]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Architecture-specific unaligned trap handling. | |
3 | * | |
4 | * Copyright (C) 1999-2002, 2004 Hewlett-Packard Co | |
5 | * Stephane Eranian <eranian@hpl.hp.com> | |
6 | * David Mosberger-Tang <davidm@hpl.hp.com> | |
7 | * | |
8 | * 2002/12/09 Fix rotating register handling (off-by-1 error, missing fr-rotation). Fix | |
9 | * get_rse_reg() to not leak kernel bits to user-level (reading an out-of-frame | |
10 | * stacked register returns an undefined value; it does NOT trigger a | |
11 | * "rsvd register fault"). | |
12 | * 2001/10/11 Fix unaligned access to rotating registers in s/w pipelined loops. | |
13 | * 2001/08/13 Correct size of extended floats (float_fsz) from 16 to 10 bytes. | |
14 | * 2001/01/17 Add support emulation of unaligned kernel accesses. | |
15 | */ | |
5cf1f7ce | 16 | #include <linux/jiffies.h> |
1da177e4 LT |
17 | #include <linux/kernel.h> |
18 | #include <linux/sched.h> | |
1da177e4 | 19 | #include <linux/tty.h> |
7683a3f9 | 20 | #include <linux/ratelimit.h> |
82ed1ac9 | 21 | #include <linux/uaccess.h> |
1da177e4 LT |
22 | |
23 | #include <asm/intrinsics.h> | |
24 | #include <asm/processor.h> | |
25 | #include <asm/rse.h> | |
82ed1ac9 | 26 | #include <asm/exception.h> |
1da177e4 LT |
27 | #include <asm/unaligned.h> |
28 | ||
620de2f5 | 29 | extern int die_if_kernel(char *str, struct pt_regs *regs, long err); |
1da177e4 LT |
30 | |
31 | #undef DEBUG_UNALIGNED_TRAP | |
32 | ||
33 | #ifdef DEBUG_UNALIGNED_TRAP | |
d4ed8084 | 34 | # define DPRINT(a...) do { printk("%s %u: ", __func__, __LINE__); printk (a); } while (0) |
1da177e4 LT |
35 | # define DDUMP(str,vp,len) dump(str, vp, len) |
36 | ||
37 | static void | |
38 | dump (const char *str, void *vp, size_t len) | |
39 | { | |
40 | unsigned char *cp = vp; | |
41 | int i; | |
42 | ||
43 | printk("%s", str); | |
44 | for (i = 0; i < len; ++i) | |
45 | printk (" %02x", *cp++); | |
46 | printk("\n"); | |
47 | } | |
48 | #else | |
49 | # define DPRINT(a...) | |
50 | # define DDUMP(str,vp,len) | |
51 | #endif | |
52 | ||
53 | #define IA64_FIRST_STACKED_GR 32 | |
54 | #define IA64_FIRST_ROTATING_FR 32 | |
55 | #define SIGN_EXT9 0xffffffffffffff00ul | |
56 | ||
d2b176ed JS |
57 | /* |
58 | * sysctl settable hook which tells the kernel whether to honor the | |
59 | * IA64_THREAD_UAC_NOPRINT prctl. Because this is user settable, we want | |
60 | * to allow the super user to enable/disable this for security reasons | |
61 | * (i.e. don't allow attacker to fill up logs with unaligned accesses). | |
62 | */ | |
63 | int no_unaligned_warning; | |
88fc241f | 64 | int unaligned_dump_stack; |
d2b176ed | 65 | |
1da177e4 LT |
66 | /* |
67 | * For M-unit: | |
68 | * | |
69 | * opcode | m | x6 | | |
70 | * --------|------|---------| | |
71 | * [40-37] | [36] | [35:30] | | |
72 | * --------|------|---------| | |
73 | * 4 | 1 | 6 | = 11 bits | |
74 | * -------------------------- | |
75 | * However bits [31:30] are not directly useful to distinguish between | |
76 | * load/store so we can use [35:32] instead, which gives the following | |
77 | * mask ([40:32]) using 9 bits. The 'e' comes from the fact that we defer | |
78 | * checking the m-bit until later in the load/store emulation. | |
79 | */ | |
80 | #define IA64_OPCODE_MASK 0x1ef | |
81 | #define IA64_OPCODE_SHIFT 32 | |
82 | ||
83 | /* | |
84 | * Table C-28 Integer Load/Store | |
85 | * | |
86 | * We ignore [35:32]= 0x6, 0x7, 0xE, 0xF | |
87 | * | |
88 | * ld8.fill, st8.fill MUST be aligned because the RNATs are based on | |
89 | * the address (bits [8:3]), so we must failed. | |
90 | */ | |
91 | #define LD_OP 0x080 | |
92 | #define LDS_OP 0x081 | |
93 | #define LDA_OP 0x082 | |
94 | #define LDSA_OP 0x083 | |
95 | #define LDBIAS_OP 0x084 | |
96 | #define LDACQ_OP 0x085 | |
97 | /* 0x086, 0x087 are not relevant */ | |
98 | #define LDCCLR_OP 0x088 | |
99 | #define LDCNC_OP 0x089 | |
100 | #define LDCCLRACQ_OP 0x08a | |
101 | #define ST_OP 0x08c | |
102 | #define STREL_OP 0x08d | |
103 | /* 0x08e,0x8f are not relevant */ | |
104 | ||
105 | /* | |
106 | * Table C-29 Integer Load +Reg | |
107 | * | |
108 | * we use the ld->m (bit [36:36]) field to determine whether or not we have | |
109 | * a load/store of this form. | |
110 | */ | |
111 | ||
112 | /* | |
113 | * Table C-30 Integer Load/Store +Imm | |
114 | * | |
115 | * We ignore [35:32]= 0x6, 0x7, 0xE, 0xF | |
116 | * | |
117 | * ld8.fill, st8.fill must be aligned because the Nat register are based on | |
118 | * the address, so we must fail and the program must be fixed. | |
119 | */ | |
120 | #define LD_IMM_OP 0x0a0 | |
121 | #define LDS_IMM_OP 0x0a1 | |
122 | #define LDA_IMM_OP 0x0a2 | |
123 | #define LDSA_IMM_OP 0x0a3 | |
124 | #define LDBIAS_IMM_OP 0x0a4 | |
125 | #define LDACQ_IMM_OP 0x0a5 | |
126 | /* 0x0a6, 0xa7 are not relevant */ | |
127 | #define LDCCLR_IMM_OP 0x0a8 | |
128 | #define LDCNC_IMM_OP 0x0a9 | |
129 | #define LDCCLRACQ_IMM_OP 0x0aa | |
130 | #define ST_IMM_OP 0x0ac | |
131 | #define STREL_IMM_OP 0x0ad | |
132 | /* 0x0ae,0xaf are not relevant */ | |
133 | ||
134 | /* | |
135 | * Table C-32 Floating-point Load/Store | |
136 | */ | |
137 | #define LDF_OP 0x0c0 | |
138 | #define LDFS_OP 0x0c1 | |
139 | #define LDFA_OP 0x0c2 | |
140 | #define LDFSA_OP 0x0c3 | |
141 | /* 0x0c6 is irrelevant */ | |
142 | #define LDFCCLR_OP 0x0c8 | |
143 | #define LDFCNC_OP 0x0c9 | |
144 | /* 0x0cb is irrelevant */ | |
145 | #define STF_OP 0x0cc | |
146 | ||
147 | /* | |
148 | * Table C-33 Floating-point Load +Reg | |
149 | * | |
150 | * we use the ld->m (bit [36:36]) field to determine whether or not we have | |
151 | * a load/store of this form. | |
152 | */ | |
153 | ||
154 | /* | |
155 | * Table C-34 Floating-point Load/Store +Imm | |
156 | */ | |
157 | #define LDF_IMM_OP 0x0e0 | |
158 | #define LDFS_IMM_OP 0x0e1 | |
159 | #define LDFA_IMM_OP 0x0e2 | |
160 | #define LDFSA_IMM_OP 0x0e3 | |
161 | /* 0x0e6 is irrelevant */ | |
162 | #define LDFCCLR_IMM_OP 0x0e8 | |
163 | #define LDFCNC_IMM_OP 0x0e9 | |
164 | #define STF_IMM_OP 0x0ec | |
165 | ||
166 | typedef struct { | |
167 | unsigned long qp:6; /* [0:5] */ | |
168 | unsigned long r1:7; /* [6:12] */ | |
169 | unsigned long imm:7; /* [13:19] */ | |
170 | unsigned long r3:7; /* [20:26] */ | |
171 | unsigned long x:1; /* [27:27] */ | |
172 | unsigned long hint:2; /* [28:29] */ | |
173 | unsigned long x6_sz:2; /* [30:31] */ | |
174 | unsigned long x6_op:4; /* [32:35], x6 = x6_sz|x6_op */ | |
175 | unsigned long m:1; /* [36:36] */ | |
176 | unsigned long op:4; /* [37:40] */ | |
177 | unsigned long pad:23; /* [41:63] */ | |
178 | } load_store_t; | |
179 | ||
180 | ||
181 | typedef enum { | |
182 | UPD_IMMEDIATE, /* ldXZ r1=[r3],imm(9) */ | |
183 | UPD_REG /* ldXZ r1=[r3],r2 */ | |
184 | } update_t; | |
185 | ||
186 | /* | |
187 | * We use tables to keep track of the offsets of registers in the saved state. | |
188 | * This way we save having big switch/case statements. | |
189 | * | |
