xref: /openbmc/linux/arch/mips/mm/tlbex.c (revision fd589a8f)
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 
30 #include <asm/mmu_context.h>
31 #include <asm/war.h>
32 
33 #include "uasm.h"
34 
35 static inline int r45k_bvahwbug(void)
36 {
37 	/* XXX: We should probe for the presence of this bug, but we don't. */
38 	return 0;
39 }
40 
41 static inline int r4k_250MHZhwbug(void)
42 {
43 	/* XXX: We should probe for the presence of this bug, but we don't. */
44 	return 0;
45 }
46 
47 static inline int __maybe_unused bcm1250_m3_war(void)
48 {
49 	return BCM1250_M3_WAR;
50 }
51 
52 static inline int __maybe_unused r10000_llsc_war(void)
53 {
54 	return R10000_LLSC_WAR;
55 }
56 
57 /*
58  * Found by experiment: At least some revisions of the 4kc throw under
59  * some circumstances a machine check exception, triggered by invalid
60  * values in the index register.  Delaying the tlbp instruction until
61  * after the next branch,  plus adding an additional nop in front of
62  * tlbwi/tlbwr avoids the invalid index register values. Nobody knows
63  * why; it's not an issue caused by the core RTL.
64  *
65  */
66 static int __cpuinit m4kc_tlbp_war(void)
67 {
68 	return (current_cpu_data.processor_id & 0xffff00) ==
69 	       (PRID_COMP_MIPS | PRID_IMP_4KC);
70 }
71 
72 /* Handle labels (which must be positive integers). */
73 enum label_id {
74 	label_second_part = 1,
75 	label_leave,
76 #ifdef MODULE_START
77 	label_module_alloc,
78 #endif
79 	label_vmalloc,
80 	label_vmalloc_done,
81 	label_tlbw_hazard,
82 	label_split,
83 	label_nopage_tlbl,
84 	label_nopage_tlbs,
85 	label_nopage_tlbm,
86 	label_smp_pgtable_change,
87 	label_r3000_write_probe_fail,
88 #ifdef CONFIG_HUGETLB_PAGE
89 	label_tlb_huge_update,
90 #endif
91 };
92 
93 UASM_L_LA(_second_part)
94 UASM_L_LA(_leave)
95 #ifdef MODULE_START
96 UASM_L_LA(_module_alloc)
97 #endif
98 UASM_L_LA(_vmalloc)
99 UASM_L_LA(_vmalloc_done)
100 UASM_L_LA(_tlbw_hazard)
101 UASM_L_LA(_split)
102 UASM_L_LA(_nopage_tlbl)
103 UASM_L_LA(_nopage_tlbs)
104 UASM_L_LA(_nopage_tlbm)
105 UASM_L_LA(_smp_pgtable_change)
106 UASM_L_LA(_r3000_write_probe_fail)
107 #ifdef CONFIG_HUGETLB_PAGE
108 UASM_L_LA(_tlb_huge_update)
109 #endif
110 
111 /*
112  * For debug purposes.
113  */
114 static inline void dump_handler(const u32 *handler, int count)
115 {
116 	int i;
117 
118 	pr_debug("\t.set push\n");
119 	pr_debug("\t.set noreorder\n");
120 
121 	for (i = 0; i < count; i++)
122 		pr_debug("\t%p\t.word 0x%08x\n", &handler[i], handler[i]);
123 
124 	pr_debug("\t.set pop\n");
125 }
126 
127 /* The only general purpose registers allowed in TLB handlers. */
128 #define K0		26
129 #define K1		27
130 
131 /* Some CP0 registers */
132 #define C0_INDEX	0, 0
133 #define C0_ENTRYLO0	2, 0
134 #define C0_TCBIND	2, 2
135 #define C0_ENTRYLO1	3, 0
136 #define C0_CONTEXT	4, 0
137 #define C0_PAGEMASK	5, 0
138 #define C0_BADVADDR	8, 0
139 #define C0_ENTRYHI	10, 0
140 #define C0_EPC		14, 0
141 #define C0_XCONTEXT	20, 0
142 
143 #ifdef CONFIG_64BIT
144 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_XCONTEXT)
145 #else
146 # define GET_CONTEXT(buf, reg) UASM_i_MFC0(buf, reg, C0_CONTEXT)
147 #endif
148 
149 /* The worst case length of the handler is around 18 instructions for
150  * R3000-style TLBs and up to 63 instructions for R4000-style TLBs.
151  * Maximum space available is 32 instructions for R3000 and 64
152  * instructions for R4000.
153  *
154  * We deliberately chose a buffer size of 128, so we won't scribble
155  * over anything important on overflow before we panic.
156  */
157 static u32 tlb_handler[128] __cpuinitdata;
158 
159 /* simply assume worst case size for labels and relocs */
160 static struct uasm_label labels[128] __cpuinitdata;
161 static struct uasm_reloc relocs[128] __cpuinitdata;
162 
163 /*
164  * The R3000 TLB handler is simple.
165  */
166 static void __cpuinit build_r3000_tlb_refill_handler(void)
167 {
168 	long pgdc = (long)pgd_current;
169 	u32 *p;
170 
171 	memset(tlb_handler, 0, sizeof(tlb_handler));
172 	p = tlb_handler;
173 
174 	uasm_i_mfc0(&p, K0, C0_BADVADDR);
175 	uasm_i_lui(&p, K1, uasm_rel_hi(pgdc)); /* cp0 delay */
176 	uasm_i_lw(&p, K1, uasm_rel_lo(pgdc), K1);
177 	uasm_i_srl(&p, K0, K0, 22); /* load delay */
178 	uasm_i_sll(&p, K0, K0, 2);
179 	uasm_i_addu(&p, K1, K1, K0);
180 	uasm_i_mfc0(&p, K0, C0_CONTEXT);
181 	uasm_i_lw(&p, K1, 0, K1); /* cp0 delay */
182 	uasm_i_andi(&p, K0, K0, 0xffc); /* load delay */
183 	uasm_i_addu(&p, K1, K1, K0);
184 	uasm_i_lw(&p, K0, 0, K1);
185 	uasm_i_nop(&p); /* load delay */
186 	uasm_i_mtc0(&p, K0, C0_ENTRYLO0);
187 	uasm_i_mfc0(&p, K1, C0_EPC); /* cp0 delay */
188 	uasm_i_tlbwr(&p); /* cp0 delay */
189 	uasm_i_jr(&p, K1);
190 	uasm_i_rfe(&p); /* branch delay */
191 
192 	if (p > tlb_handler + 32)
193 		panic("TLB refill handler space exceeded");
194 
195 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
196 		 (unsigned int)(p - tlb_handler));
197 
198 	memcpy((void *)ebase, tlb_handler, 0x80);
199 
200 	dump_handler((u32 *)ebase, 32);
201 }
202 
203 /*
204  * The R4000 TLB handler is much more complicated. We have two
205  * consecutive handler areas with 32 instructions space each.
206  * Since they aren't used at the same time, we can overflow in the
207  * other one.To keep things simple, we first assume linear space,
208  * then we relocate it to the final handler layout as needed.
209  */
210 static u32 final_handler[64] __cpuinitdata;
211 
212 /*
213  * Hazards
214  *
215  * From the IDT errata for the QED RM5230 (Nevada), processor revision 1.0:
216  * 2. A timing hazard exists for the TLBP instruction.
