?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
mm: fix exec activate_mm vs TLB shootdown and lazy tlb switching race
[linux/fpc-iii.git] / arch / tile / kernel / unaligned.c
blob8149c38f67b6cc5ea6fa4daafae908e330d5dcf1
1 /*
2 * Copyright 2013 Tilera Corporation. All Rights Reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation, version 2.
7 *
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11 * NON INFRINGEMENT. See the GNU General Public License for
12 * more details.
13 *
14 * A code-rewriter that handles unaligned exception.
15 */
16
17 #include <linux/smp.h>
18 #include <linux/ptrace.h>
19 #include <linux/slab.h>
20 #include <linux/sched/debug.h>
21 #include <linux/sched/task.h>
22 #include <linux/thread_info.h>
23 #include <linux/uaccess.h>
24 #include <linux/mman.h>
25 #include <linux/types.h>
26 #include <linux/err.h>
27 #include <linux/extable.h>
28 #include <linux/compat.h>
29 #include <linux/prctl.h>
30 #include <asm/cacheflush.h>
31 #include <asm/traps.h>
32 #include <linux/uaccess.h>
33 #include <asm/unaligned.h>
34 #include <arch/abi.h>
35 #include <arch/spr_def.h>
36 #include <arch/opcode.h>
37
38
39 /*
40 * This file handles unaligned exception for tile-Gx. The tilepro's unaligned
41 * exception is supported out of single_step.c
42 */
43
44 int unaligned_printk;
45
46 static int __init setup_unaligned_printk(char *str)
47 {
48 long val;
49 if (kstrtol(str, 0, &val) != 0)
50 return 0;
51 unaligned_printk = val;
52 pr_info("Printk for each unaligned data accesses is %s\n",
53 unaligned_printk ? "enabled" : "disabled");
54 return 1;
55 }
56 __setup("unaligned_printk=", setup_unaligned_printk);
57
58 unsigned int unaligned_fixup_count;
59
60 #ifdef __tilegx__
61
62 /*
63 * Unalign data jit fixup code fragement. Reserved space is 128 bytes.
64 * The 1st 64-bit word saves fault PC address, 2nd word is the fault
65 * instruction bundle followed by 14 JIT bundles.
66 */
67
68 struct unaligned_jit_fragment {
69 unsigned long pc;
70 tilegx_bundle_bits bundle;
71 tilegx_bundle_bits insn[14];
72 };
73
74 /*
75 * Check if a nop or fnop at bundle's pipeline X0.
76 */
77
78 static bool is_bundle_x0_nop(tilegx_bundle_bits bundle)
79 {
80 return (((get_UnaryOpcodeExtension_X0(bundle) ==
81 NOP_UNARY_OPCODE_X0) &&
82 (get_RRROpcodeExtension_X0(bundle) ==
83 UNARY_RRR_0_OPCODE_X0) &&
84 (get_Opcode_X0(bundle) ==
85 RRR_0_OPCODE_X0)) ||
86 ((get_UnaryOpcodeExtension_X0(bundle) ==
87 FNOP_UNARY_OPCODE_X0) &&
88 (get_RRROpcodeExtension_X0(bundle) ==
89 UNARY_RRR_0_OPCODE_X0) &&
90 (get_Opcode_X0(bundle) ==
91 RRR_0_OPCODE_X0)));
92 }
93
94 /*
95 * Check if nop or fnop at bundle's pipeline X1.
96 */
97
98 static bool is_bundle_x1_nop(tilegx_bundle_bits bundle)
99 {
100 return (((get_UnaryOpcodeExtension_X1(bundle) ==
101 NOP_UNARY_OPCODE_X1) &&
102 (get_RRROpcodeExtension_X1(bundle) ==
103 UNARY_RRR_0_OPCODE_X1) &&
104 (get_Opcode_X1(bundle) ==
105 RRR_0_OPCODE_X1)) ||
106 ((get_UnaryOpcodeExtension_X1(bundle) ==
107 FNOP_UNARY_OPCODE_X1) &&
108 (get_RRROpcodeExtension_X1(bundle) ==
109 UNARY_RRR_0_OPCODE_X1) &&
110 (get_Opcode_X1(bundle) ==
111 RRR_0_OPCODE_X1)));
112 }
113
114 /*
115 * Check if nop or fnop at bundle's Y0 pipeline.
116 */
117
118 static bool is_bundle_y0_nop(tilegx_bundle_bits bundle)
119 {
120 return (((get_UnaryOpcodeExtension_Y0(bundle) ==
121 NOP_UNARY_OPCODE_Y0) &&
122 (get_RRROpcodeExtension_Y0(bundle) ==
123 UNARY_RRR_1_OPCODE_Y0) &&
124 (get_Opcode_Y0(bundle) ==
125 RRR_1_OPCODE_Y0)) ||
126 ((get_UnaryOpcodeExtension_Y0(bundle) ==
127 FNOP_UNARY_OPCODE_Y0) &&
128 (get_RRROpcodeExtension_Y0(bundle) ==
129 UNARY_RRR_1_OPCODE_Y0) &&
130 (get_Opcode_Y0(bundle) ==
131 RRR_1_OPCODE_Y0)));
132 }
133
134 /*
135 * Check if nop or fnop at bundle's pipeline Y1.
136 */
137
138 static bool is_bundle_y1_nop(tilegx_bundle_bits bundle)
139 {
140 return (((get_UnaryOpcodeExtension_Y1(bundle) ==
141 NOP_UNARY_OPCODE_Y1) &&
142 (get_RRROpcodeExtension_Y1(bundle) ==
143 UNARY_RRR_1_OPCODE_Y1) &&
144 (get_Opcode_Y1(bundle) ==
145 RRR_1_OPCODE_Y1)) ||
146 ((get_UnaryOpcodeExtension_Y1(bundle) ==
147 FNOP_UNARY_OPCODE_Y1) &&
148 (get_RRROpcodeExtension_Y1(bundle) ==
149 UNARY_RRR_1_OPCODE_Y1) &&
150 (get_Opcode_Y1(bundle) ==
151 RRR_1_OPCODE_Y1)));
152 }
153
154 /*
155 * Test if a bundle's y0 and y1 pipelines are both nop or fnop.
156 */
157
158 static bool is_y0_y1_nop(tilegx_bundle_bits bundle)
159 {
160 return is_bundle_y0_nop(bundle) && is_bundle_y1_nop(bundle);
161 }
162
163 /*
164 * Test if a bundle's x0 and x1 pipelines are both nop or fnop.
165 */
166
167 static bool is_x0_x1_nop(tilegx_bundle_bits bundle)
168 {
169 return is_bundle_x0_nop(bundle) && is_bundle_x1_nop(bundle);
170 }
171
172 /*
173 * Find the destination, source registers of fault unalign access instruction
174 * at X1 or Y2. Also, allocate up to 3 scratch registers clob1, clob2 and
175 * clob3, which are guaranteed different from any register used in the fault
176 * bundle. r_alias is used to return if the other instructions other than the
177 * unalign load/store shares same register with ra, rb and rd.
178 */
179
180 static void find_regs(tilegx_bundle_bits bundle, uint64_t *rd, uint64_t *ra,
181 uint64_t *rb, uint64_t *clob1, uint64_t *clob2,
182 uint64_t *clob3, bool *r_alias)
183 {
184 int i;
185 uint64_t reg;
186 uint64_t reg_map = 0, alias_reg_map = 0, map;
187 bool alias = false;
188
189 /*
190 * Parse fault bundle, find potential used registers and mark
191 * corresponding bits in reg_map and alias_map. These 2 bit maps
192 * are used to find the scratch registers and determine if there
193 * is register alias.