190 | * We use bit 0 to indicate switch_stack or pt_regs. | |
191 | * The offset is simply shifted by 1 bit. | |
192 | * A 2-byte value should be enough to hold any kind of offset | |
193 | * | |
194 | * In case the calling convention changes (and thus pt_regs/switch_stack) | |
195 | * simply use RSW instead of RPT or vice-versa. | |
196 | */ | |
197 | ||
198 | #define RPO(x) ((size_t) &((struct pt_regs *)0)->x) | |
199 | #define RSO(x) ((size_t) &((struct switch_stack *)0)->x) | |
200 | ||
201 | #define RPT(x) (RPO(x) << 1) | |
202 | #define RSW(x) (1| RSO(x)<<1) | |
203 | ||
204 | #define GR_OFFS(x) (gr_info[x]>>1) | |
205 | #define GR_IN_SW(x) (gr_info[x] & 0x1) | |
206 | ||
207 | #define FR_OFFS(x) (fr_info[x]>>1) | |
208 | #define FR_IN_SW(x) (fr_info[x] & 0x1) | |
209 | ||
210 | static u16 gr_info[32]={ | |
211 | 0, /* r0 is read-only : WE SHOULD NEVER GET THIS */ | |
212 | ||
213 | RPT(r1), RPT(r2), RPT(r3), | |
214 | ||
215 | RSW(r4), RSW(r5), RSW(r6), RSW(r7), | |
216 | ||
217 | RPT(r8), RPT(r9), RPT(r10), RPT(r11), | |
218 | RPT(r12), RPT(r13), RPT(r14), RPT(r15), | |
219 | ||
220 | RPT(r16), RPT(r17), RPT(r18), RPT(r19), | |
221 | RPT(r20), RPT(r21), RPT(r22), RPT(r23), | |
222 | RPT(r24), RPT(r25), RPT(r26), RPT(r27), | |
223 | RPT(r28), RPT(r29), RPT(r30), RPT(r31) | |
224 | }; | |
225 | ||
226 | static u16 fr_info[32]={ | |
227 | 0, /* constant : WE SHOULD NEVER GET THIS */ | |
228 | 0, /* constant : WE SHOULD NEVER GET THIS */ | |
229 | ||
230 | RSW(f2), RSW(f3), RSW(f4), RSW(f5), | |
231 | ||
232 | RPT(f6), RPT(f7), RPT(f8), RPT(f9), | |
233 | RPT(f10), RPT(f11), | |
234 | ||
235 | RSW(f12), RSW(f13), RSW(f14), | |
236 | RSW(f15), RSW(f16), RSW(f17), RSW(f18), RSW(f19), | |
237 | RSW(f20), RSW(f21), RSW(f22), RSW(f23), RSW(f24), | |
238 | RSW(f25), RSW(f26), RSW(f27), RSW(f28), RSW(f29), | |
239 | RSW(f30), RSW(f31) | |
240 | }; | |
241 | ||
242 | /* Invalidate ALAT entry for integer register REGNO. */ | |
243 | static void | |
244 | invala_gr (int regno) | |
245 | { | |
246 | # define F(reg) case reg: ia64_invala_gr(reg); break | |
247 | ||
248 | switch (regno) { | |
249 | F( 0); F( 1); F( 2); F( 3); F( 4); F( 5); F( 6); F( 7); | |
250 | F( 8); F( 9); F( 10); F( 11); F( 12); F( 13); F( 14); F( 15); | |
251 | F( 16); F( 17); F( 18); F( 19); F( 20); F( 21); F( 22); F( 23); | |
252 | F( 24); F( 25); F( 26); F( 27); F( 28); F( 29); F( 30); F( 31); | |
253 | F( 32); F( 33); F( 34); F( 35); F( 36); F( 37); F( 38); F( 39); | |
254 | F( 40); F( 41); F( 42); F( 43); F( 44); F( 45); F( 46); F( 47); | |
255 | F( 48); F( 49); F( 50); F( 51); F( 52); F( 53); F( 54); F( 55); | |
256 | F( 56); F( 57); F( 58); F( 59); F( 60); F( 61); F( 62); F( 63); | |
257 | F( 64); F( 65); F( 66); F( 67); F( 68); F( 69); F( 70); F( 71); | |
258 | F( 72); F( 73); F( 74); F( 75); F( 76); F( 77); F( 78); F( 79); | |
259 | F( 80); F( 81); F( 82); F( 83); F( 84); F( 85); F( 86); F( 87); | |
260 | F( 88); F( 89); F( 90); F( 91); F( 92); F( 93); F( 94); F( 95); | |
261 | F( 96); F( 97); F( 98); F( 99); F(100); F(101); F(102); F(103); | |
262 | F(104); F(105); F(106); F(107); F(108); F(109); F(110); F(111); | |
263 | F(112); F(113); F(114); F(115); F(116); F(117); F(118); F(119); | |
264 | F(120); F(121); F(122); F(123); F(124); F(125); F(126); F(127); | |
265 | } | |
266 | # undef F | |
267 | } | |
268 | ||
269 | /* Invalidate ALAT entry for floating-point register REGNO. */ | |
270 | static void | |
271 | invala_fr (int regno) | |
272 | { | |
273 | # define F(reg) case reg: ia64_invala_fr(reg); break | |
274 | ||
275 | switch (regno) { | |
276 | F( 0); F( 1); F( 2); F( 3); F( 4); F( 5); F( 6); F( 7); | |
277 | F( 8); F( 9); F( 10); F( 11); F( 12); F( 13); F( 14); F( 15); | |
278 | F( 16); F( 17); F( 18); F( 19); F( 20); F( 21); F( 22); F( 23); | |
279 | F( 24); F( 25); F( 26); F( 27); F( 28); F( 29); F( 30); F( 31); | |
280 | F( 32); F( 33); F( 34); F( 35); F( 36); F( 37); F( 38); F( 39); | |
281 | F( 40); F( 41); F( 42); F( 43); F( 44); F( 45); F( 46); F( 47); | |
282 | F( 48); F( 49); F( 50); F( 51); F( 52); F( 53); F( 54); F( 55); | |
283 | F( 56); F( 57); F( 58); F( 59); F( 60); F( 61); F( 62); F( 63); | |
284 | F( 64); F( 65); F( 66); F( 67); F( 68); F( 69); F( 70); F( 71); | |
285 | F( 72); F( 73); F( 74); F( 75); F( 76); F( 77); F( 78); F( 79); | |
286 | F( 80); F( 81); F( 82); F( 83); F( 84); F( 85); F( 86); F( 87); | |
287 | F( 88); F( 89); F( 90); F( 91); F( 92); F( 93); F( 94); F( 95); | |
288 | F( 96); F( 97); F( 98); F( 99); F(100); F(101); F(102); F(103); | |
289 | F(104); F(105); F(106); F(107); F(108); F(109); F(110); F(111); | |
290 | F(112); F(113); F(114); F(115); F(116); F(117); F(118); F(119); | |
291 | F(120); F(121); F(122); F(123); F(124); F(125); F(126); F(127); | |
292 | } | |
293 | # undef F | |
294 | } | |
295 | ||
296 | static inline unsigned long | |
297 | rotate_reg (unsigned long sor, unsigned long rrb, unsigned long reg) | |
298 | { | |
299 | reg += rrb; | |
300 | if (reg >= sor) | |
301 | reg -= sor; | |
302 | return reg; | |
303 | } | |
304 | ||
305 | static void | |
306 | set_rse_reg (struct pt_regs *regs, unsigned long r1, unsigned long val, int nat) | |
307 | { | |
308 | struct switch_stack *sw = (struct switch_stack *) regs - 1; | |
309 | unsigned long *bsp, *bspstore, *addr, *rnat_addr, *ubs_end; | |
310 | unsigned long *kbs = (void *) current + IA64_RBS_OFFSET; | |
311 | unsigned long rnats, nat_mask; | |
312 | unsigned long on_kbs; | |
313 | long sof = (regs->cr_ifs) & 0x7f; | |
314 | long sor = 8 * ((regs->cr_ifs >> 14) & 0xf); | |
315 | long rrb_gr = (regs->cr_ifs >> 18) & 0x7f; | |
316 | long ridx = r1 - 32; | |
317 | ||
318 | if (ridx >= sof) { | |
319 | /* this should never happen, as the "rsvd register fault" has higher priority */ | |
320 | DPRINT("ignoring write to r%lu; only %lu registers are allocated!\n", r1, sof); | |
321 | return; | |
322 | } | |
323 | ||
324 | if (ridx < sor) | |
325 | ridx = rotate_reg(sor, rrb_gr, ridx); | |
326 | ||
327 | DPRINT("r%lu, sw.bspstore=%lx pt.bspstore=%lx sof=%ld sol=%ld ridx=%ld\n", | |
328 | r1, sw->ar_bspstore, regs->ar_bspstore, sof, (regs->cr_ifs >> 7) & 0x7f, ridx); | |
329 | ||
330 | on_kbs = ia64_rse_num_regs(kbs, (unsigned long *) sw->ar_bspstore); | |
331 | addr = ia64_rse_skip_regs((unsigned long *) sw->ar_bspstore, -sof + ridx); | |
332 | if (addr >= kbs) { | |
333 | /* the register is on the kernel backing store: easy... */ | |
334 | rnat_addr = ia64_rse_rnat_addr(addr); | |
335 | if ((unsigned long) rnat_addr >= sw->ar_bspstore) | |
336 | rnat_addr = &sw->ar_rnat; | |
337 | nat_mask = 1UL << ia64_rse_slot_num(addr); | |
338 | ||
339 | *addr = val; | |
340 | if (nat) | |
341 | *rnat_addr |= nat_mask; | |
342 | else | |
343 | *rnat_addr &= ~nat_mask; | |
344 | return; | |
345 | } | |
346 | ||
347 | if (!user_stack(current, regs)) { | |
348 | DPRINT("ignoring kernel write to r%lu; register isn't on the kernel RBS!", r1); | |
349 | return; | |
350 | } | |
351 | ||
352 | bspstore = (unsigned long *)regs->ar_bspstore; | |