217  *
218  *      stalling_instruction
219  *      TLBP
220  *
221  * The JTLB is being read for the TLBP throughout the stall generated by the
222  * previous instruction. This is not really correct as the stalling instruction
223  * can modify the address used to access the JTLB.  The failure symptom is that
224  * the TLBP instruction will use an address created for the stalling instruction
225  * and not the address held in C0_ENHI and thus report the wrong results.
226  *
227  * The software work-around is to not allow the instruction preceding the TLBP
228  * to stall - make it an NOP or some other instruction guaranteed not to stall.
229  *
230  * Errata 2 will not be fixed.  This errata is also on the R5000.
231  *
232  * As if we MIPS hackers wouldn't know how to nop pipelines happy ...
233  */
234 static void __cpuinit __maybe_unused build_tlb_probe_entry(u32 **p)
235 {
236 	switch (current_cpu_type()) {
237 	/* Found by experiment: R4600 v2.0/R4700 needs this, too.  */
238 	case CPU_R4600:
239 	case CPU_R4700:
240 	case CPU_R5000:
241 	case CPU_R5000A:
242 	case CPU_NEVADA:
243 		uasm_i_nop(p);
244 		uasm_i_tlbp(p);
245 		break;
246 
247 	default:
248 		uasm_i_tlbp(p);
249 		break;
250 	}
251 }
252 
253 /*
254  * Write random or indexed TLB entry, and care about the hazards from
255  * the preceeding mtc0 and for the following eret.
256  */
257 enum tlb_write_entry { tlb_random, tlb_indexed };
258 
259 static void __cpuinit build_tlb_write_entry(u32 **p, struct uasm_label **l,
260 					 struct uasm_reloc **r,
261 					 enum tlb_write_entry wmode)
262 {
263 	void(*tlbw)(u32 **) = NULL;
264 
265 	switch (wmode) {
266 	case tlb_random: tlbw = uasm_i_tlbwr; break;
267 	case tlb_indexed: tlbw = uasm_i_tlbwi; break;
268 	}
269 
270 	if (cpu_has_mips_r2) {
271 		if (cpu_has_mips_r2_exec_hazard)
272 			uasm_i_ehb(p);
273 		tlbw(p);
274 		return;
275 	}
276 
277 	switch (current_cpu_type()) {
278 	case CPU_R4000PC:
279 	case CPU_R4000SC:
280 	case CPU_R4000MC:
281 	case CPU_R4400PC:
282 	case CPU_R4400SC:
283 	case CPU_R4400MC:
284 		/*
285 		 * This branch uses up a mtc0 hazard nop slot and saves
286 		 * two nops after the tlbw instruction.
287 		 */
288 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
289 		tlbw(p);
290 		uasm_l_tlbw_hazard(l, *p);
291 		uasm_i_nop(p);
292 		break;
293 
294 	case CPU_R4600:
295 	case CPU_R4700:
296 	case CPU_R5000:
297 	case CPU_R5000A:
298 		uasm_i_nop(p);
299 		tlbw(p);
300 		uasm_i_nop(p);
301 		break;
302 
303 	case CPU_R4300:
304 	case CPU_5KC:
305 	case CPU_TX49XX:
306 	case CPU_PR4450:
307 		uasm_i_nop(p);
308 		tlbw(p);
309 		break;
310 
311 	case CPU_R10000:
312 	case CPU_R12000:
313 	case CPU_R14000:
314 	case CPU_4KC:
315 	case CPU_4KEC:
316 	case CPU_SB1:
317 	case CPU_SB1A:
318 	case CPU_4KSC:
319 	case CPU_20KC:
320 	case CPU_25KF:
321 	case CPU_BCM3302:
322 	case CPU_BCM4710:
323 	case CPU_LOONGSON2:
324 	case CPU_BCM6338:
325 	case CPU_BCM6345:
326 	case CPU_BCM6348:
327 	case CPU_BCM6358:
328 	case CPU_R5500:
329 		if (m4kc_tlbp_war())
330 			uasm_i_nop(p);
331 	case CPU_ALCHEMY:
332 		tlbw(p);
333 		break;
334 
335 	case CPU_NEVADA:
336 		uasm_i_nop(p); /* QED specifies 2 nops hazard */
337 		/*
338 		 * This branch uses up a mtc0 hazard nop slot and saves
339 		 * a nop after the tlbw instruction.
340 		 */
341 		uasm_il_bgezl(p, r, 0, label_tlbw_hazard);
342 		tlbw(p);
343 		uasm_l_tlbw_hazard(l, *p);
344 		break;
345 
346 	case CPU_RM7000:
347 		uasm_i_nop(p);
348 		uasm_i_nop(p);
349 		uasm_i_nop(p);
350 		uasm_i_nop(p);
351 		tlbw(p);
352 		break;
353 
354 	case CPU_RM9000:
355 		/*
356 		 * When the JTLB is updated by tlbwi or tlbwr, a subsequent
357 		 * use of the JTLB for instructions should not occur for 4
358 		 * cpu cycles and use for data translations should not occur
359 		 * for 3 cpu cycles.
360 		 */
361 		uasm_i_ssnop(p);
362 		uasm_i_ssnop(p);
363 		uasm_i_ssnop(p);
364 		uasm_i_ssnop(p);
365 		tlbw(p);
366 		uasm_i_ssnop(p);
367 		uasm_i_ssnop(p);
368 		uasm_i_ssnop(p);
369 		uasm_i_ssnop(p);
370 		break;
371 
372 	case CPU_VR4111:
373 	case CPU_VR4121:
374 	case CPU_VR4122:
375 	case CPU_VR4181:
376 	case CPU_VR4181A:
377 		uasm_i_nop(p);
378 		uasm_i_nop(p);
379 		tlbw(p);
380 		uasm_i_nop(p);
381 		uasm_i_nop(p);
382 		break;
383 
384 	case CPU_VR4131:
385 	case CPU_VR4133:
386 	case CPU_R5432:
387 		uasm_i_nop(p);
388 		uasm_i_nop(p);
389 		tlbw(p);
390 		break;
391 
392 	default:
393 		panic("No TLB refill handler yet (CPU type: %d)",
394 		      current_cpu_data.cputype);
395 		break;
396 	}
397 }
398 
399 #ifdef CONFIG_HUGETLB_PAGE
400 static __cpuinit void build_huge_tlb_write_entry(u32 **p,
401 						 struct uasm_label **l,
402 						 struct uasm_reloc **r,
403 						 unsigned int tmp,
404 						 enum tlb_write_entry wmode)
405 {
406 	/* Set huge page tlb entry size */
407 	uasm_i_lui(p, tmp, PM_HUGE_MASK >> 16);
408 	uasm_i_ori(p, tmp, tmp, PM_HUGE_MASK & 0xffff);
409 	uasm_i_mtc0(p, tmp, C0_PAGEMASK);
410 
411 	build_tlb_write_entry(p, l, r, wmode);
412 
413 	/* Reset default page size */
414 	if (PM_DEFAULT_MASK >> 16) {
415 		uasm_i_lui(p, tmp, PM_DEFAULT_MASK >> 16);
416 		uasm_i_ori(p, tmp, tmp, PM_DEFAULT_MASK & 0xffff);
417 		uasm_il_b(p, r, label_leave);
418 		uasm_i_mtc0(p, tmp, C0_PAGEMASK);
419 	} else if (PM_DEFAULT_MASK) {
420 		uasm_i_ori(p, tmp, 0, PM_DEFAULT_MASK);
421 		uasm_il_b(p, r, label_leave);
422 		uasm_i_mtc0(p, tmp, C0_PAGEMASK);
423 	} else {
424 		uasm_il_b(p, r, label_leave);
425 		uasm_i_mtc0(p, 0, C0_PAGEMASK);
426 	}
427 }
428 
429 /*
430  * Check if Huge PTE is present, if so then jump to LABEL.