194 */
195 if (bundle & TILEGX_BUNDLE_MODE_MASK) { /* Y Mode Bundle. */
196
197 reg = get_SrcA_Y2(bundle);
198 reg_map |= 1ULL << reg;
199 *ra = reg;
200 reg = get_SrcBDest_Y2(bundle);
201 reg_map |= 1ULL << reg;
202
203 if (rd) {
204 /* Load. */
205 *rd = reg;
206 alias_reg_map = (1ULL << *rd) | (1ULL << *ra);
207 } else {
208 /* Store. */
209 *rb = reg;
210 alias_reg_map = (1ULL << *ra) | (1ULL << *rb);
211 }
212
213 if (!is_bundle_y1_nop(bundle)) {
214 reg = get_SrcA_Y1(bundle);
215 reg_map |= (1ULL << reg);
216 map = (1ULL << reg);
217
218 reg = get_SrcB_Y1(bundle);
219 reg_map |= (1ULL << reg);
220 map |= (1ULL << reg);
221
222 reg = get_Dest_Y1(bundle);
223 reg_map |= (1ULL << reg);
224 map |= (1ULL << reg);
225
226 if (map & alias_reg_map)
227 alias = true;
228 }
229
230 if (!is_bundle_y0_nop(bundle)) {
231 reg = get_SrcA_Y0(bundle);
232 reg_map |= (1ULL << reg);
233 map = (1ULL << reg);
234
235 reg = get_SrcB_Y0(bundle);
236 reg_map |= (1ULL << reg);
237 map |= (1ULL << reg);
238
239 reg = get_Dest_Y0(bundle);
240 reg_map |= (1ULL << reg);
241 map |= (1ULL << reg);
242
243 if (map & alias_reg_map)
244 alias = true;
245 }
246 } else { /* X Mode Bundle. */
247
248 reg = get_SrcA_X1(bundle);
249 reg_map |= (1ULL << reg);
250 *ra = reg;
251 if (rd) {
252 /* Load. */
253 reg = get_Dest_X1(bundle);
254 reg_map |= (1ULL << reg);
255 *rd = reg;
256 alias_reg_map = (1ULL << *rd) | (1ULL << *ra);
257 } else {
258 /* Store. */
259 reg = get_SrcB_X1(bundle);
260 reg_map |= (1ULL << reg);
261 *rb = reg;
262 alias_reg_map = (1ULL << *ra) | (1ULL << *rb);
263 }
264
265 if (!is_bundle_x0_nop(bundle)) {
266 reg = get_SrcA_X0(bundle);
267 reg_map |= (1ULL << reg);
268 map = (1ULL << reg);
269
270 reg = get_SrcB_X0(bundle);
271 reg_map |= (1ULL << reg);
272 map |= (1ULL << reg);
273
274 reg = get_Dest_X0(bundle);
275 reg_map |= (1ULL << reg);
276 map |= (1ULL << reg);
277
278 if (map & alias_reg_map)
279 alias = true;
280 }
281 }
282
283 /*
284 * "alias" indicates if the unalign access registers have collision
285 * with others in the same bundle. We jsut simply test all register
286 * operands case (RRR), ignored the case with immidate. If a bundle
287 * has no register alias, we may do fixup in a simple or fast manner.
288 * So if an immidata field happens to hit with a register, we may end
289 * up fall back to the generic handling.
290 */
291
292 *r_alias = alias;
293
294 /* Flip bits on reg_map. */
295 reg_map ^= -1ULL;
296
297 /* Scan reg_map lower 54(TREG_SP) bits to find 3 set bits. */
298 for (i = 0; i < TREG_SP; i++) {
299 if (reg_map & (0x1ULL << i)) {
300 if (*clob1 == -1) {
301 *clob1 = i;
302 } else if (*clob2 == -1) {
303 *clob2 = i;
304 } else if (*clob3 == -1) {
305 *clob3 = i;
306 return;
307 }
308 }
309 }
310 }
311
312 /*
313 * Sanity check for register ra, rb, rd, clob1/2/3. Return true if any of them
314 * is unexpected.
315 */
316
317 static bool check_regs(uint64_t rd, uint64_t ra, uint64_t rb,
318 uint64_t clob1, uint64_t clob2, uint64_t clob3)
319 {
320 bool unexpected = false;
321 if ((ra >= 56) && (ra != TREG_ZERO))
322 unexpected = true;
323
324 if ((clob1 >= 56) || (clob2 >= 56) || (clob3 >= 56))
325 unexpected = true;
326
327 if (rd != -1) {
328 if ((rd >= 56) && (rd != TREG_ZERO))
329 unexpected = true;
330 } else {
331 if ((rb >= 56) && (rb != TREG_ZERO))
332 unexpected = true;
333 }
334 return unexpected;
335 }
336
337
338 #define GX_INSN_X0_MASK ((1ULL << 31) - 1)
339 #define GX_INSN_X1_MASK (((1ULL << 31) - 1) << 31)
340 #define GX_INSN_Y0_MASK ((0xFULL << 27) | (0xFFFFFULL))
341 #define GX_INSN_Y1_MASK (GX_INSN_Y0_MASK << 31)
342 #define GX_INSN_Y2_MASK ((0x7FULL << 51) | (0x7FULL << 20))
343
344 #ifdef __LITTLE_ENDIAN
345 #define GX_INSN_BSWAP(_bundle_) (_bundle_)
346 #else
347 #define GX_INSN_BSWAP(_bundle_) swab64(_bundle_)
348 #endif /* __LITTLE_ENDIAN */
349
350 /*
351 * __JIT_CODE(.) creates template bundles in .rodata.unalign_data section.
352 * The corresponding static function jix_x#_###(.) generates partial or
353 * whole bundle based on the template and given arguments.
354 */
355
356 #define __JIT_CODE(_X_) \
357 asm (".pushsection .rodata.unalign_data, \"a\"\n" \
358 _X_"\n" \
359 ".popsection\n")
360
361 __JIT_CODE("__unalign_jit_x1_mtspr: {mtspr 0, r0}");
362 static tilegx_bundle_bits jit_x1_mtspr(int spr, int reg)
363 {
364 extern tilegx_bundle_bits __unalign_jit_x1_mtspr;
365 return (GX_INSN_BSWAP(__unalign_jit_x1_mtspr) & GX_INSN_X1_MASK) |
366 create_MT_Imm14_X1(spr) | create_SrcA_X1(reg);
367 }
368
369 __JIT_CODE("__unalign_jit_x1_mfspr: {mfspr r0, 0}");
370 static tilegx_bundle_bits jit_x1_mfspr(int reg, int spr)
371 {
372 extern tilegx_bundle_bits __unalign_jit_x1_mfspr;
373 return (GX_INSN_BSWAP(__unalign_jit_x1_mfspr) & GX_INSN_X1_MASK) |
374 create_MF_Imm14_X1(spr) | create_Dest_X1(reg);
375 }
376
377 __JIT_CODE("__unalign_jit_x0_addi: {addi r0, r0, 0; iret}");
378 static tilegx_bundle_bits jit_x0_addi(int rd, int ra, int imm8)
379 {
380 extern tilegx_bundle_bits __unalign_jit_x0_addi;
381 return (GX_INSN_BSWAP(__unalign_jit_x0_addi) & GX_INSN_X0_MASK) |
382 create_Dest_X0(rd) | create_SrcA_X0(ra) |
383 create_Imm8_X0(imm8);
384 }
385
386 __JIT_CODE("__unalign_jit_x1_ldna: {ldna r0, r0}");
387 static tilegx_bundle_bits jit_x1_ldna(int rd, int ra)
388 {
389 extern tilegx_bundle_bits __unalign_jit_x1_ldna;
390 return (GX_INSN_BSWAP(__unalign_jit_x1_ldna) & GX_INSN_X1_MASK) |
391 create_Dest_X1(rd) | create_SrcA_X1(ra);
392 }