353 | ubs_end = ia64_rse_skip_regs(bspstore, on_kbs); | |
354 | bsp = ia64_rse_skip_regs(ubs_end, -sof); | |
355 | addr = ia64_rse_skip_regs(bsp, ridx); | |
356 | ||
357 | DPRINT("ubs_end=%p bsp=%p addr=%p\n", (void *) ubs_end, (void *) bsp, (void *) addr); | |
358 | ||
359 | ia64_poke(current, sw, (unsigned long) ubs_end, (unsigned long) addr, val); | |
360 | ||
361 | rnat_addr = ia64_rse_rnat_addr(addr); | |
362 | ||
363 | ia64_peek(current, sw, (unsigned long) ubs_end, (unsigned long) rnat_addr, &rnats); | |
364 | DPRINT("rnat @%p = 0x%lx nat=%d old nat=%ld\n", | |
365 | (void *) rnat_addr, rnats, nat, (rnats >> ia64_rse_slot_num(addr)) & 1); | |
366 | ||
367 | nat_mask = 1UL << ia64_rse_slot_num(addr); | |
368 | if (nat) | |
369 | rnats |= nat_mask; | |
370 | else | |
371 | rnats &= ~nat_mask; | |
372 | ia64_poke(current, sw, (unsigned long) ubs_end, (unsigned long) rnat_addr, rnats); | |
373 | ||
374 | DPRINT("rnat changed to @%p = 0x%lx\n", (void *) rnat_addr, rnats); | |
375 | } | |
376 | ||
377 | ||
378 | static void | |
379 | get_rse_reg (struct pt_regs *regs, unsigned long r1, unsigned long *val, int *nat) | |
380 | { | |
381 | struct switch_stack *sw = (struct switch_stack *) regs - 1; | |
382 | unsigned long *bsp, *addr, *rnat_addr, *ubs_end, *bspstore; | |
383 | unsigned long *kbs = (void *) current + IA64_RBS_OFFSET; | |
384 | unsigned long rnats, nat_mask; | |
385 | unsigned long on_kbs; | |
386 | long sof = (regs->cr_ifs) & 0x7f; | |
387 | long sor = 8 * ((regs->cr_ifs >> 14) & 0xf); | |
388 | long rrb_gr = (regs->cr_ifs >> 18) & 0x7f; | |
389 | long ridx = r1 - 32; | |
390 | ||
391 | if (ridx >= sof) { | |
392 | /* read of out-of-frame register returns an undefined value; 0 in our case. */ | |
393 | DPRINT("ignoring read from r%lu; only %lu registers are allocated!\n", r1, sof); | |
394 | goto fail; | |
395 | } | |
396 | ||
397 | if (ridx < sor) | |
398 | ridx = rotate_reg(sor, rrb_gr, ridx); | |
399 | ||
400 | DPRINT("r%lu, sw.bspstore=%lx pt.bspstore=%lx sof=%ld sol=%ld ridx=%ld\n", | |
401 | r1, sw->ar_bspstore, regs->ar_bspstore, sof, (regs->cr_ifs >> 7) & 0x7f, ridx); | |
402 | ||
403 | on_kbs = ia64_rse_num_regs(kbs, (unsigned long *) sw->ar_bspstore); | |
404 | addr = ia64_rse_skip_regs((unsigned long *) sw->ar_bspstore, -sof + ridx); | |
405 | if (addr >= kbs) { | |
406 | /* the register is on the kernel backing store: easy... */ | |
407 | *val = *addr; | |
408 | if (nat) { | |
409 | rnat_addr = ia64_rse_rnat_addr(addr); | |
410 | if ((unsigned long) rnat_addr >= sw->ar_bspstore) | |
411 | rnat_addr = &sw->ar_rnat; | |
412 | nat_mask = 1UL << ia64_rse_slot_num(addr); | |
413 | *nat = (*rnat_addr & nat_mask) != 0; | |
414 | } | |
415 | return; | |
416 | } | |
417 | ||
418 | if (!user_stack(current, regs)) { | |
419 | DPRINT("ignoring kernel read of r%lu; register isn't on the RBS!", r1); | |
420 | goto fail; | |
421 | } | |
422 | ||
423 | bspstore = (unsigned long *)regs->ar_bspstore; | |
424 | ubs_end = ia64_rse_skip_regs(bspstore, on_kbs); | |
425 | bsp = ia64_rse_skip_regs(ubs_end, -sof); | |
426 | addr = ia64_rse_skip_regs(bsp, ridx); | |
427 | ||
428 | DPRINT("ubs_end=%p bsp=%p addr=%p\n", (void *) ubs_end, (void *) bsp, (void *) addr); | |
429 | ||
430 | ia64_peek(current, sw, (unsigned long) ubs_end, (unsigned long) addr, val); | |
431 | ||
432 | if (nat) { | |
433 | rnat_addr = ia64_rse_rnat_addr(addr); | |
434 | nat_mask = 1UL << ia64_rse_slot_num(addr); | |
435 | ||
436 | DPRINT("rnat @%p = 0x%lx\n", (void *) rnat_addr, rnats); | |
437 | ||
438 | ia64_peek(current, sw, (unsigned long) ubs_end, (unsigned long) rnat_addr, &rnats); | |
439 | *nat = (rnats & nat_mask) != 0; | |
440 | } | |
441 | return; | |
442 | ||
443 | fail: | |
444 | *val = 0; | |
445 | if (nat) | |
446 | *nat = 0; | |
447 | return; | |
448 | } | |
449 | ||
450 | ||
451 | static void | |
452 | setreg (unsigned long regnum, unsigned long val, int nat, struct pt_regs *regs) | |
453 | { | |
454 | struct switch_stack *sw = (struct switch_stack *) regs - 1; | |
455 | unsigned long addr; | |
456 | unsigned long bitmask; | |
457 | unsigned long *unat; | |
458 | ||
459 | /* | |
460 | * First takes care of stacked registers | |
461 | */ | |
462 | if (regnum >= IA64_FIRST_STACKED_GR) { | |
463 | set_rse_reg(regs, regnum, val, nat); | |
464 | return; | |
465 | } | |
466 | ||
467 | /* | |
468 | * Using r0 as a target raises a General Exception fault which has higher priority | |
469 | * than the Unaligned Reference fault. | |
470 | */ | |
471 | ||
472 | /* | |
473 | * Now look at registers in [0-31] range and init correct UNAT | |
474 | */ | |
475 | if (GR_IN_SW(regnum)) { | |
476 | addr = (unsigned long)sw; | |
477 | unat = &sw->ar_unat; | |
478 | } else { | |
479 | addr = (unsigned long)regs; | |
480 | unat = &sw->caller_unat; | |
481 | } | |
482 | DPRINT("tmp_base=%lx switch_stack=%s offset=%d\n", | |
483 | addr, unat==&sw->ar_unat ? "yes":"no", GR_OFFS(regnum)); | |
484 | /* | |
485 | * add offset from base of struct | |
486 | * and do it ! | |
487 | */ | |
488 | addr += GR_OFFS(regnum); | |
489 | ||
490 | *(unsigned long *)addr = val; | |
491 | ||
492 | /* | |
493 | * We need to clear the corresponding UNAT bit to fully emulate the load | |
494 | * UNAT bit_pos = GR[r3]{8:3} form EAS-2.4 | |
495 | */ | |
496 | bitmask = 1UL << (addr >> 3 & 0x3f); | |
497 | DPRINT("*0x%lx=0x%lx NaT=%d prev_unat @%p=%lx\n", addr, val, nat, (void *) unat, *unat); | |
498 | if (nat) { | |
499 | *unat |= bitmask; | |
500 | } else { | |
501 | *unat &= ~bitmask; | |
502 | } | |
503 | DPRINT("*0x%lx=0x%lx NaT=%d new unat: %p=%lx\n", addr, val, nat, (void *) unat,*unat); | |
504 | } | |
505 | ||
506 | /* | |
507 | * Return the (rotated) index for floating point register REGNUM (REGNUM must be in the | |
508 | * range from 32-127, result is in the range from 0-95. | |
509 | */ | |
510 | static inline unsigned long | |
511 | fph_index (struct pt_regs *regs, long regnum) | |
512 | { | |
513 | unsigned long rrb_fr = (regs->cr_ifs >> 25) & 0x7f; | |
514 | return rotate_reg(96, rrb_fr, (regnum - IA64_FIRST_ROTATING_FR)); | |
515 | } | |
516 | ||
517 | static void | |
518 | setfpreg (unsigned long regnum, struct ia64_fpreg *fpval, struct pt_regs *regs) | |
519 | { | |
520 | struct switch_stack *sw = (struct switch_stack *)regs - 1; | |
521 | unsigned long addr; | |
522 | ||
523 | /* | |
524 | * From EAS-2.5: FPDisableFault has higher priority than Unaligned | |
525 | * Fault. Thus, when we get here, we know the partition is enabled. | |
526 | * To update f32-f127, there are three choices: | |
527 | * | |
528 | * (1) save f32-f127 to thread.fph and update the values there | |
529 | * (2) use a gigantic switch statement to directly access the registers | |
530 | * (3) generate code on the fly to update the desired register | |
531 | * | |
532 | * For now, we are using approach (1). | |
533 | */ | |
534 | if (regnum >= IA64_FIRST_ROTATING_FR) { | |
535 | ia64_sync_fph(current); | |