431  */
432 static void __cpuinit
433 build_is_huge_pte(u32 **p, struct uasm_reloc **r, unsigned int tmp,
434 		unsigned int pmd, int lid)
435 {
436 	UASM_i_LW(p, tmp, 0, pmd);
437 	uasm_i_andi(p, tmp, tmp, _PAGE_HUGE);
438 	uasm_il_bnez(p, r, tmp, lid);
439 }
440 
441 static __cpuinit void build_huge_update_entries(u32 **p,
442 						unsigned int pte,
443 						unsigned int tmp)
444 {
445 	int small_sequence;
446 
447 	/*
448 	 * A huge PTE describes an area the size of the
449 	 * configured huge page size. This is twice the
450 	 * of the large TLB entry size we intend to use.
451 	 * A TLB entry half the size of the configured
452 	 * huge page size is configured into entrylo0
453 	 * and entrylo1 to cover the contiguous huge PTE
454 	 * address space.
455 	 */
456 	small_sequence = (HPAGE_SIZE >> 7) < 0x10000;
457 
458 	/* We can clobber tmp.  It isn't used after this.*/
459 	if (!small_sequence)
460 		uasm_i_lui(p, tmp, HPAGE_SIZE >> (7 + 16));
461 
462 	UASM_i_SRL(p, pte, pte, 6); /* convert to entrylo */
463 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* load it */
464 	/* convert to entrylo1 */
465 	if (small_sequence)
466 		UASM_i_ADDIU(p, pte, pte, HPAGE_SIZE >> 7);
467 	else
468 		UASM_i_ADDU(p, pte, pte, tmp);
469 
470 	uasm_i_mtc0(p, pte, C0_ENTRYLO1); /* load it */
471 }
472 
473 static __cpuinit void build_huge_handler_tail(u32 **p,
474 					      struct uasm_reloc **r,
475 					      struct uasm_label **l,
476 					      unsigned int pte,
477 					      unsigned int ptr)
478 {
479 #ifdef CONFIG_SMP
480 	UASM_i_SC(p, pte, 0, ptr);
481 	uasm_il_beqz(p, r, pte, label_tlb_huge_update);
482 	UASM_i_LW(p, pte, 0, ptr); /* Needed because SC killed our PTE */
483 #else
484 	UASM_i_SW(p, pte, 0, ptr);
485 #endif
486 	build_huge_update_entries(p, pte, ptr);
487 	build_huge_tlb_write_entry(p, l, r, pte, tlb_indexed);
488 }
489 #endif /* CONFIG_HUGETLB_PAGE */
490 
491 #ifdef CONFIG_64BIT
492 /*
493  * TMP and PTR are scratch.
494  * TMP will be clobbered, PTR will hold the pmd entry.
495  */
496 static void __cpuinit
497 build_get_pmde64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
498 		 unsigned int tmp, unsigned int ptr)
499 {
500 	long pgdc = (long)pgd_current;
501 
502 	/*
503 	 * The vmalloc handling is not in the hotpath.
504 	 */
505 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
506 	uasm_il_bltz(p, r, tmp, label_vmalloc);
507 	/* No uasm_i_nop needed here, since the next insn doesn't touch TMP. */
508 
509 #ifdef CONFIG_SMP
510 # ifdef  CONFIG_MIPS_MT_SMTC
511 	/*
512 	 * SMTC uses TCBind value as "CPU" index
513 	 */
514 	uasm_i_mfc0(p, ptr, C0_TCBIND);
515 	uasm_i_dsrl(p, ptr, ptr, 19);
516 # else
517 	/*
518 	 * 64 bit SMP running in XKPHYS has smp_processor_id() << 3
519 	 * stored in CONTEXT.
520 	 */
521 	uasm_i_dmfc0(p, ptr, C0_CONTEXT);
522 	uasm_i_dsrl(p, ptr, ptr, 23);
523 #endif
524 	UASM_i_LA_mostly(p, tmp, pgdc);
525 	uasm_i_daddu(p, ptr, ptr, tmp);
526 	uasm_i_dmfc0(p, tmp, C0_BADVADDR);
527 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
528 #else
529 	UASM_i_LA_mostly(p, ptr, pgdc);
530 	uasm_i_ld(p, ptr, uasm_rel_lo(pgdc), ptr);
531 #endif
532 
533 	uasm_l_vmalloc_done(l, *p);
534 
535 	if (PGDIR_SHIFT - 3 < 32)		/* get pgd offset in bytes */
536 		uasm_i_dsrl(p, tmp, tmp, PGDIR_SHIFT-3);
537 	else
538 		uasm_i_dsrl32(p, tmp, tmp, PGDIR_SHIFT - 3 - 32);
539 
540 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PGD - 1)<<3);
541 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pgd offset */
542 	uasm_i_dmfc0(p, tmp, C0_BADVADDR); /* get faulting address */
543 	uasm_i_ld(p, ptr, 0, ptr); /* get pmd pointer */
544 	uasm_i_dsrl(p, tmp, tmp, PMD_SHIFT-3); /* get pmd offset in bytes */
545 	uasm_i_andi(p, tmp, tmp, (PTRS_PER_PMD - 1)<<3);
546 	uasm_i_daddu(p, ptr, ptr, tmp); /* add in pmd offset */
547 }
548 
549 /*
550  * BVADDR is the faulting address, PTR is scratch.
551  * PTR will hold the pgd for vmalloc.
552  */
553 static void __cpuinit
554 build_get_pgd_vmalloc64(u32 **p, struct uasm_label **l, struct uasm_reloc **r,
555 			unsigned int bvaddr, unsigned int ptr)
556 {
557 	long swpd = (long)swapper_pg_dir;
558 
559 	uasm_l_vmalloc(l, *p);
560 
561 	if (uasm_in_compat_space_p(swpd) && !uasm_rel_lo(swpd)) {
562 		uasm_il_b(p, r, label_vmalloc_done);
563 		uasm_i_lui(p, ptr, uasm_rel_hi(swpd));
564 	} else {
565 		UASM_i_LA_mostly(p, ptr, swpd);
566 		uasm_il_b(p, r, label_vmalloc_done);
567 		if (uasm_in_compat_space_p(swpd))
568 			uasm_i_addiu(p, ptr, ptr, uasm_rel_lo(swpd));
569 		else
570 			uasm_i_daddiu(p, ptr, ptr, uasm_rel_lo(swpd));
571 	}
572 }
573 
574 #else /* !CONFIG_64BIT */
575 
576 /*
577  * TMP and PTR are scratch.
578  * TMP will be clobbered, PTR will hold the pgd entry.