393
394 __JIT_CODE("__unalign_jit_x0_dblalign: {dblalign r0, r0 ,r0}");
395 static tilegx_bundle_bits jit_x0_dblalign(int rd, int ra, int rb)
396 {
397 extern tilegx_bundle_bits __unalign_jit_x0_dblalign;
398 return (GX_INSN_BSWAP(__unalign_jit_x0_dblalign) & GX_INSN_X0_MASK) |
399 create_Dest_X0(rd) | create_SrcA_X0(ra) |
400 create_SrcB_X0(rb);
401 }
402
403 __JIT_CODE("__unalign_jit_x1_iret: {iret}");
404 static tilegx_bundle_bits jit_x1_iret(void)
405 {
406 extern tilegx_bundle_bits __unalign_jit_x1_iret;
407 return GX_INSN_BSWAP(__unalign_jit_x1_iret) & GX_INSN_X1_MASK;
408 }
409
410 __JIT_CODE("__unalign_jit_x01_fnop: {fnop;fnop}");
411 static tilegx_bundle_bits jit_x0_fnop(void)
412 {
413 extern tilegx_bundle_bits __unalign_jit_x01_fnop;
414 return GX_INSN_BSWAP(__unalign_jit_x01_fnop) & GX_INSN_X0_MASK;
415 }
416
417 static tilegx_bundle_bits jit_x1_fnop(void)
418 {
419 extern tilegx_bundle_bits __unalign_jit_x01_fnop;
420 return GX_INSN_BSWAP(__unalign_jit_x01_fnop) & GX_INSN_X1_MASK;
421 }
422
423 __JIT_CODE("__unalign_jit_y2_dummy: {fnop; fnop; ld zero, sp}");
424 static tilegx_bundle_bits jit_y2_dummy(void)
425 {
426 extern tilegx_bundle_bits __unalign_jit_y2_dummy;
427 return GX_INSN_BSWAP(__unalign_jit_y2_dummy) & GX_INSN_Y2_MASK;
428 }
429
430 static tilegx_bundle_bits jit_y1_fnop(void)
431 {
432 extern tilegx_bundle_bits __unalign_jit_y2_dummy;
433 return GX_INSN_BSWAP(__unalign_jit_y2_dummy) & GX_INSN_Y1_MASK;
434 }
435
436 __JIT_CODE("__unalign_jit_x1_st1_add: {st1_add r1, r0, 0}");
437 static tilegx_bundle_bits jit_x1_st1_add(int ra, int rb, int imm8)
438 {
439 extern tilegx_bundle_bits __unalign_jit_x1_st1_add;
440 return (GX_INSN_BSWAP(__unalign_jit_x1_st1_add) &
441 (~create_SrcA_X1(-1)) &
442 GX_INSN_X1_MASK) | create_SrcA_X1(ra) |
443 create_SrcB_X1(rb) | create_Dest_Imm8_X1(imm8);
444 }
445
446 __JIT_CODE("__unalign_jit_x1_st: {crc32_8 r1, r0, r0; st r0, r0}");
447 static tilegx_bundle_bits jit_x1_st(int ra, int rb)
448 {
449 extern tilegx_bundle_bits __unalign_jit_x1_st;
450 return (GX_INSN_BSWAP(__unalign_jit_x1_st) & GX_INSN_X1_MASK) |
451 create_SrcA_X1(ra) | create_SrcB_X1(rb);
452 }
453
454 __JIT_CODE("__unalign_jit_x1_st_add: {st_add r1, r0, 0}");
455 static tilegx_bundle_bits jit_x1_st_add(int ra, int rb, int imm8)
456 {
457 extern tilegx_bundle_bits __unalign_jit_x1_st_add;
458 return (GX_INSN_BSWAP(__unalign_jit_x1_st_add) &
459 (~create_SrcA_X1(-1)) &
460 GX_INSN_X1_MASK) | create_SrcA_X1(ra) |
461 create_SrcB_X1(rb) | create_Dest_Imm8_X1(imm8);
462 }
463
464 __JIT_CODE("__unalign_jit_x1_ld: {crc32_8 r1, r0, r0; ld r0, r0}");
465 static tilegx_bundle_bits jit_x1_ld(int rd, int ra)
466 {
467 extern tilegx_bundle_bits __unalign_jit_x1_ld;
468 return (GX_INSN_BSWAP(__unalign_jit_x1_ld) & GX_INSN_X1_MASK) |
469 create_Dest_X1(rd) | create_SrcA_X1(ra);
470 }
471
472 __JIT_CODE("__unalign_jit_x1_ld_add: {ld_add r1, r0, 0}");
473 static tilegx_bundle_bits jit_x1_ld_add(int rd, int ra, int imm8)
474 {
475 extern tilegx_bundle_bits __unalign_jit_x1_ld_add;
476 return (GX_INSN_BSWAP(__unalign_jit_x1_ld_add) &
477 (~create_Dest_X1(-1)) &
478 GX_INSN_X1_MASK) | create_Dest_X1(rd) |
479 create_SrcA_X1(ra) | create_Imm8_X1(imm8);
480 }
481
482 __JIT_CODE("__unalign_jit_x0_bfexts: {bfexts r0, r0, 0, 0}");
483 static tilegx_bundle_bits jit_x0_bfexts(int rd, int ra, int bfs, int bfe)
484 {
485 extern tilegx_bundle_bits __unalign_jit_x0_bfexts;
486 return (GX_INSN_BSWAP(__unalign_jit_x0_bfexts) &
487 GX_INSN_X0_MASK) |
488 create_Dest_X0(rd) | create_SrcA_X0(ra) |
489 create_BFStart_X0(bfs) | create_BFEnd_X0(bfe);
490 }
491
492 __JIT_CODE("__unalign_jit_x0_bfextu: {bfextu r0, r0, 0, 0}");
493 static tilegx_bundle_bits jit_x0_bfextu(int rd, int ra, int bfs, int bfe)
494 {
495 extern tilegx_bundle_bits __unalign_jit_x0_bfextu;
496 return (GX_INSN_BSWAP(__unalign_jit_x0_bfextu) &
497 GX_INSN_X0_MASK) |
498 create_Dest_X0(rd) | create_SrcA_X0(ra) |
499 create_BFStart_X0(bfs) | create_BFEnd_X0(bfe);
500 }
501
502 __JIT_CODE("__unalign_jit_x1_addi: {bfextu r1, r1, 0, 0; addi r0, r0, 0}");
503 static tilegx_bundle_bits jit_x1_addi(int rd, int ra, int imm8)
504 {
505 extern tilegx_bundle_bits __unalign_jit_x1_addi;
506 return (GX_INSN_BSWAP(__unalign_jit_x1_addi) & GX_INSN_X1_MASK) |
507 create_Dest_X1(rd) | create_SrcA_X1(ra) |
508 create_Imm8_X1(imm8);
509 }
510
511 __JIT_CODE("__unalign_jit_x0_shrui: {shrui r0, r0, 0; iret}");
512 static tilegx_bundle_bits jit_x0_shrui(int rd, int ra, int imm6)
513 {
514 extern tilegx_bundle_bits __unalign_jit_x0_shrui;
515 return (GX_INSN_BSWAP(__unalign_jit_x0_shrui) &
516 GX_INSN_X0_MASK) |
517 create_Dest_X0(rd) | create_SrcA_X0(ra) |
518 create_ShAmt_X0(imm6);
519 }
520
521 __JIT_CODE("__unalign_jit_x0_rotli: {rotli r0, r0, 0; iret}");
522 static tilegx_bundle_bits jit_x0_rotli(int rd, int ra, int imm6)
523 {
524 extern tilegx_bundle_bits __unalign_jit_x0_rotli;
525 return (GX_INSN_BSWAP(__unalign_jit_x0_rotli) &
526 GX_INSN_X0_MASK) |
527 create_Dest_X0(rd) | create_SrcA_X0(ra) |
528 create_ShAmt_X0(imm6);
529 }
530
531 __JIT_CODE("__unalign_jit_x1_bnezt: {bnezt r0, __unalign_jit_x1_bnezt}");
532 static tilegx_bundle_bits jit_x1_bnezt(int ra, int broff)
533 {
534 extern tilegx_bundle_bits __unalign_jit_x1_bnezt;
535 return (GX_INSN_BSWAP(__unalign_jit_x1_bnezt) &
536 GX_INSN_X1_MASK) |
537 create_SrcA_X1(ra) | create_BrOff_X1(broff);
538 }
539
540 #undef __JIT_CODE
541
542 /*
543 * This function generates unalign fixup JIT.