536 | current->thread.fph[fph_index(regs, regnum)] = *fpval; | |
537 | } else { | |
538 | /* | |
539 | * pt_regs or switch_stack ? | |
540 | */ | |
541 | if (FR_IN_SW(regnum)) { | |
542 | addr = (unsigned long)sw; | |
543 | } else { | |
544 | addr = (unsigned long)regs; | |
545 | } | |
546 | ||
547 | DPRINT("tmp_base=%lx offset=%d\n", addr, FR_OFFS(regnum)); | |
548 | ||
549 | addr += FR_OFFS(regnum); | |
550 | *(struct ia64_fpreg *)addr = *fpval; | |
551 | ||
552 | /* | |
553 | * mark the low partition as being used now | |
554 | * | |
555 | * It is highly unlikely that this bit is not already set, but | |
556 | * let's do it for safety. | |
557 | */ | |
558 | regs->cr_ipsr |= IA64_PSR_MFL; | |
559 | } | |
560 | } | |
561 | ||
562 | /* | |
563 | * Those 2 inline functions generate the spilled versions of the constant floating point | |
564 | * registers which can be used with stfX | |
565 | */ | |
566 | static inline void | |
567 | float_spill_f0 (struct ia64_fpreg *final) | |
568 | { | |
569 | ia64_stf_spill(final, 0); | |
570 | } | |
571 | ||
572 | static inline void | |
573 | float_spill_f1 (struct ia64_fpreg *final) | |
574 | { | |
575 | ia64_stf_spill(final, 1); | |
576 | } | |
577 | ||
578 | static void | |
579 | getfpreg (unsigned long regnum, struct ia64_fpreg *fpval, struct pt_regs *regs) | |
580 | { | |
581 | struct switch_stack *sw = (struct switch_stack *) regs - 1; | |
582 | unsigned long addr; | |
583 | ||
584 | /* | |
585 | * From EAS-2.5: FPDisableFault has higher priority than | |
586 | * Unaligned Fault. Thus, when we get here, we know the partition is | |
587 | * enabled. | |
588 | * | |
589 | * When regnum > 31, the register is still live and we need to force a save | |
590 | * to current->thread.fph to get access to it. See discussion in setfpreg() | |
591 | * for reasons and other ways of doing this. | |
592 | */ | |
593 | if (regnum >= IA64_FIRST_ROTATING_FR) { | |
594 | ia64_flush_fph(current); | |
595 | *fpval = current->thread.fph[fph_index(regs, regnum)]; | |
596 | } else { | |
597 | /* | |
598 | * f0 = 0.0, f1= 1.0. Those registers are constant and are thus | |
599 | * not saved, we must generate their spilled form on the fly | |
600 | */ | |
601 | switch(regnum) { | |
602 | case 0: | |
603 | float_spill_f0(fpval); | |
604 | break; | |
605 | case 1: | |
606 | float_spill_f1(fpval); | |
607 | break; | |
608 | default: | |
609 | /* | |
610 | * pt_regs or switch_stack ? | |
611 | */ | |
612 | addr = FR_IN_SW(regnum) ? (unsigned long)sw | |
613 | : (unsigned long)regs; | |
614 | ||
615 | DPRINT("is_sw=%d tmp_base=%lx offset=0x%x\n", | |
616 | FR_IN_SW(regnum), addr, FR_OFFS(regnum)); | |
617 | ||
618 | addr += FR_OFFS(regnum); | |
619 | *fpval = *(struct ia64_fpreg *)addr; | |
620 | } | |
621 | } | |
622 | } | |
623 | ||
624 | ||
625 | static void | |
626 | getreg (unsigned long regnum, unsigned long *val, int *nat, struct pt_regs *regs) | |
627 | { | |
628 | struct switch_stack *sw = (struct switch_stack *) regs - 1; | |
629 | unsigned long addr, *unat; | |
630 | ||
631 | if (regnum >= IA64_FIRST_STACKED_GR) { | |
632 | get_rse_reg(regs, regnum, val, nat); | |
633 | return; | |
634 | } | |
635 | ||
636 | /* | |
637 | * take care of r0 (read-only always evaluate to 0) | |
638 | */ | |
639 | if (regnum == 0) { | |
640 | *val = 0; | |
641 | if (nat) | |
642 | *nat = 0; | |
643 | return; | |
644 | } | |
645 | ||
646 | /* | |
647 | * Now look at registers in [0-31] range and init correct UNAT | |
648 | */ | |
649 | if (GR_IN_SW(regnum)) { | |
650 | addr = (unsigned long)sw; | |
651 | unat = &sw->ar_unat; | |
652 | } else { | |
653 | addr = (unsigned long)regs; | |
654 | unat = &sw->caller_unat; | |
655 | } | |
656 | ||
657 | DPRINT("addr_base=%lx offset=0x%x\n", addr, GR_OFFS(regnum)); | |
658 | ||
659 | addr += GR_OFFS(regnum); | |
660 | ||
661 | *val = *(unsigned long *)addr; | |
662 | ||
663 | /* | |
664 | * do it only when requested | |
665 | */ | |
666 | if (nat) | |
667 | *nat = (*unat >> (addr >> 3 & 0x3f)) & 0x1UL; | |
668 | } | |
669 | ||
670 | static void | |
671 | emulate_load_updates (update_t type, load_store_t ld, struct pt_regs *regs, unsigned long ifa) | |
672 | { | |
673 | /* | |
674 | * IMPORTANT: | |
675 | * Given the way we handle unaligned speculative loads, we should | |
676 | * not get to this point in the code but we keep this sanity check, | |
677 | * just in case. | |
678 | */ | |
679 | if (ld.x6_op == 1 || ld.x6_op == 3) { | |
d4ed8084 | 680 | printk(KERN_ERR "%s: register update on speculative load, error\n", __func__); |
620de2f5 JB |
681 | if (die_if_kernel("unaligned reference on speculative load with register update\n", |
682 | regs, 30)) | |
683 | return; | |
1da177e4 LT |
684 | } |
685 | ||
686 | ||
687 | /* | |
688 | * at this point, we know that the base register to update is valid i.e., | |
689 | * it's not r0 | |
690 | */ | |
691 | if (type == UPD_IMMEDIATE) { | |
692 | unsigned long imm; | |
693 | ||
694 | /* | |
695 | * Load +Imm: ldXZ r1=[r3],imm(9) | |
696 | * | |
697 | * | |
698 | * form imm9: [13:19] contain the first 7 bits | |
699 | */ | |
700 | imm = ld.x << 7 | ld.imm; | |
701 | ||
702 | /* | |
703 | * sign extend (1+8bits) if m set | |
704 | */ | |
705 | if (ld.m) imm |= SIGN_EXT9; | |
706 | ||
707 | /* | |
708 | * ifa == r3 and we know that the NaT bit on r3 was clear so | |
709 | * we can directly use ifa. | |
710 | */ | |
711 | ifa += imm; | |
712 | ||
713 | setreg(ld.r3, ifa, 0, regs); | |
714 | ||
715 | DPRINT("ld.x=%d ld.m=%d imm=%ld r3=0x%lx\n", ld.x, ld.m, imm, ifa); | |
716 | ||
717 | } else if (ld.m) { | |
718 | unsigned long r2; | |
719 | int nat_r2; | |
720 | ||
721 | /* | |
722 | * Load +Reg Opcode: ldXZ r1=[r3],r2 | |
723 | * | |
724 | * Note: that we update r3 even in the case of ldfX.a | |
725 | * (where the load does not happen) | |
726 | * | |
727 | * The way the load algorithm works, we know that r3 does not | |
728 | * have its NaT bit set (would have gotten NaT consumption | |
729 | * before getting the unaligned fault). So we can use ifa | |
730 | * which equals r3 at this point. | |
731 | * | |
732 | * IMPORTANT: | |
733 | * The above statement holds ONLY because we know that we | |
734 | * never reach this code when trying to do a ldX.s. | |
735 | * If we ever make it to here on an ldfX.s then | |
736 | */ | |
737 | getreg(ld.imm, &r2, &nat_r2, regs); | |
738 | ||
739 | ifa += r2; | |
740 | ||
741 | /* | |
742 | * propagate Nat r2 -> r3 | |
743 | */ | |
744 | setreg(ld.r3, ifa, nat_r2, regs); | |
745 | ||
746 | DPRINT("imm=%d r2=%ld r3=0x%lx nat_r2=%d\n",ld.imm, r2, ifa, nat_r2); | |
747 | } | |
748 | } | |
749 | ||
750 | ||
751 | static int | |
752 | emulate_load_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs) | |
753 | { | |
754 | unsigned int len = 1 << ld.x6_sz; | |
755 | unsigned long val = 0; | |
756 | ||
757 | /* | |
758 | * r0, as target, doesn't need to be checked because Illegal Instruction | |
759 | * faults have higher priority than unaligned faults. | |