579  */
580 static void __cpuinit __maybe_unused
581 build_get_pgde32(u32 **p, unsigned int tmp, unsigned int ptr)
582 {
583 	long pgdc = (long)pgd_current;
584 
585 	/* 32 bit SMP has smp_processor_id() stored in CONTEXT. */
586 #ifdef CONFIG_SMP
587 #ifdef  CONFIG_MIPS_MT_SMTC
588 	/*
589 	 * SMTC uses TCBind value as "CPU" index
590 	 */
591 	uasm_i_mfc0(p, ptr, C0_TCBIND);
592 	UASM_i_LA_mostly(p, tmp, pgdc);
593 	uasm_i_srl(p, ptr, ptr, 19);
594 #else
595 	/*
596 	 * smp_processor_id() << 3 is stored in CONTEXT.
597          */
598 	uasm_i_mfc0(p, ptr, C0_CONTEXT);
599 	UASM_i_LA_mostly(p, tmp, pgdc);
600 	uasm_i_srl(p, ptr, ptr, 23);
601 #endif
602 	uasm_i_addu(p, ptr, tmp, ptr);
603 #else
604 	UASM_i_LA_mostly(p, ptr, pgdc);
605 #endif
606 	uasm_i_mfc0(p, tmp, C0_BADVADDR); /* get faulting address */
607 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
608 	uasm_i_srl(p, tmp, tmp, PGDIR_SHIFT); /* get pgd only bits */
609 	uasm_i_sll(p, tmp, tmp, PGD_T_LOG2);
610 	uasm_i_addu(p, ptr, ptr, tmp); /* add in pgd offset */
611 }
612 
613 #endif /* !CONFIG_64BIT */
614 
615 static void __cpuinit build_adjust_context(u32 **p, unsigned int ctx)
616 {
617 	unsigned int shift = 4 - (PTE_T_LOG2 + 1) + PAGE_SHIFT - 12;
618 	unsigned int mask = (PTRS_PER_PTE / 2 - 1) << (PTE_T_LOG2 + 1);
619 
620 	switch (current_cpu_type()) {
621 	case CPU_VR41XX:
622 	case CPU_VR4111:
623 	case CPU_VR4121:
624 	case CPU_VR4122:
625 	case CPU_VR4131:
626 	case CPU_VR4181:
627 	case CPU_VR4181A:
628 	case CPU_VR4133:
629 		shift += 2;
630 		break;
631 
632 	default:
633 		break;
634 	}
635 
636 	if (shift)
637 		UASM_i_SRL(p, ctx, ctx, shift);
638 	uasm_i_andi(p, ctx, ctx, mask);
639 }
640 
641 static void __cpuinit build_get_ptep(u32 **p, unsigned int tmp, unsigned int ptr)
642 {
643 	/*
644 	 * Bug workaround for the Nevada. It seems as if under certain
645 	 * circumstances the move from cp0_context might produce a
646 	 * bogus result when the mfc0 instruction and its consumer are
647 	 * in a different cacheline or a load instruction, probably any
648 	 * memory reference, is between them.
649 	 */
650 	switch (current_cpu_type()) {
651 	case CPU_NEVADA:
652 		UASM_i_LW(p, ptr, 0, ptr);
653 		GET_CONTEXT(p, tmp); /* get context reg */
654 		break;
655 
656 	default:
657 		GET_CONTEXT(p, tmp); /* get context reg */
658 		UASM_i_LW(p, ptr, 0, ptr);
659 		break;
660 	}
661 
662 	build_adjust_context(p, tmp);
663 	UASM_i_ADDU(p, ptr, ptr, tmp); /* add in offset */
664 }
665 
666 static void __cpuinit build_update_entries(u32 **p, unsigned int tmp,
667 					unsigned int ptep)
668 {
669 	/*
670 	 * 64bit address support (36bit on a 32bit CPU) in a 32bit
671 	 * Kernel is a special case. Only a few CPUs use it.
672 	 */
673 #ifdef CONFIG_64BIT_PHYS_ADDR
674 	if (cpu_has_64bits) {
675 		uasm_i_ld(p, tmp, 0, ptep); /* get even pte */
676 		uasm_i_ld(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
677 		uasm_i_dsrl(p, tmp, tmp, 6); /* convert to entrylo0 */
678 		uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
679 		uasm_i_dsrl(p, ptep, ptep, 6); /* convert to entrylo1 */
680 		uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
681 	} else {
682 		int pte_off_even = sizeof(pte_t) / 2;
683 		int pte_off_odd = pte_off_even + sizeof(pte_t);
684 
685 		/* The pte entries are pre-shifted */
686 		uasm_i_lw(p, tmp, pte_off_even, ptep); /* get even pte */
687 		uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
688 		uasm_i_lw(p, ptep, pte_off_odd, ptep); /* get odd pte */
689 		uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
690 	}
691 #else
692 	UASM_i_LW(p, tmp, 0, ptep); /* get even pte */
693 	UASM_i_LW(p, ptep, sizeof(pte_t), ptep); /* get odd pte */
694 	if (r45k_bvahwbug())
695 		build_tlb_probe_entry(p);
696 	UASM_i_SRL(p, tmp, tmp, 6); /* convert to entrylo0 */
697 	if (r4k_250MHZhwbug())
698 		uasm_i_mtc0(p, 0, C0_ENTRYLO0);
699 	uasm_i_mtc0(p, tmp, C0_ENTRYLO0); /* load it */
700 	UASM_i_SRL(p, ptep, ptep, 6); /* convert to entrylo1 */
701 	if (r45k_bvahwbug())
702 		uasm_i_mfc0(p, tmp, C0_INDEX);
703 	if (r4k_250MHZhwbug())
704 		uasm_i_mtc0(p, 0, C0_ENTRYLO1);
705 	uasm_i_mtc0(p, ptep, C0_ENTRYLO1); /* load it */
706 #endif
707 }
708 
709 /*
710  * For a 64-bit kernel, we are using the 64-bit XTLB refill exception
711  * because EXL == 0.  If we wrap, we can also use the 32 instruction
712  * slots before the XTLB refill exception handler which belong to the
713  * unused TLB refill exception.