544 *
545 * We first find unalign load/store instruction's destination, source
546 * registers: ra, rb and rd. and 3 scratch registers by calling
547 * find_regs(...). 3 scratch clobbers should not alias with any register
548 * used in the fault bundle. Then analyze the fault bundle to determine
549 * if it's a load or store, operand width, branch or address increment etc.
550 * At last generated JIT is copied into JIT code area in user space.
551 */
552
553 static
554 void jit_bundle_gen(struct pt_regs *regs, tilegx_bundle_bits bundle,
555 int align_ctl)
556 {
557 struct thread_info *info = current_thread_info();
558 struct unaligned_jit_fragment frag;
559 struct unaligned_jit_fragment *jit_code_area;
560 tilegx_bundle_bits bundle_2 = 0;
561 /* If bundle_2_enable = false, bundle_2 is fnop/nop operation. */
562 bool bundle_2_enable = true;
563 uint64_t ra = -1, rb = -1, rd = -1, clob1 = -1, clob2 = -1, clob3 = -1;
564 /*
565 * Indicate if the unalign access
566 * instruction's registers hit with
567 * others in the same bundle.
568 */
569 bool alias = false;
570 bool load_n_store = true;
571 bool load_store_signed = false;
572 unsigned int load_store_size = 8;
573 bool y1_br = false; /* True, for a branch in same bundle at Y1.*/
574 int y1_br_reg = 0;
575 /* True for link operation. i.e. jalr or lnk at Y1 */
576 bool y1_lr = false;
577 int y1_lr_reg = 0;
578 bool x1_add = false;/* True, for load/store ADD instruction at X1*/
579 int x1_add_imm8 = 0;
580 bool unexpected = false;
581 int n = 0, k;
582
583 jit_code_area =
584 (struct unaligned_jit_fragment *)(info->unalign_jit_base);
585
586 memset((void *)&frag, 0, sizeof(frag));
587
588 /* 0: X mode, Otherwise: Y mode. */
589 if (bundle & TILEGX_BUNDLE_MODE_MASK) {
590 unsigned int mod, opcode;
591
592 if (get_Opcode_Y1(bundle) == RRR_1_OPCODE_Y1 &&
593 get_RRROpcodeExtension_Y1(bundle) ==
594 UNARY_RRR_1_OPCODE_Y1) {
595
596 opcode = get_UnaryOpcodeExtension_Y1(bundle);
597
598 /*
599 * Test "jalr", "jalrp", "jr", "jrp" instruction at Y1
600 * pipeline.
601 */
602 switch (opcode) {
603 case JALR_UNARY_OPCODE_Y1:
604 case JALRP_UNARY_OPCODE_Y1:
605 y1_lr = true;
606 y1_lr_reg = 55; /* Link register. */
607 /* FALLTHROUGH */
608 case JR_UNARY_OPCODE_Y1:
609 case JRP_UNARY_OPCODE_Y1:
610 y1_br = true;
611 y1_br_reg = get_SrcA_Y1(bundle);
612 break;
613 case LNK_UNARY_OPCODE_Y1:
614 /* "lnk" at Y1 pipeline. */
615 y1_lr = true;
616 y1_lr_reg = get_Dest_Y1(bundle);
617 break;
618 }
619 }
620
621 opcode = get_Opcode_Y2(bundle);
622 mod = get_Mode(bundle);
623
624 /*
625 * bundle_2 is bundle after making Y2 as a dummy operation
626 * - ld zero, sp
627 */
628 bundle_2 = (bundle & (~GX_INSN_Y2_MASK)) | jit_y2_dummy();
629
630 /* Make Y1 as fnop if Y1 is a branch or lnk operation. */
631 if (y1_br || y1_lr) {
632 bundle_2 &= ~(GX_INSN_Y1_MASK);
633 bundle_2 |= jit_y1_fnop();
634 }
635
636 if (is_y0_y1_nop(bundle_2))
637 bundle_2_enable = false;
638
639 if (mod == MODE_OPCODE_YC2) {
640 /* Store. */
641 load_n_store = false;
642 load_store_size = 1 << opcode;
643 load_store_signed = false;
644 find_regs(bundle, 0, &ra, &rb, &clob1, &clob2,
645 &clob3, &alias);
646 if (load_store_size > 8)
647 unexpected = true;
648 } else {
649 /* Load. */
650 load_n_store = true;
651 if (mod == MODE_OPCODE_YB2) {
652 switch (opcode) {
653 case LD_OPCODE_Y2:
654 load_store_signed = false;
655 load_store_size = 8;
656 break;
657 case LD4S_OPCODE_Y2:
658 load_store_signed = true;
659 load_store_size = 4;
660 break;
661 case LD4U_OPCODE_Y2:
662 load_store_signed = false;
663 load_store_size = 4;
664 break;
665 default:
666 unexpected = true;
667 }
668 } else if (mod == MODE_OPCODE_YA2) {
669 if (opcode == LD2S_OPCODE_Y2) {
670 load_store_signed = true;
671 load_store_size = 2;
672 } else if (opcode == LD2U_OPCODE_Y2) {
673 load_store_signed = false;
674 load_store_size = 2;
675 } else
676 unexpected = true;
677 } else
678 unexpected = true;
679 find_regs(bundle, &rd, &ra, &rb, &clob1, &clob2,
680 &clob3, &alias);
681 }
682 } else {
683 unsigned int opcode;
684
685 /* bundle_2 is bundle after making X1 as "fnop". */
686 bundle_2 = (bundle & (~GX_INSN_X1_MASK)) | jit_x1_fnop();
687
688 if (is_x0_x1_nop(bundle_2))
689 bundle_2_enable = false;
690
691 if (get_Opcode_X1(bundle) == RRR_0_OPCODE_X1) {
692 opcode = get_UnaryOpcodeExtension_X1(bundle);
693
694 if (get_RRROpcodeExtension_X1(bundle) ==
695 UNARY_RRR_0_OPCODE_X1) {
696 load_n_store = true;
697 find_regs(bundle, &rd, &ra, &rb, &clob1,
698 &clob2, &clob3, &alias);
699
700 switch (opcode) {
701 case LD_UNARY_OPCODE_X1:
702 load_store_signed = false;
703 load_store_size = 8;
704 break;
705 case LD4S_UNARY_OPCODE_X1:
706 load_store_signed = true;
707 /* FALLTHROUGH */
708 case LD4U_UNARY_OPCODE_X1:
709 load_store_size = 4;
710 break;
711
712 case LD2S_UNARY_OPCODE_X1:
713 load_store_signed = true;
714 /* FALLTHROUGH */
715 case LD2U_UNARY_OPCODE_X1:
716 load_store_size = 2;
717 break;
718 default:
719 unexpected = true;
720 }
721 } else {
722 load_n_store = false;
723 load_store_signed = false;
724 find_regs(bundle, 0, &ra, &rb,
725 &clob1, &clob2, &clob3,
726 &alias);
727
728 opcode = get_RRROpcodeExtension_X1(bundle);
729 switch (opcode) {
730 case ST_RRR_0_OPCODE_X1:
731 load_store_size = 8;
732 break;
733 case ST4_RRR_0_OPCODE_X1:
734 load_store_size = 4;
735 break;
736 case ST2_RRR_0_OPCODE_X1:
737 load_store_size = 2;
738 break;
739 default:
740 unexpected = true;
741 }
742 }
743 } else if (get_Opcode_X1(bundle) == IMM8_OPCODE_X1) {
744 load_n_store = true;
745 opcode = get_Imm8OpcodeExtension_X1(bundle);
746 switch (opcode) {
747 case LD_ADD_IMM8_OPCODE_X1:
748 load_store_size = 8;
749 break;
750
751 case LD4S_ADD_IMM8_OPCODE_X1:
752 load_store_signed = true;
753 /* FALLTHROUGH */
754 case LD4U_ADD_IMM8_OPCODE_X1:
755 load_store_size = 4;
756 break;
757
758 case LD2S_ADD_IMM8_OPCODE_X1:
759 load_store_signed = true;
760 /* FALLTHROUGH */
761 case LD2U_ADD_IMM8_OPCODE_X1:
762 load_store_size = 2;
763 break;
764
765 case ST_ADD_IMM8_OPCODE_X1:
766 load_n_store = false;
767 load_store_size = 8;
768 break;
769 case ST4_ADD_IMM8_OPCODE_X1:
770 load_n_store = false;
771 load_store_size = 4;
772 break;
773 case ST2_ADD_IMM8_OPCODE_X1:
774 load_n_store = false;
775 load_store_size = 2;
776 break;
777 default:
778 unexpected = true;
779 }
780
781 if (!unexpected) {
782 x1_add = true;
783 if (load_n_store)
784 x1_add_imm8 = get_Imm8_X1(bundle);
785 else
786 x1_add_imm8 = get_Dest_Imm8_X1(bundle);
787 }
788
789 find_regs(bundle, load_n_store ? (&rd) : NULL,
790 &ra, &rb, &clob1, &clob2, &clob3, &alias);
791 } else
792 unexpected = true;
793 }
794
795 /*
796 * Some sanity check for register numbers extracted from fault bundle.