760 | * | |
761 | * r0 cannot be found as the base as it would never generate an | |
762 | * unaligned reference. | |
763 | */ | |
764 | ||
765 | /* | |
766 | * ldX.a we will emulate load and also invalidate the ALAT entry. | |
767 | * See comment below for explanation on how we handle ldX.a | |
768 | */ | |
769 | ||
770 | if (len != 2 && len != 4 && len != 8) { | |
771 | DPRINT("unknown size: x6=%d\n", ld.x6_sz); | |
772 | return -1; | |
773 | } | |
774 | /* this assumes little-endian byte-order: */ | |
775 | if (copy_from_user(&val, (void __user *) ifa, len)) | |
776 | return -1; | |
777 | setreg(ld.r1, val, 0, regs); | |
778 | ||
779 | /* | |
780 | * check for updates on any kind of loads | |
781 | */ | |
782 | if (ld.op == 0x5 || ld.m) | |
783 | emulate_load_updates(ld.op == 0x5 ? UPD_IMMEDIATE: UPD_REG, ld, regs, ifa); | |
784 | ||
785 | /* | |
786 | * handling of various loads (based on EAS2.4): | |
787 | * | |
788 | * ldX.acq (ordered load): | |
789 | * - acquire semantics would have been used, so force fence instead. | |
790 | * | |
791 | * ldX.c.clr (check load and clear): | |
792 | * - if we get to this handler, it's because the entry was not in the ALAT. | |
793 | * Therefore the operation reverts to a normal load | |
794 | * | |
795 | * ldX.c.nc (check load no clear): | |
796 | * - same as previous one | |
797 | * | |
798 | * ldX.c.clr.acq (ordered check load and clear): | |
799 | * - same as above for c.clr part. The load needs to have acquire semantics. So | |
800 | * we use the fence semantics which is stronger and thus ensures correctness. | |
801 | * | |
802 | * ldX.a (advanced load): | |
803 | * - suppose ldX.a r1=[r3]. If we get to the unaligned trap it's because the | |
804 | * address doesn't match requested size alignment. This means that we would | |
805 | * possibly need more than one load to get the result. | |
806 | * | |
807 | * The load part can be handled just like a normal load, however the difficult | |
808 | * part is to get the right thing into the ALAT. The critical piece of information | |
809 | * in the base address of the load & size. To do that, a ld.a must be executed, | |
810 | * clearly any address can be pushed into the table by using ld1.a r1=[r3]. Now | |
811 | * if we use the same target register, we will be okay for the check.a instruction. | |
812 | * If we look at the store, basically a stX [r3]=r1 checks the ALAT for any entry | |
813 | * which would overlap within [r3,r3+X] (the size of the load was store in the | |
814 | * ALAT). If such an entry is found the entry is invalidated. But this is not good | |
815 | * enough, take the following example: | |
816 | * r3=3 | |
817 | * ld4.a r1=[r3] | |
818 | * | |
819 | * Could be emulated by doing: | |
820 | * ld1.a r1=[r3],1 | |
821 | * store to temporary; | |
822 | * ld1.a r1=[r3],1 | |
823 | * store & shift to temporary; | |
824 | * ld1.a r1=[r3],1 | |
825 | * store & shift to temporary; | |
826 | * ld1.a r1=[r3] | |
827 | * store & shift to temporary; | |
828 | * r1=temporary | |
829 | * | |
830 | * So in this case, you would get the right value is r1 but the wrong info in | |
831 | * the ALAT. Notice that you could do it in reverse to finish with address 3 | |
832 | * but you would still get the size wrong. To get the size right, one needs to | |
833 | * execute exactly the same kind of load. You could do it from a aligned | |
834 | * temporary location, but you would get the address wrong. | |
835 | * | |
836 | * So no matter what, it is not possible to emulate an advanced load | |
837 | * correctly. But is that really critical ? | |
838 | * | |
839 | * We will always convert ld.a into a normal load with ALAT invalidated. This | |
840 | * will enable compiler to do optimization where certain code path after ld.a | |
841 | * is not required to have ld.c/chk.a, e.g., code path with no intervening stores. | |
842 | * | |
843 | * If there is a store after the advanced load, one must either do a ld.c.* or | |
844 | * chk.a.* to reuse the value stored in the ALAT. Both can "fail" (meaning no | |
845 | * entry found in ALAT), and that's perfectly ok because: | |
846 | * | |
847 | * - ld.c.*, if the entry is not present a normal load is executed | |
848 | * - chk.a.*, if the entry is not present, execution jumps to recovery code | |
849 | * | |
850 | * In either case, the load can be potentially retried in another form. | |
851 | * | |
852 | * ALAT must be invalidated for the register (so that chk.a or ld.c don't pick | |
853 | * up a stale entry later). The register base update MUST also be performed. | |
854 | */ | |
855 | ||
856 | /* | |
857 | * when the load has the .acq completer then | |
858 | * use ordering fence. | |
859 | */ | |
860 | if (ld.x6_op == 0x5 || ld.x6_op == 0xa) | |
861 | mb(); | |
862 | ||
863 | /* | |
864 | * invalidate ALAT entry in case of advanced load | |
865 | */ | |
866 | if (ld.x6_op == 0x2) | |
867 | invala_gr(ld.r1); | |
868 | ||
869 | return 0; | |
870 | } | |
871 | ||
872 | static int | |
873 | emulate_store_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs) | |
874 | { | |
875 | unsigned long r2; | |
876 | unsigned int len = 1 << ld.x6_sz; | |
877 | ||
878 | /* | |
879 | * if we get to this handler, Nat bits on both r3 and r2 have already | |
880 | * been checked. so we don't need to do it | |
881 | * | |
882 | * extract the value to be stored | |
883 | */ | |
884 | getreg(ld.imm, &r2, NULL, regs); | |
885 | ||
886 | /* | |
887 | * we rely on the macros in unaligned.h for now i.e., | |
888 | * we let the compiler figure out how to read memory gracefully. | |
889 | * | |
890 | * We need this switch/case because the way the inline function | |
891 | * works. The code is optimized by the compiler and looks like | |
892 | * a single switch/case. | |
893 | */ | |
894 | DPRINT("st%d [%lx]=%lx\n", len, ifa, r2); | |
895 | ||
896 | if (len != 2 && len != 4 && len != 8) { | |
897 | DPRINT("unknown size: x6=%d\n", ld.x6_sz); | |
898 | return -1; | |
899 | } | |
900 | ||
901 | /* this assumes little-endian byte-order: */ | |
902 | if (copy_to_user((void __user *) ifa, &r2, len)) | |
903 | return -1; | |
904 | ||
905 | /* | |
906 | * stX [r3]=r2,imm(9) | |
907 | * | |
908 | * NOTE: | |
909 | * ld.r3 can never be r0, because r0 would not generate an | |
910 | * unaligned access. | |
911 | */ | |
912 | if (ld.op == 0x5) { | |
913 | unsigned long imm; | |
914 | ||
915 | /* | |
916 | * form imm9: [12:6] contain first 7bits | |
917 | */ | |
918 | imm = ld.x << 7 | ld.r1; | |
919 | /* | |
920 | * sign extend (8bits) if m set | |
921 | */ | |
922 | if (ld.m) imm |= SIGN_EXT9; | |
923 | /* | |
924 | * ifa == r3 (NaT is necessarily cleared) | |
925 | */ | |
926 | ifa += imm; | |
927 | ||
928 | DPRINT("imm=%lx r3=%lx\n", imm, ifa); | |
929 | ||
930 | setreg(ld.r3, ifa, 0, regs); | |
931 | } | |
932 | /* | |
933 | * we don't have alat_invalidate_multiple() so we need | |
934 | * to do the complete flush :-<< | |
935 | */ | |