714  */
715 #define MIPS64_REFILL_INSNS 32
716 
717 static void __cpuinit build_r4000_tlb_refill_handler(void)
718 {
719 	u32 *p = tlb_handler;
720 	struct uasm_label *l = labels;
721 	struct uasm_reloc *r = relocs;
722 	u32 *f;
723 	unsigned int final_len;
724 
725 	memset(tlb_handler, 0, sizeof(tlb_handler));
726 	memset(labels, 0, sizeof(labels));
727 	memset(relocs, 0, sizeof(relocs));
728 	memset(final_handler, 0, sizeof(final_handler));
729 
730 	/*
731 	 * create the plain linear handler
732 	 */
733 	if (bcm1250_m3_war()) {
734 		UASM_i_MFC0(&p, K0, C0_BADVADDR);
735 		UASM_i_MFC0(&p, K1, C0_ENTRYHI);
736 		uasm_i_xor(&p, K0, K0, K1);
737 		UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
738 		uasm_il_bnez(&p, &r, K0, label_leave);
739 		/* No need for uasm_i_nop */
740 	}
741 
742 #ifdef CONFIG_64BIT
743 	build_get_pmde64(&p, &l, &r, K0, K1); /* get pmd in K1 */
744 #else
745 	build_get_pgde32(&p, K0, K1); /* get pgd in K1 */
746 #endif
747 
748 #ifdef CONFIG_HUGETLB_PAGE
749 	build_is_huge_pte(&p, &r, K0, K1, label_tlb_huge_update);
750 #endif
751 
752 	build_get_ptep(&p, K0, K1);
753 	build_update_entries(&p, K0, K1);
754 	build_tlb_write_entry(&p, &l, &r, tlb_random);
755 	uasm_l_leave(&l, p);
756 	uasm_i_eret(&p); /* return from trap */
757 
758 #ifdef CONFIG_HUGETLB_PAGE
759 	uasm_l_tlb_huge_update(&l, p);
760 	UASM_i_LW(&p, K0, 0, K1);
761 	build_huge_update_entries(&p, K0, K1);
762 	build_huge_tlb_write_entry(&p, &l, &r, K0, tlb_random);
763 #endif
764 
765 #ifdef CONFIG_64BIT
766 	build_get_pgd_vmalloc64(&p, &l, &r, K0, K1);
767 #endif
768 
769 	/*
770 	 * Overflow check: For the 64bit handler, we need at least one
771 	 * free instruction slot for the wrap-around branch. In worst
772 	 * case, if the intended insertion point is a delay slot, we
773 	 * need three, with the second nop'ed and the third being
774 	 * unused.
775 	 */
776 	/* Loongson2 ebase is different than r4k, we have more space */
777 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
778 	if ((p - tlb_handler) > 64)
779 		panic("TLB refill handler space exceeded");
780 #else
781 	if (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 1)
782 	    || (((p - tlb_handler) > (MIPS64_REFILL_INSNS * 2) - 3)
783 		&& uasm_insn_has_bdelay(relocs,
784 					tlb_handler + MIPS64_REFILL_INSNS - 3)))
785 		panic("TLB refill handler space exceeded");
786 #endif
787 
788 	/*
789 	 * Now fold the handler in the TLB refill handler space.
790 	 */
791 #if defined(CONFIG_32BIT) || defined(CONFIG_CPU_LOONGSON2)
792 	f = final_handler;
793 	/* Simplest case, just copy the handler. */
794 	uasm_copy_handler(relocs, labels, tlb_handler, p, f);
795 	final_len = p - tlb_handler;
796 #else /* CONFIG_64BIT */
797 	f = final_handler + MIPS64_REFILL_INSNS;
798 	if ((p - tlb_handler) <= MIPS64_REFILL_INSNS) {
799 		/* Just copy the handler. */
800 		uasm_copy_handler(relocs, labels, tlb_handler, p, f);
801 		final_len = p - tlb_handler;
802 	} else {
803 #if defined(CONFIG_HUGETLB_PAGE)
804 		const enum label_id ls = label_tlb_huge_update;
805 #elif defined(MODULE_START)
806 		const enum label_id ls = label_module_alloc;
807 #else
808 		const enum label_id ls = label_vmalloc;
809 #endif
810 		u32 *split;
811 		int ov = 0;
812 		int i;
813 
814 		for (i = 0; i < ARRAY_SIZE(labels) && labels[i].lab != ls; i++)
815 			;
816 		BUG_ON(i == ARRAY_SIZE(labels));
817 		split = labels[i].addr;
818 
819 		/*
820 		 * See if we have overflown one way or the other.
821 		 */
822 		if (split > tlb_handler + MIPS64_REFILL_INSNS ||
823 		    split < p - MIPS64_REFILL_INSNS)
824 			ov = 1;
825 
826 		if (ov) {
827 			/*
828 			 * Split two instructions before the end.  One
829 			 * for the branch and one for the instruction
830 			 * in the delay slot.
831 			 */
832 			split = tlb_handler + MIPS64_REFILL_INSNS - 2;
833 
834 			/*
835 			 * If the branch would fall in a delay slot,
836 			 * we must back up an additional instruction
837 			 * so that it is no longer in a delay slot.
838 			 */
839 			if (uasm_insn_has_bdelay(relocs, split - 1))
840 				split--;
841 		}
842 		/* Copy first part of the handler. */
843 		uasm_copy_handler(relocs, labels, tlb_handler, split, f);
844 		f += split - tlb_handler;
845 
846 		if (ov) {
847 			/* Insert branch. */
848 			uasm_l_split(&l, final_handler);
849 			uasm_il_b(&f, &r, label_split);
850 			if (uasm_insn_has_bdelay(relocs, split))
851 				uasm_i_nop(&f);
852 			else {
853 				uasm_copy_handler(relocs, labels,
854 						  split, split + 1, f);
855 				uasm_move_labels(labels, f, f + 1, -1);
856 				f++;
857 				split++;
858 			}
859 		}
860 
861 		/* Copy the rest of the handler. */
862 		uasm_copy_handler(relocs, labels, split, p, final_handler);
863 		final_len = (f - (final_handler + MIPS64_REFILL_INSNS)) +
864 			    (p - split);
865 	}
866 #endif /* CONFIG_64BIT */
867 
868 	uasm_resolve_relocs(relocs, labels);
869 	pr_debug("Wrote TLB refill handler (%u instructions).\n",
870 		 final_len);
871 
872 	memcpy((void *)ebase, final_handler, 0x100);
873 
874 	dump_handler((u32 *)ebase, 64);
875 }
876 
877 /*
878  * TLB load/store/modify handlers.
879  *
880  * Only the fastpath gets synthesized at runtime, the slowpath for
881  * do_page_fault remains normal asm.
882  */
883 extern void tlb_do_page_fault_0(void);
884 extern void tlb_do_page_fault_1(void);
885 
886 /*
887  * 128 instructions for the fastpath handler is generous and should
888  * never be exceeded.