797 */
798 if (check_regs(rd, ra, rb, clob1, clob2, clob3) == true)
799 unexpected = true;
800
801 /* Give warning if register ra has an aligned address. */
802 if (!unexpected)
803 WARN_ON(!((load_store_size - 1) & (regs->regs[ra])));
804
805
806 /*
807 * Fault came from kernel space, here we only need take care of
808 * unaligned "get_user/put_user" macros defined in "uaccess.h".
809 * Basically, we will handle bundle like this:
810 * {ld/2u/4s rd, ra; movei rx, 0} or {st/2/4 ra, rb; movei rx, 0}
811 * (Refer to file "arch/tile/include/asm/uaccess.h" for details).
812 * For either load or store, byte-wise operation is performed by calling
813 * get_user() or put_user(). If the macro returns non-zero value,
814 * set the value to rx, otherwise set zero to rx. Finally make pc point
815 * to next bundle and return.
816 */
817
818 if (EX1_PL(regs->ex1) != USER_PL) {
819
820 unsigned long rx = 0;
821 unsigned long x = 0, ret = 0;
822
823 if (y1_br || y1_lr || x1_add ||
824 (load_store_signed !=
825 (load_n_store && load_store_size == 4))) {
826 /* No branch, link, wrong sign-ext or load/store add. */
827 unexpected = true;
828 } else if (!unexpected) {
829 if (bundle & TILEGX_BUNDLE_MODE_MASK) {
830 /*
831 * Fault bundle is Y mode.
832 * Check if the Y1 and Y0 is the form of
833 * { movei rx, 0; nop/fnop }, if yes,
834 * find the rx.
835 */
836
837 if ((get_Opcode_Y1(bundle) == ADDI_OPCODE_Y1)
838 && (get_SrcA_Y1(bundle) == TREG_ZERO) &&
839 (get_Imm8_Y1(bundle) == 0) &&
840 is_bundle_y0_nop(bundle)) {
841 rx = get_Dest_Y1(bundle);
842 } else if ((get_Opcode_Y0(bundle) ==
843 ADDI_OPCODE_Y0) &&
844 (get_SrcA_Y0(bundle) == TREG_ZERO) &&
845 (get_Imm8_Y0(bundle) == 0) &&
846 is_bundle_y1_nop(bundle)) {
847 rx = get_Dest_Y0(bundle);
848 } else {
849 unexpected = true;
850 }
851 } else {
852 /*
853 * Fault bundle is X mode.
854 * Check if the X0 is 'movei rx, 0',
855 * if yes, find the rx.
856 */
857
858 if ((get_Opcode_X0(bundle) == IMM8_OPCODE_X0)
859 && (get_Imm8OpcodeExtension_X0(bundle) ==
860 ADDI_IMM8_OPCODE_X0) &&
861 (get_SrcA_X0(bundle) == TREG_ZERO) &&
862 (get_Imm8_X0(bundle) == 0)) {
863 rx = get_Dest_X0(bundle);
864 } else {
865 unexpected = true;
866 }
867 }
868
869 /* rx should be less than 56. */
870 if (!unexpected && (rx >= 56))
871 unexpected = true;
872 }
873
874 if (!search_exception_tables(regs->pc)) {
875 /* No fixup in the exception tables for the pc. */
876 unexpected = true;
877 }
878
879 if (unexpected) {
880 /* Unexpected unalign kernel fault. */
881 struct task_struct *tsk = validate_current();
882
883 bust_spinlocks(1);
884
885 show_regs(regs);
886
887 if (unlikely(tsk->pid < 2)) {
888 panic("Kernel unalign fault running %s!",
889 tsk->pid ? "init" : "the idle task");
890 }
891 #ifdef SUPPORT_DIE
892 die("Oops", regs);
893 #endif
894 bust_spinlocks(1);
895
896 do_group_exit(SIGKILL);
897
898 } else {
899 unsigned long i, b = 0;
900 unsigned char *ptr =
901 (unsigned char *)regs->regs[ra];
902 if (load_n_store) {
903 /* handle get_user(x, ptr) */
904 for (i = 0; i < load_store_size; i++) {
905 ret = get_user(b, ptr++);
906 if (!ret) {
907 /* Success! update x. */
908 #ifdef __LITTLE_ENDIAN
909 x |= (b << (8 * i));
910 #else
911 x <<= 8;
912 x |= b;
913 #endif /* __LITTLE_ENDIAN */
914 } else {
915 x = 0;
916 break;
917 }
918 }
919
920 /* Sign-extend 4-byte loads. */
921 if (load_store_size == 4)
922 x = (long)(int)x;
923
924 /* Set register rd. */
925 regs->regs[rd] = x;
926
927 /* Set register rx. */
928 regs->regs[rx] = ret;
929
930 /* Bump pc. */
931 regs->pc += 8;
932
933 } else {
934 /* Handle put_user(x, ptr) */
935 x = regs->regs[rb];
936 #ifdef __LITTLE_ENDIAN
937 b = x;
938 #else
939 /*
940 * Swap x in order to store x from low
941 * to high memory same as the
942 * little-endian case.
943 */
944 switch (load_store_size) {
945 case 8:
946 b = swab64(x);
947 break;
948 case 4:
949 b = swab32(x);
950 break;
951 case 2:
952 b = swab16(x);
953 break;
954 }
955 #endif /* __LITTLE_ENDIAN */
956 for (i = 0; i < load_store_size; i++) {
957 ret = put_user(b, ptr++);
958 if (ret)
959 break;
960 /* Success! shift 1 byte. */
961 b >>= 8;
962 }
963 /* Set register rx. */
964 regs->regs[rx] = ret;
965
966 /* Bump pc. */
967 regs->pc += 8;
968 }
969 }
970
971 unaligned_fixup_count++;
972
973 if (unaligned_printk) {
974 pr_info("%s/%d - Unalign fixup for kernel access to userspace %lx\n",
975 current->comm, current->pid, regs->regs[ra]);
976 }
977
978 /* Done! Return to the exception handler. */
979 return;
980 }
981
982 if ((align_ctl == 0) || unexpected) {
983 siginfo_t info = {
984 .si_signo = SIGBUS,
985 .si_code = BUS_ADRALN,
986 .si_addr = (unsigned char __user *)0
987 };
988 if (unaligned_printk)
989 pr_info("Unalign bundle: unexp @%llx, %llx\n",
990 (unsigned long long)regs->pc,
991 (unsigned long long)bundle);
992
993 if (ra < 56) {
994 unsigned long uaa = (unsigned long)regs->regs[ra];
995 /* Set bus Address. */
996 info.si_addr = (unsigned char __user *)uaa;
997 }
998
999 unaligned_fixup_count++;
1000
1001 trace_unhandled_signal("unaligned fixup trap", regs,
1002 (unsigned long)info.si_addr, SIGBUS);
1003 force_sig_info(info.si_signo, &info, current);
1004 return;
1005 }
1006
1007 #ifdef __LITTLE_ENDIAN
1008 #define UA_FIXUP_ADDR_DELTA 1
1009 #define UA_FIXUP_BFEXT_START(_B_) 0
1010 #define UA_FIXUP_BFEXT_END(_B_) (8 * (_B_) - 1)
1011 #else /* __BIG_ENDIAN */
1012 #define UA_FIXUP_ADDR_DELTA -1
1013 #define UA_FIXUP_BFEXT_START(_B_) (64 - 8 * (_B_))
1014 #define UA_FIXUP_BFEXT_END(_B_) 63
1015 #endif /* __LITTLE_ENDIAN */
1016
1017
1018
1019 if ((ra != rb) && (rd != TREG_SP) && !alias &&
1020 !y1_br && !y1_lr && !x1_add) {
1021 /*
1022 * Simple case: ra != rb and no register alias found,
1023 * and no branch or link. This will be the majority.