936 | ia64_invala(); | |
937 | ||
938 | /* | |
939 | * stX.rel: use fence instead of release | |
940 | */ | |
941 | if (ld.x6_op == 0xd) | |
942 | mb(); | |
943 | ||
944 | return 0; | |
945 | } | |
946 | ||
947 | /* | |
948 | * floating point operations sizes in bytes | |
949 | */ | |
950 | static const unsigned char float_fsz[4]={ | |
951 | 10, /* extended precision (e) */ | |
952 | 8, /* integer (8) */ | |
953 | 4, /* single precision (s) */ | |
954 | 8 /* double precision (d) */ | |
955 | }; | |
956 | ||
957 | static inline void | |
958 | mem2float_extended (struct ia64_fpreg *init, struct ia64_fpreg *final) | |
959 | { | |
960 | ia64_ldfe(6, init); | |
961 | ia64_stop(); | |
962 | ia64_stf_spill(final, 6); | |
963 | } | |
964 | ||
965 | static inline void | |
966 | mem2float_integer (struct ia64_fpreg *init, struct ia64_fpreg *final) | |
967 | { | |
968 | ia64_ldf8(6, init); | |
969 | ia64_stop(); | |
970 | ia64_stf_spill(final, 6); | |
971 | } | |
972 | ||
973 | static inline void | |
974 | mem2float_single (struct ia64_fpreg *init, struct ia64_fpreg *final) | |
975 | { | |
976 | ia64_ldfs(6, init); | |
977 | ia64_stop(); | |
978 | ia64_stf_spill(final, 6); | |
979 | } | |
980 | ||
981 | static inline void | |
982 | mem2float_double (struct ia64_fpreg *init, struct ia64_fpreg *final) | |
983 | { | |
984 | ia64_ldfd(6, init); | |
985 | ia64_stop(); | |
986 | ia64_stf_spill(final, 6); | |
987 | } | |
988 | ||
989 | static inline void | |
990 | float2mem_extended (struct ia64_fpreg *init, struct ia64_fpreg *final) | |
991 | { | |
992 | ia64_ldf_fill(6, init); | |
993 | ia64_stop(); | |
994 | ia64_stfe(final, 6); | |
995 | } | |
996 | ||
997 | static inline void | |
998 | float2mem_integer (struct ia64_fpreg *init, struct ia64_fpreg *final) | |
999 | { | |
1000 | ia64_ldf_fill(6, init); | |
1001 | ia64_stop(); | |
1002 | ia64_stf8(final, 6); | |
1003 | } | |
1004 | ||
1005 | static inline void | |
1006 | float2mem_single (struct ia64_fpreg *init, struct ia64_fpreg *final) | |
1007 | { | |
1008 | ia64_ldf_fill(6, init); | |
1009 | ia64_stop(); | |
1010 | ia64_stfs(final, 6); | |
1011 | } | |
1012 | ||
1013 | static inline void | |
1014 | float2mem_double (struct ia64_fpreg *init, struct ia64_fpreg *final) | |
1015 | { | |
1016 | ia64_ldf_fill(6, init); | |
1017 | ia64_stop(); | |
1018 | ia64_stfd(final, 6); | |
1019 | } | |
1020 | ||
1021 | static int | |
1022 | emulate_load_floatpair (unsigned long ifa, load_store_t ld, struct pt_regs *regs) | |
1023 | { | |
1024 | struct ia64_fpreg fpr_init[2]; | |
1025 | struct ia64_fpreg fpr_final[2]; | |
1026 | unsigned long len = float_fsz[ld.x6_sz]; | |
1027 | ||
1028 | /* | |
1029 | * fr0 & fr1 don't need to be checked because Illegal Instruction faults have | |
1030 | * higher priority than unaligned faults. | |
1031 | * | |
1032 | * r0 cannot be found as the base as it would never generate an unaligned | |
1033 | * reference. | |
1034 | */ | |
1035 | ||
1036 | /* | |
1037 | * make sure we get clean buffers | |
1038 | */ | |
1039 | memset(&fpr_init, 0, sizeof(fpr_init)); | |
1040 | memset(&fpr_final, 0, sizeof(fpr_final)); | |
1041 | ||
1042 | /* | |
1043 | * ldfpX.a: we don't try to emulate anything but we must | |
1044 | * invalidate the ALAT entry and execute updates, if any. | |
1045 | */ | |
1046 | if (ld.x6_op != 0x2) { | |
1047 | /* | |
1048 | * This assumes little-endian byte-order. Note that there is no "ldfpe" | |
1049 | * instruction: | |
1050 | */ | |
1051 | if (copy_from_user(&fpr_init[0], (void __user *) ifa, len) | |
1052 | || copy_from_user(&fpr_init[1], (void __user *) (ifa + len), len)) | |
1053 | return -1; | |
1054 | ||
1055 | DPRINT("ld.r1=%d ld.imm=%d x6_sz=%d\n", ld.r1, ld.imm, ld.x6_sz); | |
1056 | DDUMP("frp_init =", &fpr_init, 2*len); | |
1057 | /* | |
1058 | * XXX fixme | |
1059 | * Could optimize inlines by using ldfpX & 2 spills | |
1060 | */ | |
1061 | switch( ld.x6_sz ) { | |
1062 | case 0: | |
1063 | mem2float_extended(&fpr_init[0], &fpr_final[0]); | |
1064 | mem2float_extended(&fpr_init[1], &fpr_final[1]); | |
1065 | break; | |
1066 | case 1: | |
1067 | mem2float_integer(&fpr_init[0], &fpr_final[0]); | |
1068 | mem2float_integer(&fpr_init[1], &fpr_final[1]); | |
1069 | break; | |
1070 | case 2: | |
1071 | mem2float_single(&fpr_init[0], &fpr_final[0]); | |
1072 | mem2float_single(&fpr_init[1], &fpr_final[1]); | |
1073 | break; | |
1074 | case 3: | |
1075 | mem2float_double(&fpr_init[0], &fpr_final[0]); | |
1076 | mem2float_double(&fpr_init[1], &fpr_final[1]); | |
1077 | break; | |
1078 | } | |
1079 | DDUMP("fpr_final =", &fpr_final, 2*len); | |
1080 | /* | |
1081 | * XXX fixme | |
1082 | * | |
1083 | * A possible optimization would be to drop fpr_final and directly | |
1084 | * use the storage from the saved context i.e., the actual final | |
1085 | * destination (pt_regs, switch_stack or thread structure). | |
1086 | */ | |
1087 | setfpreg(ld.r1, &fpr_final[0], regs); | |
1088 | setfpreg(ld.imm, &fpr_final[1], regs); | |
1089 | } | |
1090 | ||
1091 | /* | |
1092 | * Check for updates: only immediate updates are available for this | |
1093 | * instruction. | |
1094 | */ | |
1095 | if (ld.m) { | |
1096 | /* | |
1097 | * the immediate is implicit given the ldsz of the operation: | |
1098 | * single: 8 (2x4) and for all others it's 16 (2x8) | |
1099 | */ | |
1100 | ifa += len<<1; | |
1101 | ||
1102 | /* | |
1103 | * IMPORTANT: | |
1104 | * the fact that we force the NaT of r3 to zero is ONLY valid | |
1105 | * as long as we don't come here with a ldfpX.s. | |
1106 | * For this reason we keep this sanity check | |
1107 | */ | |
1108 | if (ld.x6_op == 1 || ld.x6_op == 3) | |
1109 | printk(KERN_ERR "%s: register update on speculative load pair, error\n", | |
d4ed8084 | 1110 | __func__); |
1da177e4 LT |
1111 | |
1112 | setreg(ld.r3, ifa, 0, regs); | |
1113 | } | |
1114 | ||
1115 | /* | |
1116 | * Invalidate ALAT entries, if any, for both registers. | |
1117 | */ | |
1118 | if (ld.x6_op == 0x2) { | |
1119 | invala_fr(ld.r1); | |
1120 | invala_fr(ld.imm); | |
1121 | } | |
1122 | return 0; | |
1123 | } | |
1124 | ||
1125 | ||
1126 | static int | |
1127 | emulate_load_float (unsigned long ifa, load_store_t ld, struct pt_regs *regs) | |
1128 | { | |
1129 | struct ia64_fpreg fpr_init; | |
1130 | struct ia64_fpreg fpr_final; | |
1131 | unsigned long len = float_fsz[ld.x6_sz]; | |
1132 | ||
1133 | /* | |
1134 | * fr0 & fr1 don't need to be checked because Illegal Instruction | |
1135 | * faults have higher priority than unaligned faults. | |
1136 | * | |
1137 | * r0 cannot be found as the base as it would never generate an | |
1138 | * unaligned reference. | |
1139 | */ | |
1140 | ||
1141 | /* | |
1142 | * make sure we get clean buffers | |
1143 | */ | |
1144 | memset(&fpr_init,0, sizeof(fpr_init)); | |
1145 | memset(&fpr_final,0, sizeof(fpr_final)); | |
1146 | ||
1147 | /* | |
1148 | * ldfX.a we don't try to emulate anything but we must | |
1149 | * invalidate the ALAT entry. | |