889  */
890 #define FASTPATH_SIZE 128
891 
892 u32 handle_tlbl[FASTPATH_SIZE] __cacheline_aligned;
893 u32 handle_tlbs[FASTPATH_SIZE] __cacheline_aligned;
894 u32 handle_tlbm[FASTPATH_SIZE] __cacheline_aligned;
895 
896 static void __cpuinit
897 iPTE_LW(u32 **p, unsigned int pte, unsigned int ptr)
898 {
899 #ifdef CONFIG_SMP
900 # ifdef CONFIG_64BIT_PHYS_ADDR
901 	if (cpu_has_64bits)
902 		uasm_i_lld(p, pte, 0, ptr);
903 	else
904 # endif
905 		UASM_i_LL(p, pte, 0, ptr);
906 #else
907 # ifdef CONFIG_64BIT_PHYS_ADDR
908 	if (cpu_has_64bits)
909 		uasm_i_ld(p, pte, 0, ptr);
910 	else
911 # endif
912 		UASM_i_LW(p, pte, 0, ptr);
913 #endif
914 }
915 
916 static void __cpuinit
917 iPTE_SW(u32 **p, struct uasm_reloc **r, unsigned int pte, unsigned int ptr,
918 	unsigned int mode)
919 {
920 #ifdef CONFIG_64BIT_PHYS_ADDR
921 	unsigned int hwmode = mode & (_PAGE_VALID | _PAGE_DIRTY);
922 #endif
923 
924 	uasm_i_ori(p, pte, pte, mode);
925 #ifdef CONFIG_SMP
926 # ifdef CONFIG_64BIT_PHYS_ADDR
927 	if (cpu_has_64bits)
928 		uasm_i_scd(p, pte, 0, ptr);
929 	else
930 # endif
931 		UASM_i_SC(p, pte, 0, ptr);
932 
933 	if (r10000_llsc_war())
934 		uasm_il_beqzl(p, r, pte, label_smp_pgtable_change);
935 	else
936 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
937 
938 # ifdef CONFIG_64BIT_PHYS_ADDR
939 	if (!cpu_has_64bits) {
940 		/* no uasm_i_nop needed */
941 		uasm_i_ll(p, pte, sizeof(pte_t) / 2, ptr);
942 		uasm_i_ori(p, pte, pte, hwmode);
943 		uasm_i_sc(p, pte, sizeof(pte_t) / 2, ptr);
944 		uasm_il_beqz(p, r, pte, label_smp_pgtable_change);
945 		/* no uasm_i_nop needed */
946 		uasm_i_lw(p, pte, 0, ptr);
947 	} else
948 		uasm_i_nop(p);
949 # else
950 	uasm_i_nop(p);
951 # endif
952 #else
953 # ifdef CONFIG_64BIT_PHYS_ADDR
954 	if (cpu_has_64bits)
955 		uasm_i_sd(p, pte, 0, ptr);
956 	else
957 # endif
958 		UASM_i_SW(p, pte, 0, ptr);
959 
960 # ifdef CONFIG_64BIT_PHYS_ADDR
961 	if (!cpu_has_64bits) {
962 		uasm_i_lw(p, pte, sizeof(pte_t) / 2, ptr);
963 		uasm_i_ori(p, pte, pte, hwmode);
964 		uasm_i_sw(p, pte, sizeof(pte_t) / 2, ptr);
965 		uasm_i_lw(p, pte, 0, ptr);
966 	}
967 # endif
968 #endif
969 }
970 
971 /*
972  * Check if PTE is present, if not then jump to LABEL. PTR points to
973  * the page table where this PTE is located, PTE will be re-loaded
974  * with it's original value.
975  */
976 static void __cpuinit
977 build_pte_present(u32 **p, struct uasm_reloc **r,
978 		  unsigned int pte, unsigned int ptr, enum label_id lid)
979 {
980 	uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
981 	uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_READ);
982 	uasm_il_bnez(p, r, pte, lid);
983 	iPTE_LW(p, pte, ptr);
984 }
985 
986 /* Make PTE valid, store result in PTR. */
987 static void __cpuinit
988 build_make_valid(u32 **p, struct uasm_reloc **r, unsigned int pte,
989 		 unsigned int ptr)
990 {
991 	unsigned int mode = _PAGE_VALID | _PAGE_ACCESSED;
992 
993 	iPTE_SW(p, r, pte, ptr, mode);
994 }
995 
996 /*
997  * Check if PTE can be written to, if not branch to LABEL. Regardless
998  * restore PTE with value from PTR when done.
999  */
1000 static void __cpuinit
1001 build_pte_writable(u32 **p, struct uasm_reloc **r,
1002 		   unsigned int pte, unsigned int ptr, enum label_id lid)
1003 {
1004 	uasm_i_andi(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1005 	uasm_i_xori(p, pte, pte, _PAGE_PRESENT | _PAGE_WRITE);
1006 	uasm_il_bnez(p, r, pte, lid);
1007 	iPTE_LW(p, pte, ptr);
1008 }
1009 
1010 /* Make PTE writable, update software status bits as well, then store
1011  * at PTR.
1012  */
1013 static void __cpuinit
1014 build_make_write(u32 **p, struct uasm_reloc **r, unsigned int pte,
1015 		 unsigned int ptr)
1016 {
1017 	unsigned int mode = (_PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID
1018 			     | _PAGE_DIRTY);
1019 
1020 	iPTE_SW(p, r, pte, ptr, mode);
1021 }
1022 
1023 /*
1024  * Check if PTE can be modified, if not branch to LABEL. Regardless
1025  * restore PTE with value from PTR when done.
1026  */
1027 static void __cpuinit
1028 build_pte_modifiable(u32 **p, struct uasm_reloc **r,
1029 		     unsigned int pte, unsigned int ptr, enum label_id lid)
1030 {
1031 	uasm_i_andi(p, pte, pte, _PAGE_WRITE);
1032 	uasm_il_beqz(p, r, pte, lid);
1033 	iPTE_LW(p, pte, ptr);
1034 }
1035 
1036 /*
1037  * R3000 style TLB load/store/modify handlers.
1038  */
1039 
1040 /*
1041  * This places the pte into ENTRYLO0 and writes it with tlbwi.
1042  * Then it returns.
1043  */
1044 static void __cpuinit
1045 build_r3000_pte_reload_tlbwi(u32 **p, unsigned int pte, unsigned int tmp)
1046 {
1047 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1048 	uasm_i_mfc0(p, tmp, C0_EPC); /* cp0 delay */
1049 	uasm_i_tlbwi(p);
1050 	uasm_i_jr(p, tmp);
1051 	uasm_i_rfe(p); /* branch delay */
1052 }
1053 
1054 /*
1055  * This places the pte into ENTRYLO0 and writes it with tlbwi
1056  * or tlbwr as appropriate.  This is because the index register
1057  * may have the probe fail bit set as a result of a trap on a
1058  * kseg2 access, i.e. without refill.  Then it returns.