1024 * We can do a little better for simplae case than the
1025 * generic scheme below.
1026 */
1027 if (!load_n_store) {
1028 /*
1029 * Simple store: ra != rb, no need for scratch register.
1030 * Just store and rotate to right bytewise.
1031 */
1032 #ifdef __BIG_ENDIAN
1033 frag.insn[n++] =
1034 jit_x0_addi(ra, ra, load_store_size - 1) |
1035 jit_x1_fnop();
1036 #endif /* __BIG_ENDIAN */
1037 for (k = 0; k < load_store_size; k++) {
1038 /* Store a byte. */
1039 frag.insn[n++] =
1040 jit_x0_rotli(rb, rb, 56) |
1041 jit_x1_st1_add(ra, rb,
1042 UA_FIXUP_ADDR_DELTA);
1043 }
1044 #ifdef __BIG_ENDIAN
1045 frag.insn[n] = jit_x1_addi(ra, ra, 1);
1046 #else
1047 frag.insn[n] = jit_x1_addi(ra, ra,
1048 -1 * load_store_size);
1049 #endif /* __LITTLE_ENDIAN */
1050
1051 if (load_store_size == 8) {
1052 frag.insn[n] |= jit_x0_fnop();
1053 } else if (load_store_size == 4) {
1054 frag.insn[n] |= jit_x0_rotli(rb, rb, 32);
1055 } else { /* = 2 */
1056 frag.insn[n] |= jit_x0_rotli(rb, rb, 16);
1057 }
1058 n++;
1059 if (bundle_2_enable)
1060 frag.insn[n++] = bundle_2;
1061 frag.insn[n++] = jit_x0_fnop() | jit_x1_iret();
1062 } else {
1063 if (rd == ra) {
1064 /* Use two clobber registers: clob1/2. */
1065 frag.insn[n++] =
1066 jit_x0_addi(TREG_SP, TREG_SP, -16) |
1067 jit_x1_fnop();
1068 frag.insn[n++] =
1069 jit_x0_addi(clob1, ra, 7) |
1070 jit_x1_st_add(TREG_SP, clob1, -8);
1071 frag.insn[n++] =
1072 jit_x0_addi(clob2, ra, 0) |
1073 jit_x1_st(TREG_SP, clob2);
1074 frag.insn[n++] =
1075 jit_x0_fnop() |
1076 jit_x1_ldna(rd, ra);
1077 frag.insn[n++] =
1078 jit_x0_fnop() |
1079 jit_x1_ldna(clob1, clob1);
1080 /*
1081 * Note: we must make sure that rd must not
1082 * be sp. Recover clob1/2 from stack.
1083 */
1084 frag.insn[n++] =
1085 jit_x0_dblalign(rd, clob1, clob2) |
1086 jit_x1_ld_add(clob2, TREG_SP, 8);
1087 frag.insn[n++] =
1088 jit_x0_fnop() |
1089 jit_x1_ld_add(clob1, TREG_SP, 16);
1090 } else {
1091 /* Use one clobber register: clob1 only. */
1092 frag.insn[n++] =
1093 jit_x0_addi(TREG_SP, TREG_SP, -16) |
1094 jit_x1_fnop();
1095 frag.insn[n++] =
1096 jit_x0_addi(clob1, ra, 7) |
1097 jit_x1_st(TREG_SP, clob1);
1098 frag.insn[n++] =
1099 jit_x0_fnop() |
1100 jit_x1_ldna(rd, ra);
1101 frag.insn[n++] =
1102 jit_x0_fnop() |
1103 jit_x1_ldna(clob1, clob1);
1104 /*
1105 * Note: we must make sure that rd must not
1106 * be sp. Recover clob1 from stack.
1107 */
1108 frag.insn[n++] =
1109 jit_x0_dblalign(rd, clob1, ra) |
1110 jit_x1_ld_add(clob1, TREG_SP, 16);
1111 }
1112
1113 if (bundle_2_enable)
1114 frag.insn[n++] = bundle_2;
1115 /*
1116 * For non 8-byte load, extract corresponding bytes and
1117 * signed extension.
1118 */
1119 if (load_store_size == 4) {
1120 if (load_store_signed)
1121 frag.insn[n++] =
1122 jit_x0_bfexts(
1123 rd, rd,
1124 UA_FIXUP_BFEXT_START(4),
1125 UA_FIXUP_BFEXT_END(4)) |
1126 jit_x1_fnop();
1127 else
1128 frag.insn[n++] =
1129 jit_x0_bfextu(
1130 rd, rd,
1131 UA_FIXUP_BFEXT_START(4),
1132 UA_FIXUP_BFEXT_END(4)) |
1133 jit_x1_fnop();
1134 } else if (load_store_size == 2) {
1135 if (load_store_signed)
1136 frag.insn[n++] =
1137 jit_x0_bfexts(
1138 rd, rd,
1139 UA_FIXUP_BFEXT_START(2),
1140 UA_FIXUP_BFEXT_END(2)) |
1141 jit_x1_fnop();
1142 else
1143 frag.insn[n++] =
1144 jit_x0_bfextu(
1145 rd, rd,
1146 UA_FIXUP_BFEXT_START(2),
1147 UA_FIXUP_BFEXT_END(2)) |
1148 jit_x1_fnop();
1149 }
1150
1151 frag.insn[n++] =
1152 jit_x0_fnop() |
1153 jit_x1_iret();
1154 }
1155 } else if (!load_n_store) {
1156
1157 /*
1158 * Generic memory store cases: use 3 clobber registers.
1159 *
1160 * Alloc space for saveing clob2,1,3 on user's stack.
1161 * register clob3 points to where clob2 saved, followed by
1162 * clob1 and 3 from high to low memory.
1163 */
1164 frag.insn[n++] =
1165 jit_x0_addi(TREG_SP, TREG_SP, -32) |
1166 jit_x1_fnop();
1167 frag.insn[n++] =
1168 jit_x0_addi(clob3, TREG_SP, 16) |
1169 jit_x1_st_add(TREG_SP, clob3, 8);
1170 #ifdef __LITTLE_ENDIAN
1171 frag.insn[n++] =
1172 jit_x0_addi(clob1, ra, 0) |
1173 jit_x1_st_add(TREG_SP, clob1, 8);
1174 #else
1175 frag.insn[n++] =
1176 jit_x0_addi(clob1, ra, load_store_size - 1) |
1177 jit_x1_st_add(TREG_SP, clob1, 8);
1178 #endif
1179 if (load_store_size == 8) {
1180 /*
1181 * We save one byte a time, not for fast, but compact
1182 * code. After each store, data source register shift
1183 * right one byte. unchanged after 8 stores.
1184 */
1185 frag.insn[n++] =
1186 jit_x0_addi(clob2, TREG_ZERO, 7) |
1187 jit_x1_st_add(TREG_SP, clob2, 16);
1188 frag.insn[n++] =
1189 jit_x0_rotli(rb, rb, 56) |
1190 jit_x1_st1_add(clob1, rb, UA_FIXUP_ADDR_DELTA);
1191 frag.insn[n++] =
1192 jit_x0_addi(clob2, clob2, -1) |
1193 jit_x1_bnezt(clob2, -1);
1194 frag.insn[n++] =
1195 jit_x0_fnop() |
1196 jit_x1_addi(clob2, y1_br_reg, 0);
1197 } else if (load_store_size == 4) {
1198 frag.insn[n++] =
1199 jit_x0_addi(clob2, TREG_ZERO, 3) |
1200 jit_x1_st_add(TREG_SP, clob2, 16);
1201 frag.insn[n++] =
1202 jit_x0_rotli(rb, rb, 56) |
1203 jit_x1_st1_add(clob1, rb, UA_FIXUP_ADDR_DELTA);
1204 frag.insn[n++] =
1205 jit_x0_addi(clob2, clob2, -1) |
1206 jit_x1_bnezt(clob2, -1);
1207 /*
1208 * same as 8-byte case, but need shift another 4
1209 * byte to recover rb for 4-byte store.
