopage_tlbl(&l, p);
1981 	build_restore_work_registers(&p);
1982 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1983 	uasm_i_nop(&p);
1984 
1985 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1986 		panic("TLB load handler fastpath space exceeded");
1987 
1988 	uasm_resolve_relocs(relocs, labels);
1989 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1990 		 (unsigned int)(p - handle_tlbl));
1991 
1992 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1993 }
1994 
1995 static void __cpuinit build_r4000_tlb_store_handler(void)
1996 {
1997 	u32 *p = handle_tlbs;
1998 	struct uasm_label *l = labels;
1999 	struct uasm_reloc *r = relocs;
2000 	struct work_registers wr;
2001 
2002 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
2003 	memset(labels, 0, sizeof(labels));
2004 	memset(relocs, 0, sizeof(relocs));
2005 
2006 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2007 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2008 	if (m4kc_tlbp_war())
2009 		build_tlb_probe_entry(&p);
2010 	build_make_write(&p, &r, wr.r1, wr.r2);
2011 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2012 
2013 #ifdef CONFIG_HUGETLB_PAGE
2014 	/*
2015 	 * This is the entry point when
2016 	 * build_r4000_tlbchange_handler_head spots a huge page.
2017 	 */
2018 	uasm_l_tlb_huge_update(&l, p);
2019 	iPTE_LW(&p, wr.r1, wr.r2);
2020 	build_pte_writable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbs);
2021 	build_tlb_probe_entry(&p);
2022 	uasm_i_ori(&p, wr.r1, wr.r1,
2023 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2024 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2025 #endif
2026 
2027 	uasm_l_nopage_tlbs(&l, p);
2028 	build_restore_work_registers(&p);
2029 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2030 	uasm_i_nop(&p);
2031 
2032 	if ((p - handle_tlbs) > FASTPATH_SIZE)
2033 		panic("TLB store handler fastpath space exceeded");
2034 
2035 	uasm_resolve_relocs(relocs, labels);
2036 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
2037 		 (unsigned int)(p - handle_tlbs));
2038 
2039 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
2040 }
2041 
2042 static void __cpuinit build_r4000_tlb_modify_handler(void)
2043 {
2044 	u32 *p = handle_tlbm;
2045 	struct uasm_label *l = labels;
2046 	struct uasm_reloc *r = relocs;
2047 	struct work_registers wr;
2048 
2049 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
2050 	memset(labels, 0, sizeof(labels));
2051 	memset(relocs, 0, sizeof(relocs));
2052 
2053 	wr = build_r4000_tlbchange_handler_head(&p, &l, &r);
2054 	build_pte_modifiable(&p, &r, wr.r1, wr.r2, wr.r3, label_nopage_tlbm);
2055 	if (m4kc_tlbp_war())
2056 		build_tlb_probe_entry(&p);
2057 	/* Present and writable bits set, set accessed and dirty bits. */
2058 	build_make_write(&p, &r, wr.r1, wr.r2);
2059 	build_r4000_tlbchange_handler_tail(&p, &l, &r, wr.r1, wr.r2);
2060 
2061 #ifdef CONFIG_HUGETLB_PAGE
2062 	/*
2063 	 * This is the entry point when
2064 	 * build_r4000_tlbchange_handler_head spots a huge page.
2065 	 */
2066 	uasm_l_tlb_huge_update(&l, p);
2067 	iPTE_LW(&p, wr.r1, wr.r2);
2068 	build_pte_modifiable(&p, &r, wr.r1, wr.r2,  wr.r3, label_nopage_tlbm);
2069 	build_tlb_probe_entry(&p);
2070 	uasm_i_ori(&p, wr.r1, wr.r1,
2071 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
2072 	build_huge_handler_tail(&p, &r, &l, wr.r1, wr.r2);
2073 #endif
2074 
2075 	uasm_l_nopage_tlbm(&l, p);
2076 	build_restore_work_registers(&p);
2077 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
2078 	uasm_i_nop(&p);
2079 
2080 	if ((p - handle_tlbm) > FASTPATH_SIZE)
2081 		panic("TLB modify handler fastpath space exceeded");
2082 
2083 	uasm_resolve_relocs(relocs, labels);
2084 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
2085 		 (unsigned int)(p - handle_tlbm));
2086 
2087 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
2088 }
2089 
2090 void __cpuinit build_tlb_refill_handler(void)
2091 {
2092 	/*
2093 	 * The refill handler is generated per-CPU, multi-node systems
2094 	 * may have local storage for it. The other handlers are only
2095 	 * needed once.
2096 	 */
2097 	static int run_once = 0;
2098 
2099 #ifdef CONFIG_64BIT
2100 	check_for_high_segbits = current_cpu_data.vmbits > (PGDIR_SHIFT + PGD_ORDER + PAGE_SHIFT - 3);
2101 #endif
2102 
2103 	switch (current_cpu_type()) {
2104 	case CPU_R2000:
2105 	case CPU_R3000:
2106 	case CPU_R3000A:
2107 	case CPU_R3081E:
2108 	case CPU_TX3912:
2109 	case CPU_TX3922:
2110 	case CPU_TX3927:
2111 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
2112 		build_r3000_tlb_refill_handler();
2113 		if (!run_once) {
2114 			build_r3000_tlb_load_handler();
2115 			build_r3000_tlb_store_handler();
2116 			build_r3000_tlb_modify_handler();
2117 			run_once++;
2118 		}
2119 #else
2120 		panic("No R3000 TLB refill handler");
2121 #endif
2122 		break;
2123 
2124 	case CPU_R6000:
2125 	case CPU_R6000A:
2126 		panic("No R6000 TLB refill handler yet");
2127 		break;
2128 
2129 	case CPU_R8000:
2130 		panic("No R8000 TLB refill handler yet");
2131 		break;
2132 
2133 	default:
2134 		if (!run_once) {
2135 			scratch_reg = allocate_kscratch();
2136 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2137 			build_r4000_setup_pgd();
2138 #endif
2139 			build_r4000_tlb_load_handler();
2140 			build_r4000_tlb_store_handler();
2141 			build_r4000_tlb_modify_handler();
2142 			run_once++;
2143 		}
2144 		build_r4000_tlb_refill_handler();
2145 	}
2146 }
2147 
2148 void __cpuinit flush_tlb_handlers(void)
2149 {
2150 	local_flush_icache_range((unsigned long)handle_tlbl,
2151 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
2152 	local_flush_icache_range((unsigned long)handle_tlbs,
2153 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
2154 	local_flush_icache_range((unsigned long)handle_tlbm,
2155 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
2156 #ifdef CONFIG_MIPS_PGD_C0_CONTEXT
2157 	local_flush_icache_range((unsigned long)tlbmiss_handler_setup_pgd,
2158 			   (unsigned long)tlbmiss_handler_setup_pgd + sizeof(handle_tlbm));
2159 #endif
2160 }
2161