1059  */
1060 static void __cpuinit
1061 build_r3000_tlb_reload_write(u32 **p, struct uasm_label **l,
1062 			     struct uasm_reloc **r, unsigned int pte,
1063 			     unsigned int tmp)
1064 {
1065 	uasm_i_mfc0(p, tmp, C0_INDEX);
1066 	uasm_i_mtc0(p, pte, C0_ENTRYLO0); /* cp0 delay */
1067 	uasm_il_bltz(p, r, tmp, label_r3000_write_probe_fail); /* cp0 delay */
1068 	uasm_i_mfc0(p, tmp, C0_EPC); /* branch delay */
1069 	uasm_i_tlbwi(p); /* cp0 delay */
1070 	uasm_i_jr(p, tmp);
1071 	uasm_i_rfe(p); /* branch delay */
1072 	uasm_l_r3000_write_probe_fail(l, *p);
1073 	uasm_i_tlbwr(p); /* cp0 delay */
1074 	uasm_i_jr(p, tmp);
1075 	uasm_i_rfe(p); /* branch delay */
1076 }
1077 
1078 static void __cpuinit
1079 build_r3000_tlbchange_handler_head(u32 **p, unsigned int pte,
1080 				   unsigned int ptr)
1081 {
1082 	long pgdc = (long)pgd_current;
1083 
1084 	uasm_i_mfc0(p, pte, C0_BADVADDR);
1085 	uasm_i_lui(p, ptr, uasm_rel_hi(pgdc)); /* cp0 delay */
1086 	uasm_i_lw(p, ptr, uasm_rel_lo(pgdc), ptr);
1087 	uasm_i_srl(p, pte, pte, 22); /* load delay */
1088 	uasm_i_sll(p, pte, pte, 2);
1089 	uasm_i_addu(p, ptr, ptr, pte);
1090 	uasm_i_mfc0(p, pte, C0_CONTEXT);
1091 	uasm_i_lw(p, ptr, 0, ptr); /* cp0 delay */
1092 	uasm_i_andi(p, pte, pte, 0xffc); /* load delay */
1093 	uasm_i_addu(p, ptr, ptr, pte);
1094 	uasm_i_lw(p, pte, 0, ptr);
1095 	uasm_i_tlbp(p); /* load delay */
1096 }
1097 
1098 static void __cpuinit build_r3000_tlb_load_handler(void)
1099 {
1100 	u32 *p = handle_tlbl;
1101 	struct uasm_label *l = labels;
1102 	struct uasm_reloc *r = relocs;
1103 
1104 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1105 	memset(labels, 0, sizeof(labels));
1106 	memset(relocs, 0, sizeof(relocs));
1107 
1108 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1109 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1110 	uasm_i_nop(&p); /* load delay */
1111 	build_make_valid(&p, &r, K0, K1);
1112 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1113 
1114 	uasm_l_nopage_tlbl(&l, p);
1115 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1116 	uasm_i_nop(&p);
1117 
1118 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1119 		panic("TLB load handler fastpath space exceeded");
1120 
1121 	uasm_resolve_relocs(relocs, labels);
1122 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1123 		 (unsigned int)(p - handle_tlbl));
1124 
1125 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1126 }
1127 
1128 static void __cpuinit build_r3000_tlb_store_handler(void)
1129 {
1130 	u32 *p = handle_tlbs;
1131 	struct uasm_label *l = labels;
1132 	struct uasm_reloc *r = relocs;
1133 
1134 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1135 	memset(labels, 0, sizeof(labels));
1136 	memset(relocs, 0, sizeof(relocs));
1137 
1138 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1139 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1140 	uasm_i_nop(&p); /* load delay */
1141 	build_make_write(&p, &r, K0, K1);
1142 	build_r3000_tlb_reload_write(&p, &l, &r, K0, K1);
1143 
1144 	uasm_l_nopage_tlbs(&l, p);
1145 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1146 	uasm_i_nop(&p);
1147 
1148 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1149 		panic("TLB store handler fastpath space exceeded");
1150 
1151 	uasm_resolve_relocs(relocs, labels);
1152 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1153 		 (unsigned int)(p - handle_tlbs));
1154 
1155 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1156 }
1157 
1158 static void __cpuinit build_r3000_tlb_modify_handler(void)
1159 {
1160 	u32 *p = handle_tlbm;
1161 	struct uasm_label *l = labels;
1162 	struct uasm_reloc *r = relocs;
1163 
1164 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1165 	memset(labels, 0, sizeof(labels));
1166 	memset(relocs, 0, sizeof(relocs));
1167 
1168 	build_r3000_tlbchange_handler_head(&p, K0, K1);
1169 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1170 	uasm_i_nop(&p); /* load delay */
1171 	build_make_write(&p, &r, K0, K1);
1172 	build_r3000_pte_reload_tlbwi(&p, K0, K1);
1173 
1174 	uasm_l_nopage_tlbm(&l, p);
1175 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1176 	uasm_i_nop(&p);
1177 
1178 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1179 		panic("TLB modify handler fastpath space exceeded");
1180 
1181 	uasm_resolve_relocs(relocs, labels);
1182 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1183 		 (unsigned int)(p - handle_tlbm));
1184 
1185 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1186 }
1187 
1188 /*
1189  * R4000 style TLB load/store/modify handlers.
1190  */
1191 static void __cpuinit
1192 build_r4000_tlbchange_handler_head(u32 **p, struct uasm_label **l,
1193 				   struct uasm_reloc **r, unsigned int pte,
1194 				   unsigned int ptr)
1195 {
1196 #ifdef CONFIG_64BIT
1197 	build_get_pmde64(p, l, r, pte, ptr); /* get pmd in ptr */
1198 #else
1199 	build_get_pgde32(p, pte, ptr); /* get pgd in ptr */
1200 #endif
1201 
1202 #ifdef CONFIG_HUGETLB_PAGE
1203 	/*
1204 	 * For huge tlb entries, pmd doesn't contain an address but
1205 	 * instead contains the tlb pte. Check the PAGE_HUGE bit and
1206 	 * see if we need to jump to huge tlb processing.
1207 	 */
1208 	build_is_huge_pte(p, r, pte, ptr, label_tlb_huge_update);
1209 #endif
1210 
1211 	UASM_i_MFC0(p, pte, C0_BADVADDR);
1212 	UASM_i_LW(p, ptr, 0, ptr);
1213 	UASM_i_SRL(p, pte, pte, PAGE_SHIFT + PTE_ORDER - PTE_T_LOG2);
1214 	uasm_i_andi(p, pte, pte, (PTRS_PER_PTE - 1) << PTE_T_LOG2);
1215 	UASM_i_ADDU(p, ptr, ptr, pte);
1216 
1217 #ifdef CONFIG_SMP
1218 	uasm_l_smp_pgtable_change(l, *p);
1219 #endif
1220 	iPTE_LW(p, pte, ptr); /* get even pte */
1221 	if (!m4kc_tlbp_war())
1222 		build_tlb_probe_entry(p);
1223 }
1224 
1225 static void __cpuinit
1226 build_r4000_tlbchange_handler_tail(u32 **p, struct uasm_label **l,
1227 				   struct uasm_reloc **r, unsigned int tmp,
1228 				   unsigned int ptr)
1229 {
1230 	uasm_i_ori(p, ptr, ptr, sizeof(pte_t));
1231 	uasm_i_xori(p, ptr, ptr, sizeof(pte_t));
1232 	build_update_entries(p, tmp, ptr);
1233 	build_tlb_write_entry(p, l, r, tlb_indexed);
1234 	uasm_l_leave(l, *p);
1235 	uasm_i_eret(p); /* return from trap */
1236 
1237 #ifdef CONFIG_64BIT
1238 	build_get_pgd_vmalloc64(p, l, r, tmp, ptr);
1239 #endif
1240 }
1241 
1242 static void __cpuinit build_r4000_tlb_load_handler(void)
1243 {
1244 	u32 *p = handle_tlbl;
1245 	struct uasm_label *l = labels;
1246 	struct uasm_reloc *r = relocs;
1247 
1248 	memset(handle_tlbl, 0, sizeof(handle_tlbl));
1249 	memset(labels, 0, sizeof(labels));
1250 	memset(relocs, 0, sizeof(relocs));
1251 
1252 	if (bcm1250_m3_war()) {
1253 		UASM_i_MFC0(&p, K0, C0_BADVADDR);
1254 		UASM_i_MFC0(&p, K1, C0_ENTRYHI);
1255 		uasm_i_xor(&p, K0, K0, K1);
1256 		UASM_i_SRL(&p, K0, K0, PAGE_SHIFT + 1);
1257 		uasm_il_bnez(&p, &r, K0, label_leave);
1258 		/* No need for uasm_i_nop */
1259 	}
1260 
1261 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1262 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1263 	if (m4kc_tlbp_war())
1264 		build_tlb_probe_entry(&p);
1265 	build_make_valid(&p, &r, K0, K1);
1266 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1267 
1268 #ifdef CONFIG_HUGETLB_PAGE
1269 	/*
1270 	 * This is the entry point when build_r4000_tlbchange_handler_head
1271 	 * spots a huge page.