1210 */
1211 frag.insn[n++] = jit_x0_rotli(rb, rb, 32) |
1212 jit_x1_addi(clob2, y1_br_reg, 0);
1213 } else { /* =2 */
1214 frag.insn[n++] =
1215 jit_x0_addi(clob2, rb, 0) |
1216 jit_x1_st_add(TREG_SP, clob2, 16);
1217 for (k = 0; k < 2; k++) {
1218 frag.insn[n++] =
1219 jit_x0_shrui(rb, rb, 8) |
1220 jit_x1_st1_add(clob1, rb,
1221 UA_FIXUP_ADDR_DELTA);
1222 }
1223 frag.insn[n++] =
1224 jit_x0_addi(rb, clob2, 0) |
1225 jit_x1_addi(clob2, y1_br_reg, 0);
1226 }
1227
1228 if (bundle_2_enable)
1229 frag.insn[n++] = bundle_2;
1230
1231 if (y1_lr) {
1232 frag.insn[n++] =
1233 jit_x0_fnop() |
1234 jit_x1_mfspr(y1_lr_reg,
1235 SPR_EX_CONTEXT_0_0);
1236 }
1237 if (y1_br) {
1238 frag.insn[n++] =
1239 jit_x0_fnop() |
1240 jit_x1_mtspr(SPR_EX_CONTEXT_0_0,
1241 clob2);
1242 }
1243 if (x1_add) {
1244 frag.insn[n++] =
1245 jit_x0_addi(ra, ra, x1_add_imm8) |
1246 jit_x1_ld_add(clob2, clob3, -8);
1247 } else {
1248 frag.insn[n++] =
1249 jit_x0_fnop() |
1250 jit_x1_ld_add(clob2, clob3, -8);
1251 }
1252 frag.insn[n++] =
1253 jit_x0_fnop() |
1254 jit_x1_ld_add(clob1, clob3, -8);
1255 frag.insn[n++] = jit_x0_fnop() | jit_x1_ld(clob3, clob3);
1256 frag.insn[n++] = jit_x0_fnop() | jit_x1_iret();
1257
1258 } else {
1259 /*
1260 * Generic memory load cases.
1261 *
1262 * Alloc space for saveing clob1,2,3 on user's stack.
1263 * register clob3 points to where clob1 saved, followed
1264 * by clob2 and 3 from high to low memory.
1265 */
1266
1267 frag.insn[n++] =
1268 jit_x0_addi(TREG_SP, TREG_SP, -32) |
1269 jit_x1_fnop();
1270 frag.insn[n++] =
1271 jit_x0_addi(clob3, TREG_SP, 16) |
1272 jit_x1_st_add(TREG_SP, clob3, 8);
1273 frag.insn[n++] =
1274 jit_x0_addi(clob2, ra, 0) |
1275 jit_x1_st_add(TREG_SP, clob2, 8);
1276
1277 if (y1_br) {
1278 frag.insn[n++] =
1279 jit_x0_addi(clob1, y1_br_reg, 0) |
1280 jit_x1_st_add(TREG_SP, clob1, 16);
1281 } else {
1282 frag.insn[n++] =
1283 jit_x0_fnop() |
1284 jit_x1_st_add(TREG_SP, clob1, 16);
1285 }
1286
1287 if (bundle_2_enable)
1288 frag.insn[n++] = bundle_2;
1289
1290 if (y1_lr) {
1291 frag.insn[n++] =
1292 jit_x0_fnop() |
1293 jit_x1_mfspr(y1_lr_reg,
1294 SPR_EX_CONTEXT_0_0);
1295 }
1296
1297 if (y1_br) {
1298 frag.insn[n++] =
1299 jit_x0_fnop() |
1300 jit_x1_mtspr(SPR_EX_CONTEXT_0_0,
1301 clob1);
1302 }
1303
1304 frag.insn[n++] =
1305 jit_x0_addi(clob1, clob2, 7) |
1306 jit_x1_ldna(rd, clob2);
1307 frag.insn[n++] =
1308 jit_x0_fnop() |
1309 jit_x1_ldna(clob1, clob1);
1310 frag.insn[n++] =
1311 jit_x0_dblalign(rd, clob1, clob2) |
1312 jit_x1_ld_add(clob1, clob3, -8);
1313 if (x1_add) {
1314 frag.insn[n++] =
1315 jit_x0_addi(ra, ra, x1_add_imm8) |
1316 jit_x1_ld_add(clob2, clob3, -8);
1317 } else {
1318 frag.insn[n++] =
1319 jit_x0_fnop() |
1320 jit_x1_ld_add(clob2, clob3, -8);
1321 }
1322
1323 frag.insn[n++] =
1324 jit_x0_fnop() |
1325 jit_x1_ld(clob3, clob3);
1326
1327 if (load_store_size == 4) {
1328 if (load_store_signed)
1329 frag.insn[n++] =
1330 jit_x0_bfexts(
1331 rd, rd,
1332 UA_FIXUP_BFEXT_START(4),
1333 UA_FIXUP_BFEXT_END(4)) |
1334 jit_x1_fnop();
1335 else
1336 frag.insn[n++] =
1337 jit_x0_bfextu(
1338 rd, rd,
1339 UA_FIXUP_BFEXT_START(4),
1340 UA_FIXUP_BFEXT_END(4)) |
1341 jit_x1_fnop();
1342 } else if (load_store_size == 2) {
1343 if (load_store_signed)
1344 frag.insn[n++] =
1345 jit_x0_bfexts(
1346 rd, rd,
1347 UA_FIXUP_BFEXT_START(2),
1348 UA_FIXUP_BFEXT_END(2)) |
1349 jit_x1_fnop();
1350 else
1351 frag.insn[n++] =
1352 jit_x0_bfextu(
1353 rd, rd,
1354 UA_FIXUP_BFEXT_START(2),
1355 UA_FIXUP_BFEXT_END(2)) |
1356 jit_x1_fnop();
1357 }
1358
1359 frag.insn[n++] = jit_x0_fnop() | jit_x1_iret();
1360 }
1361
1362 /* Max JIT bundle count is 14. */
1363 WARN_ON(n > 14);
1364
1365 if (!unexpected) {
1366 int status = 0;
1367 int idx = (regs->pc >> 3) &
1368 ((1ULL << (PAGE_SHIFT - UNALIGN_JIT_SHIFT)) - 1);
1369
1370 frag.pc = regs->pc;
1371 frag.bundle = bundle;
1372
1373 if (unaligned_printk) {
1374 pr_info("%s/%d, Unalign fixup: pc=%lx bundle=%lx %d %d %d %d %d %d %d %d\n",
1375 current->comm, current->pid,
1376 (unsigned long)frag.pc,
1377 (unsigned long)frag.bundle,
1378 (int)alias, (int)rd, (int)ra,
1379 (int)rb, (int)bundle_2_enable,
1380 (int)y1_lr, (int)y1_br, (int)x1_add);
1381
1382 for (k = 0; k < n; k += 2)
1383 pr_info("[%d] %016llx %016llx\n",
1384 k, (unsigned long long)frag.insn[k],
1385 (unsigned long long)frag.insn[k+1]);
1386 }
1387
1388 /* Swap bundle byte order for big endian sys. */
1389 #ifdef __BIG_ENDIAN
1390 frag.bundle = GX_INSN_BSWAP(frag.bundle);
1391 for (k = 0; k < n; k++)
1392 frag.insn[k] = GX_INSN_BSWAP(frag.insn[k]);
1393 #endif /* __BIG_ENDIAN */
1394
1395 status = copy_to_user((void __user *)&jit_code_area[idx],
1396 &frag, sizeof(frag));
1397 if (status) {
1398 /* Fail to copy JIT into user land. send SIGSEGV. */
1399 siginfo_t info = {
1400 .si_signo = SIGSEGV,
1401 .si_code = SEGV_MAPERR,
1402 .si_addr = (void __user *)&jit_code_area[idx]
1403 };
1404
1405 pr_warn("Unalign fixup: pid=%d %s jit_code_area=%llx\n",
1406 current->pid, current->comm,
1407 (unsigned long long)&jit_code_area[idx]);
1408
1409 trace_unhandled_signal("segfault in unalign fixup",
1410 regs,
1411 (unsigned long)info.si_addr,
1412 SIGSEGV);
1413 force_sig_info(info.si_signo, &info, current);
1414 return;
1415 }
1416
1417
1418 /* Do a cheaper increment, not accurate. */
1419 unaligned_fixup_count++;
1420 __flush_icache_range((unsigned long)&jit_code_area[idx],
1421 (unsigned long)&jit_code_area[idx] +
1422 sizeof(frag));
1423
1424 /* Setup SPR_EX_CONTEXT_0_0/1 for returning to user program.*/
1425 __insn_mtspr(SPR_EX_CONTEXT_0_0, regs->pc + 8);
1426 __insn_mtspr(SPR_EX_CONTEXT_0_1, PL_ICS_EX1(USER_PL, 0));
1427
1428 /* Modify pc at the start of new JIT. */
1429 regs->pc = (unsigned long)&jit_code_area[idx].insn[0];
1430 /* Set ICS in SPR_EX_CONTEXT_K_1. */
1431 regs->ex1 = PL_ICS_EX1(USER_PL, 1);
1432 }
1433 }
1434
1435
1436 /*
1437 * C function to generate unalign data JIT. Called from unalign data
1438 * interrupt handler.