1150 | * See comments in ldX for descriptions on how the various loads are handled. | |
1151 | */ | |
1152 | if (ld.x6_op != 0x2) { | |
1153 | if (copy_from_user(&fpr_init, (void __user *) ifa, len)) | |
1154 | return -1; | |
1155 | ||
1156 | DPRINT("ld.r1=%d x6_sz=%d\n", ld.r1, ld.x6_sz); | |
1157 | DDUMP("fpr_init =", &fpr_init, len); | |
1158 | /* | |
1159 | * we only do something for x6_op={0,8,9} | |
1160 | */ | |
1161 | switch( ld.x6_sz ) { | |
1162 | case 0: | |
1163 | mem2float_extended(&fpr_init, &fpr_final); | |
1164 | break; | |
1165 | case 1: | |
1166 | mem2float_integer(&fpr_init, &fpr_final); | |
1167 | break; | |
1168 | case 2: | |
1169 | mem2float_single(&fpr_init, &fpr_final); | |
1170 | break; | |
1171 | case 3: | |
1172 | mem2float_double(&fpr_init, &fpr_final); | |
1173 | break; | |
1174 | } | |
1175 | DDUMP("fpr_final =", &fpr_final, len); | |
1176 | /* | |
1177 | * XXX fixme | |
1178 | * | |
1179 | * A possible optimization would be to drop fpr_final and directly | |
1180 | * use the storage from the saved context i.e., the actual final | |
1181 | * destination (pt_regs, switch_stack or thread structure). | |
1182 | */ | |
1183 | setfpreg(ld.r1, &fpr_final, regs); | |
1184 | } | |
1185 | ||
1186 | /* | |
1187 | * check for updates on any loads | |
1188 | */ | |
1189 | if (ld.op == 0x7 || ld.m) | |
1190 | emulate_load_updates(ld.op == 0x7 ? UPD_IMMEDIATE: UPD_REG, ld, regs, ifa); | |
1191 | ||
1192 | /* | |
1193 | * invalidate ALAT entry in case of advanced floating point loads | |
1194 | */ | |
1195 | if (ld.x6_op == 0x2) | |
1196 | invala_fr(ld.r1); | |
1197 | ||
1198 | return 0; | |
1199 | } | |
1200 | ||
1201 | ||
1202 | static int | |
1203 | emulate_store_float (unsigned long ifa, load_store_t ld, struct pt_regs *regs) | |
1204 | { | |
1205 | struct ia64_fpreg fpr_init; | |
1206 | struct ia64_fpreg fpr_final; | |
1207 | unsigned long len = float_fsz[ld.x6_sz]; | |
1208 | ||
1209 | /* | |
1210 | * make sure we get clean buffers | |
1211 | */ | |
1212 | memset(&fpr_init,0, sizeof(fpr_init)); | |
1213 | memset(&fpr_final,0, sizeof(fpr_final)); | |
1214 | ||
1215 | /* | |
1216 | * if we get to this handler, Nat bits on both r3 and r2 have already | |
1217 | * been checked. so we don't need to do it | |
1218 | * | |
1219 | * extract the value to be stored | |
1220 | */ | |
1221 | getfpreg(ld.imm, &fpr_init, regs); | |
1222 | /* | |
1223 | * during this step, we extract the spilled registers from the saved | |
1224 | * context i.e., we refill. Then we store (no spill) to temporary | |
1225 | * aligned location | |
1226 | */ | |
1227 | switch( ld.x6_sz ) { | |
1228 | case 0: | |
1229 | float2mem_extended(&fpr_init, &fpr_final); | |
1230 | break; | |
1231 | case 1: | |
1232 | float2mem_integer(&fpr_init, &fpr_final); | |
1233 | break; | |
1234 | case 2: | |
1235 | float2mem_single(&fpr_init, &fpr_final); | |
1236 | break; | |
1237 | case 3: | |
1238 | float2mem_double(&fpr_init, &fpr_final); | |
1239 | break; | |
1240 | } | |
1241 | DPRINT("ld.r1=%d x6_sz=%d\n", ld.r1, ld.x6_sz); | |
1242 | DDUMP("fpr_init =", &fpr_init, len); | |
1243 | DDUMP("fpr_final =", &fpr_final, len); | |
1244 | ||
1245 | if (copy_to_user((void __user *) ifa, &fpr_final, len)) | |
1246 | return -1; | |
1247 | ||
1248 | /* | |
1249 | * stfX [r3]=r2,imm(9) | |
1250 | * | |
1251 | * NOTE: | |
1252 | * ld.r3 can never be r0, because r0 would not generate an | |
1253 | * unaligned access. | |
1254 | */ | |
1255 | if (ld.op == 0x7) { | |
1256 | unsigned long imm; | |
1257 | ||
1258 | /* | |
1259 | * form imm9: [12:6] contain first 7bits | |
1260 | */ | |
1261 | imm = ld.x << 7 | ld.r1; | |
1262 | /* | |
1263 | * sign extend (8bits) if m set | |
1264 | */ | |
1265 | if (ld.m) | |
1266 | imm |= SIGN_EXT9; | |
1267 | /* | |
1268 | * ifa == r3 (NaT is necessarily cleared) | |
1269 | */ | |
1270 | ifa += imm; | |
1271 | ||
1272 | DPRINT("imm=%lx r3=%lx\n", imm, ifa); | |
1273 | ||
1274 | setreg(ld.r3, ifa, 0, regs); | |
1275 | } | |
1276 | /* | |
1277 | * we don't have alat_invalidate_multiple() so we need | |
1278 | * to do the complete flush :-<< | |
1279 | */ | |
1280 | ia64_invala(); | |
1281 | ||
1282 | return 0; | |
1283 | } | |
1284 | ||
1285 | /* | |
1286 | * Make sure we log the unaligned access, so that user/sysadmin can notice it and | |
1287 | * eventually fix the program. However, we don't want to do that for every access so we | |
7683a3f9 | 1288 | * pace it with jiffies. |
1da177e4 | 1289 | */ |
7683a3f9 | 1290 | static DEFINE_RATELIMIT_STATE(logging_rate_limit, 5 * HZ, 5); |
1da177e4 LT |
1291 | |
1292 | void | |
1293 | ia64_handle_unaligned (unsigned long ifa, struct pt_regs *regs) | |
1294 | { | |
1295 | struct ia64_psr *ipsr = ia64_psr(regs); | |
1296 | mm_segment_t old_fs = get_fs(); | |
1297 | unsigned long bundle[2]; | |
1298 | unsigned long opcode; | |
1299 | struct siginfo si; | |
1300 | const struct exception_table_entry *eh = NULL; | |
1301 | union { | |
1302 | unsigned long l; | |
1303 | load_store_t insn; | |
1304 | } u; | |
1305 | int ret = -1; | |
1306 | ||
1307 | if (ia64_psr(regs)->be) { | |
1308 | /* we don't support big-endian accesses */ | |
620de2f5 JB |
1309 | if (die_if_kernel("big-endian unaligned accesses are not supported", regs, 0)) |
1310 | return; | |
1da177e4 LT |
1311 | goto force_sigbus; |
1312 | } | |
1313 | ||
1314 | /* | |
1315 | * Treat kernel accesses for which there is an exception handler entry the same as | |
1316 | * user-level unaligned accesses. Otherwise, a clever program could trick this | |
1317 | * handler into reading an arbitrary kernel addresses... | |
1318 | */ | |
1319 | if (!user_mode(regs)) | |
1320 | eh = search_exception_tables(regs->cr_iip + ia64_psr(regs)->ri); | |
1321 | if (user_mode(regs) || eh) { | |
1322 | if ((current->thread.flags & IA64_THREAD_UAC_SIGBUS) != 0) | |
1323 | goto force_sigbus; | |
1324 | ||
d2b176ed JS |
1325 | if (!no_unaligned_warning && |
1326 | !(current->thread.flags & IA64_THREAD_UAC_NOPRINT) && | |
7683a3f9 | 1327 | __ratelimit(&logging_rate_limit)) |
1da177e4 LT |
1328 | { |
1329 | char buf[200]; /* comm[] is at most 16 bytes... */ | |
1330 | size_t len; | |
1331 | ||
1332 | len = sprintf(buf, "%s(%d): unaligned access to 0x%016lx, " | |
19c5870c AD |
1333 | "ip=0x%016lx\n\r", current->comm, |
1334 | task_pid_nr(current), | |
1da177e4 LT |
1335 | ifa, regs->cr_iip + ipsr->ri); |
1336 | /* | |
1337 | * Don't call tty_write_message() if we're in the kernel; we might | |
1338 | * be holding locks... | |
1339 | */ | |
02f14c79 PH |
1340 | if (user_mode(regs)) { |
1341 | struct tty_struct *tty = get_current_tty(); | |
1342 | tty_write_message(tty, buf); | |
1343 | tty_kref_put(tty); | |
1344 | } | |
1da177e4 | 1345 | buf[len-1] = '\0'; /* drop '\r' */ |
d2b176ed JS |
1346 | /* watch for command names containing %s */ |
1347 | printk(KERN_WARNING "%s", buf); | |
1348 | } else { | |
54f8dd3c MS |
1349 | if (no_unaligned_warning) { |
1350 | printk_once(KERN_WARNING "%s(%d) encountered an " | |