1272 	 */
1273 	uasm_l_tlb_huge_update(&l, p);
1274 	iPTE_LW(&p, K0, K1);
1275 	build_pte_present(&p, &r, K0, K1, label_nopage_tlbl);
1276 	build_tlb_probe_entry(&p);
1277 	uasm_i_ori(&p, K0, K0, (_PAGE_ACCESSED | _PAGE_VALID));
1278 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1279 #endif
1280 
1281 	uasm_l_nopage_tlbl(&l, p);
1282 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_0 & 0x0fffffff);
1283 	uasm_i_nop(&p);
1284 
1285 	if ((p - handle_tlbl) > FASTPATH_SIZE)
1286 		panic("TLB load handler fastpath space exceeded");
1287 
1288 	uasm_resolve_relocs(relocs, labels);
1289 	pr_debug("Wrote TLB load handler fastpath (%u instructions).\n",
1290 		 (unsigned int)(p - handle_tlbl));
1291 
1292 	dump_handler(handle_tlbl, ARRAY_SIZE(handle_tlbl));
1293 }
1294 
1295 static void __cpuinit build_r4000_tlb_store_handler(void)
1296 {
1297 	u32 *p = handle_tlbs;
1298 	struct uasm_label *l = labels;
1299 	struct uasm_reloc *r = relocs;
1300 
1301 	memset(handle_tlbs, 0, sizeof(handle_tlbs));
1302 	memset(labels, 0, sizeof(labels));
1303 	memset(relocs, 0, sizeof(relocs));
1304 
1305 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1306 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1307 	if (m4kc_tlbp_war())
1308 		build_tlb_probe_entry(&p);
1309 	build_make_write(&p, &r, K0, K1);
1310 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1311 
1312 #ifdef CONFIG_HUGETLB_PAGE
1313 	/*
1314 	 * This is the entry point when
1315 	 * build_r4000_tlbchange_handler_head spots a huge page.
1316 	 */
1317 	uasm_l_tlb_huge_update(&l, p);
1318 	iPTE_LW(&p, K0, K1);
1319 	build_pte_writable(&p, &r, K0, K1, label_nopage_tlbs);
1320 	build_tlb_probe_entry(&p);
1321 	uasm_i_ori(&p, K0, K0,
1322 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1323 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1324 #endif
1325 
1326 	uasm_l_nopage_tlbs(&l, p);
1327 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1328 	uasm_i_nop(&p);
1329 
1330 	if ((p - handle_tlbs) > FASTPATH_SIZE)
1331 		panic("TLB store handler fastpath space exceeded");
1332 
1333 	uasm_resolve_relocs(relocs, labels);
1334 	pr_debug("Wrote TLB store handler fastpath (%u instructions).\n",
1335 		 (unsigned int)(p - handle_tlbs));
1336 
1337 	dump_handler(handle_tlbs, ARRAY_SIZE(handle_tlbs));
1338 }
1339 
1340 static void __cpuinit build_r4000_tlb_modify_handler(void)
1341 {
1342 	u32 *p = handle_tlbm;
1343 	struct uasm_label *l = labels;
1344 	struct uasm_reloc *r = relocs;
1345 
1346 	memset(handle_tlbm, 0, sizeof(handle_tlbm));
1347 	memset(labels, 0, sizeof(labels));
1348 	memset(relocs, 0, sizeof(relocs));
1349 
1350 	build_r4000_tlbchange_handler_head(&p, &l, &r, K0, K1);
1351 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1352 	if (m4kc_tlbp_war())
1353 		build_tlb_probe_entry(&p);
1354 	/* Present and writable bits set, set accessed and dirty bits. */
1355 	build_make_write(&p, &r, K0, K1);
1356 	build_r4000_tlbchange_handler_tail(&p, &l, &r, K0, K1);
1357 
1358 #ifdef CONFIG_HUGETLB_PAGE
1359 	/*
1360 	 * This is the entry point when
1361 	 * build_r4000_tlbchange_handler_head spots a huge page.
1362 	 */
1363 	uasm_l_tlb_huge_update(&l, p);
1364 	iPTE_LW(&p, K0, K1);
1365 	build_pte_modifiable(&p, &r, K0, K1, label_nopage_tlbm);
1366 	build_tlb_probe_entry(&p);
1367 	uasm_i_ori(&p, K0, K0,
1368 		   _PAGE_ACCESSED | _PAGE_MODIFIED | _PAGE_VALID | _PAGE_DIRTY);
1369 	build_huge_handler_tail(&p, &r, &l, K0, K1);
1370 #endif
1371 
1372 	uasm_l_nopage_tlbm(&l, p);
1373 	uasm_i_j(&p, (unsigned long)tlb_do_page_fault_1 & 0x0fffffff);
1374 	uasm_i_nop(&p);
1375 
1376 	if ((p - handle_tlbm) > FASTPATH_SIZE)
1377 		panic("TLB modify handler fastpath space exceeded");
1378 
1379 	uasm_resolve_relocs(relocs, labels);
1380 	pr_debug("Wrote TLB modify handler fastpath (%u instructions).\n",
1381 		 (unsigned int)(p - handle_tlbm));
1382 
1383 	dump_handler(handle_tlbm, ARRAY_SIZE(handle_tlbm));
1384 }
1385 
1386 void __cpuinit build_tlb_refill_handler(void)
1387 {
1388 	/*
1389 	 * The refill handler is generated per-CPU, multi-node systems
1390 	 * may have local storage for it. The other handlers are only
1391 	 * needed once.
1392 	 */
1393 	static int run_once = 0;
1394 
1395 	switch (current_cpu_type()) {
1396 	case CPU_R2000:
1397 	case CPU_R3000:
1398 	case CPU_R3000A:
1399 	case CPU_R3081E:
1400 	case CPU_TX3912:
1401 	case CPU_TX3922:
1402 	case CPU_TX3927:
1403 		build_r3000_tlb_refill_handler();
1404 		if (!run_once) {
1405 			build_r3000_tlb_load_handler();
1406 			build_r3000_tlb_store_handler();
1407 			build_r3000_tlb_modify_handler();
1408 			run_once++;
1409 		}
1410 		break;
1411 
1412 	case CPU_R6000:
1413 	case CPU_R6000A:
1414 		panic("No R6000 TLB refill handler yet");
1415 		break;
1416 
1417 	case CPU_R8000:
1418 		panic("No R8000 TLB refill handler yet");
1419 		break;
1420 
1421 	default:
1422 		build_r4000_tlb_refill_handler();
1423 		if (!run_once) {
1424 			build_r4000_tlb_load_handler();
1425 			build_r4000_tlb_store_handler();
1426 			build_r4000_tlb_modify_handler();
1427 			run_once++;
1428 		}
1429 	}
1430 }
1431 
1432 void __cpuinit flush_tlb_handlers(void)
1433 {
1434 	local_flush_icache_range((unsigned long)handle_tlbl,
1435 			   (unsigned long)handle_tlbl + sizeof(handle_tlbl));
1436 	local_flush_icache_range((unsigned long)handle_tlbs,
1437 			   (unsigned long)handle_tlbs + sizeof(handle_tlbs));
1438 	local_flush_icache_range((unsigned long)handle_tlbm,
1439 			   (unsigned long)handle_tlbm + sizeof(handle_tlbm));
1440 }
1441