1439 *
1440 * First check if unalign fix is disabled or exception did not not come from
1441 * user space or sp register points to unalign address, if true, generate a
1442 * SIGBUS. Then map a page into user space as JIT area if it is not mapped
1443 * yet. Genenerate JIT code by calling jit_bundle_gen(). After that return
1444 * back to exception handler.
1445 *
1446 * The exception handler will "iret" to new generated JIT code after
1447 * restoring caller saved registers. In theory, the JIT code will perform
1448 * another "iret" to resume user's program.
1449 */
1450
1451 void do_unaligned(struct pt_regs *regs, int vecnum)
1452 {
1453 tilegx_bundle_bits __user *pc;
1454 tilegx_bundle_bits bundle;
1455 struct thread_info *info = current_thread_info();
1456 int align_ctl;
1457
1458 /* Checks the per-process unaligned JIT flags */
1459 align_ctl = unaligned_fixup;
1460 switch (task_thread_info(current)->align_ctl) {
1461 case PR_UNALIGN_NOPRINT:
1462 align_ctl = 1;
1463 break;
1464 case PR_UNALIGN_SIGBUS:
1465 align_ctl = 0;
1466 break;
1467 }
1468
1469 /* Enable iterrupt in order to access user land. */
1470 local_irq_enable();
1471
1472 /*
1473 * The fault came from kernel space. Two choices:
1474 * (a) unaligned_fixup < 1, we will first call get/put_user fixup
1475 * to return -EFAULT. If no fixup, simply panic the kernel.
1476 * (b) unaligned_fixup >=1, we will try to fix the unaligned access
1477 * if it was triggered by get_user/put_user() macros. Panic the
1478 * kernel if it is not fixable.
1479 */
1480
1481 if (EX1_PL(regs->ex1) != USER_PL) {
1482
1483 if (align_ctl < 1) {
1484 unaligned_fixup_count++;
1485 /* If exception came from kernel, try fix it up. */
1486 if (fixup_exception(regs)) {
1487 if (unaligned_printk)
1488 pr_info("Unalign fixup: %d %llx @%llx\n",
1489 (int)unaligned_fixup,
1490 (unsigned long long)regs->ex1,
1491 (unsigned long long)regs->pc);
1492 } else {
1493 /* Not fixable. Go panic. */
1494 panic("Unalign exception in Kernel. pc=%lx",
1495 regs->pc);
1496 }
1497 } else {
1498 /*
1499 * Try to fix the exception. If we can't, panic the
1500 * kernel.
1501 */
1502 bundle = GX_INSN_BSWAP(
1503 *((tilegx_bundle_bits *)(regs->pc)));
1504 jit_bundle_gen(regs, bundle, align_ctl);
1505 }
1506 return;
1507 }
1508
1509 /*
1510 * Fault came from user with ICS or stack is not aligned.
1511 * If so, we will trigger SIGBUS.
1512 */
1513 if ((regs->sp & 0x7) || (regs->ex1) || (align_ctl < 0)) {
1514 siginfo_t info = {
1515 .si_signo = SIGBUS,
1516 .si_code = BUS_ADRALN,
1517 .si_addr = (unsigned char __user *)0
1518 };
1519
1520 if (unaligned_printk)
1521 pr_info("Unalign fixup: %d %llx @%llx\n",
1522 (int)unaligned_fixup,
1523 (unsigned long long)regs->ex1,
1524 (unsigned long long)regs->pc);
1525
1526 unaligned_fixup_count++;
1527
1528 trace_unhandled_signal("unaligned fixup trap", regs, 0, SIGBUS);
1529 force_sig_info(info.si_signo, &info, current);
1530 return;
1531 }
1532
1533
1534 /* Read the bundle caused the exception! */
1535 pc = (tilegx_bundle_bits __user *)(regs->pc);
1536 if (get_user(bundle, pc) != 0) {
1537 /* Probably never be here since pc is valid user address.*/
1538 siginfo_t info = {
1539 .si_signo = SIGSEGV,
1540 .si_code = SEGV_MAPERR,
1541 .si_addr = (void __user *)pc
1542 };
1543 pr_err("Couldn't read instruction at %p trying to step\n", pc);
1544 trace_unhandled_signal("segfault in unalign fixup", regs,
1545 (unsigned long)info.si_addr, SIGSEGV);
1546 force_sig_info(info.si_signo, &info, current);
1547 return;
1548 }
1549
1550 if (!info->unalign_jit_base) {
1551 void __user *user_page;
1552
1553 /*
1554 * Allocate a page in userland.
1555 * For 64-bit processes we try to place the mapping far
1556 * from anything else that might be going on (specifically
1557 * 64 GB below the top of the user address space). If it
1558 * happens not to be possible to put it there, it's OK;
1559 * the kernel will choose another location and we'll
1560 * remember it for later.
1561 */
1562 if (is_compat_task())
1563 user_page = NULL;
1564 else
1565 user_page = (void __user *)(TASK_SIZE - (1UL << 36)) +
1566 (current->pid << PAGE_SHIFT);
1567
1568 user_page = (void __user *) vm_mmap(NULL,
1569 (unsigned long)user_page,
1570 PAGE_SIZE,
1571 PROT_EXEC | PROT_READ |
1572 PROT_WRITE,
1573 #ifdef CONFIG_HOMECACHE
1574 MAP_CACHE_HOME_TASK |
1575 #endif
1576 MAP_PRIVATE |
1577 MAP_ANONYMOUS,
1578 0);
1579
1580 if (IS_ERR((void __force *)user_page)) {
1581 pr_err("Out of kernel pages trying do_mmap\n");
1582 return;
1583 }
1584
1585 /* Save the address in the thread_info struct */
1586 info->unalign_jit_base = user_page;
1587 if (unaligned_printk)
1588 pr_info("Unalign bundle: %d:%d, allocate page @%llx\n",
1589 raw_smp_processor_id(), current->pid,
1590 (unsigned long long)user_page);
1591 }
1592
1593 /* Generate unalign JIT */
1594 jit_bundle_gen(regs, GX_INSN_BSWAP(bundle), align_ctl);
1595 }
1596
1597 #endif /* __tilegx__ */
Linux with added FPC-III support