d2b176ed JS |
1351 | "unaligned exception which required\n" |
1352 | "kernel assistance, which degrades " | |
1353 | "the performance of the application.\n" | |
1354 | "Unaligned exception warnings have " | |
1355 | "been disabled by the system " | |
1356 | "administrator\n" | |
1357 | "echo 0 > /proc/sys/kernel/ignore-" | |
1358 | "unaligned-usertrap to re-enable\n", | |
19c5870c | 1359 | current->comm, task_pid_nr(current)); |
d2b176ed | 1360 | } |
1da177e4 LT |
1361 | } |
1362 | } else { | |
7683a3f9 | 1363 | if (__ratelimit(&logging_rate_limit)) { |
1da177e4 LT |
1364 | printk(KERN_WARNING "kernel unaligned access to 0x%016lx, ip=0x%016lx\n", |
1365 | ifa, regs->cr_iip + ipsr->ri); | |
88fc241f DC |
1366 | if (unaligned_dump_stack) |
1367 | dump_stack(); | |
1368 | } | |
1da177e4 LT |
1369 | set_fs(KERNEL_DS); |
1370 | } | |
1371 | ||
1372 | DPRINT("iip=%lx ifa=%lx isr=%lx (ei=%d, sp=%d)\n", | |
1373 | regs->cr_iip, ifa, regs->cr_ipsr, ipsr->ri, ipsr->it); | |
1374 | ||
1375 | if (__copy_from_user(bundle, (void __user *) regs->cr_iip, 16)) | |
1376 | goto failure; | |
1377 | ||
1378 | /* | |
1379 | * extract the instruction from the bundle given the slot number | |
1380 | */ | |
1381 | switch (ipsr->ri) { | |
787ca32d | 1382 | default: |
1da177e4 LT |
1383 | case 0: u.l = (bundle[0] >> 5); break; |
1384 | case 1: u.l = (bundle[0] >> 46) | (bundle[1] << 18); break; | |
1385 | case 2: u.l = (bundle[1] >> 23); break; | |
1386 | } | |
1387 | opcode = (u.l >> IA64_OPCODE_SHIFT) & IA64_OPCODE_MASK; | |
1388 | ||
1389 | DPRINT("opcode=%lx ld.qp=%d ld.r1=%d ld.imm=%d ld.r3=%d ld.x=%d ld.hint=%d " | |
1390 | "ld.x6=0x%x ld.m=%d ld.op=%d\n", opcode, u.insn.qp, u.insn.r1, u.insn.imm, | |
1391 | u.insn.r3, u.insn.x, u.insn.hint, u.insn.x6_sz, u.insn.m, u.insn.op); | |
1392 | ||
1393 | /* | |
1394 | * IMPORTANT: | |
1395 | * Notice that the switch statement DOES not cover all possible instructions | |
1396 | * that DO generate unaligned references. This is made on purpose because for some | |
1397 | * instructions it DOES NOT make sense to try and emulate the access. Sometimes it | |
1398 | * is WRONG to try and emulate. Here is a list of instruction we don't emulate i.e., | |
1399 | * the program will get a signal and die: | |
1400 | * | |
1401 | * load/store: | |
1402 | * - ldX.spill | |
1403 | * - stX.spill | |
1404 | * Reason: RNATs are based on addresses | |
1405 | * - ld16 | |
1406 | * - st16 | |
1407 | * Reason: ld16 and st16 are supposed to occur in a single | |
1408 | * memory op | |
1409 | * | |
1410 | * synchronization: | |
1411 | * - cmpxchg | |
1412 | * - fetchadd | |
1413 | * - xchg | |
1414 | * Reason: ATOMIC operations cannot be emulated properly using multiple | |
1415 | * instructions. | |
1416 | * | |
1417 | * speculative loads: | |
1418 | * - ldX.sZ | |
1419 | * Reason: side effects, code must be ready to deal with failure so simpler | |
1420 | * to let the load fail. | |
1421 | * --------------------------------------------------------------------------------- | |
1422 | * XXX fixme | |
1423 | * | |
1424 | * I would like to get rid of this switch case and do something | |
1425 | * more elegant. | |
1426 | */ | |
1427 | switch (opcode) { | |
1428 | case LDS_OP: | |
1429 | case LDSA_OP: | |
1430 | if (u.insn.x) | |
1431 | /* oops, really a semaphore op (cmpxchg, etc) */ | |
1432 | goto failure; | |
1433 | /* no break */ | |
1434 | case LDS_IMM_OP: | |
1435 | case LDSA_IMM_OP: | |
1436 | case LDFS_OP: | |
1437 | case LDFSA_OP: | |
1438 | case LDFS_IMM_OP: | |
1439 | /* | |
1440 | * The instruction will be retried with deferred exceptions turned on, and | |
1441 | * we should get Nat bit installed | |
1442 | * | |
1443 | * IMPORTANT: When PSR_ED is set, the register & immediate update forms | |
1444 | * are actually executed even though the operation failed. So we don't | |
1445 | * need to take care of this. | |
1446 | */ | |
1447 | DPRINT("forcing PSR_ED\n"); | |
1448 | regs->cr_ipsr |= IA64_PSR_ED; | |
1449 | goto done; | |
1450 | ||
1451 | case LD_OP: | |
1452 | case LDA_OP: | |
1453 | case LDBIAS_OP: | |
1454 | case LDACQ_OP: | |
1455 | case LDCCLR_OP: | |
1456 | case LDCNC_OP: | |
1457 | case LDCCLRACQ_OP: | |
1458 | if (u.insn.x) | |
1459 | /* oops, really a semaphore op (cmpxchg, etc) */ | |
1460 | goto failure; | |
1461 | /* no break */ | |
1462 | case LD_IMM_OP: | |
1463 | case LDA_IMM_OP: | |
1464 | case LDBIAS_IMM_OP: | |
1465 | case LDACQ_IMM_OP: | |
1466 | case LDCCLR_IMM_OP: | |
1467 | case LDCNC_IMM_OP: | |
1468 | case LDCCLRACQ_IMM_OP: | |
1469 | ret = emulate_load_int(ifa, u.insn, regs); | |
1470 | break; | |
1471 | ||
1472 | case ST_OP: | |
1473 | case STREL_OP: | |
1474 | if (u.insn.x) | |
1475 | /* oops, really a semaphore op (cmpxchg, etc) */ | |
1476 | goto failure; | |
1477 | /* no break */ | |
1478 | case ST_IMM_OP: | |
1479 | case STREL_IMM_OP: | |
1480 | ret = emulate_store_int(ifa, u.insn, regs); | |
1481 | break; | |
1482 | ||
1483 | case LDF_OP: | |
1484 | case LDFA_OP: | |
1485 | case LDFCCLR_OP: | |
1486 | case LDFCNC_OP: | |
1da177e4 LT |
1487 | if (u.insn.x) |
1488 | ret = emulate_load_floatpair(ifa, u.insn, regs); | |
1489 | else | |
1490 | ret = emulate_load_float(ifa, u.insn, regs); | |
1491 | break; | |
1492 | ||
1a499150 TL |
1493 | case LDF_IMM_OP: |
1494 | case LDFA_IMM_OP: | |
1495 | case LDFCCLR_IMM_OP: | |
1496 | case LDFCNC_IMM_OP: | |
1497 | ret = emulate_load_float(ifa, u.insn, regs); | |
1498 | break; | |
1499 | ||
1da177e4 LT |
1500 | case STF_OP: |
1501 | case STF_IMM_OP: | |
1502 | ret = emulate_store_float(ifa, u.insn, regs); | |
1503 | break; | |
1504 | ||
1505 | default: | |
1506 | goto failure; | |
1507 | } | |
1508 | DPRINT("ret=%d\n", ret); | |
1509 | if (ret) | |
1510 | goto failure; | |
1511 | ||
1512 | if (ipsr->ri == 2) | |
1513 | /* | |
1514 | * given today's architecture this case is not likely to happen because a | |
1515 | * memory access instruction (M) can never be in the last slot of a | |
1516 | * bundle. But let's keep it for now. | |
1517 | */ | |
1518 | regs->cr_iip += 16; | |
1519 | ipsr->ri = (ipsr->ri + 1) & 0x3; | |
1520 | ||
1521 | DPRINT("ipsr->ri=%d iip=%lx\n", ipsr->ri, regs->cr_iip); | |
1522 | done: | |
1523 | set_fs(old_fs); /* restore original address limit */ | |
1524 | return; | |
1525 | ||
1526 | failure: | |
1527 | /* something went wrong... */ | |
1528 | if (!user_mode(regs)) { | |
1529 | if (eh) { | |
1530 | ia64_handle_exception(regs, eh); | |
1531 | goto done; | |
1532 | } | |
620de2f5 JB |
1533 | if (die_if_kernel("error during unaligned kernel access\n", regs, ret)) |
1534 | return; | |
1da177e4 LT |
1535 | /* NOT_REACHED */ |
1536 | } | |
1537 | force_sigbus: | |
1538 | si.si_signo = SIGBUS; | |
1539 | si.si_errno = 0; | |
1540 | si.si_code = BUS_ADRALN; | |
1541 | si.si_addr = (void __user *) ifa; | |
1542 | si.si_flags = 0; | |
1543 | si.si_isr = 0; | |
1544 | si.si_imm = 0; | |
1545 | force_sig_info(SIGBUS, &si, current); | |
1546 | goto done; | |
1547 | } |