xref: /openbmc/linux/arch/ia64/kernel/unaligned.c (revision 65844828)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Architecture-specific unaligned trap handling.
4  *
5  * Copyright (C) 1999-2002, 2004 Hewlett-Packard Co
6  *	Stephane Eranian <eranian@hpl.hp.com>
7  *	David Mosberger-Tang <davidm@hpl.hp.com>
8  *
9  * 2002/12/09   Fix rotating register handling (off-by-1 error, missing fr-rotation).  Fix
10  *		get_rse_reg() to not leak kernel bits to user-level (reading an out-of-frame
11  *		stacked register returns an undefined value; it does NOT trigger a
12  *		"rsvd register fault").
13  * 2001/10/11	Fix unaligned access to rotating registers in s/w pipelined loops.
14  * 2001/08/13	Correct size of extended floats (float_fsz) from 16 to 10 bytes.
15  * 2001/01/17	Add support emulation of unaligned kernel accesses.
16  */
17 #include <linux/jiffies.h>
18 #include <linux/kernel.h>
19 #include <linux/sched/signal.h>
20 #include <linux/tty.h>
21 #include <linux/extable.h>
22 #include <linux/ratelimit.h>
23 #include <linux/uaccess.h>
24 
25 #include <asm/intrinsics.h>
26 #include <asm/processor.h>
27 #include <asm/rse.h>
28 #include <asm/exception.h>
29 #include <asm/unaligned.h>
30 
31 extern int die_if_kernel(char *str, struct pt_regs *regs, long err);
32 
33 #undef DEBUG_UNALIGNED_TRAP
34 
35 #ifdef DEBUG_UNALIGNED_TRAP
36 # define DPRINT(a...)	do { printk("%s %u: ", __func__, __LINE__); printk (a); } while (0)
37 # define DDUMP(str,vp,len)	dump(str, vp, len)
38 
39 static void
40 dump (const char *str, void *vp, size_t len)
41 {
42 	unsigned char *cp = vp;
43 	int i;
44 
45 	printk("%s", str);
46 	for (i = 0; i < len; ++i)
47 		printk (" %02x", *cp++);
48 	printk("\n");
49 }
50 #else
51 # define DPRINT(a...)
52 # define DDUMP(str,vp,len)
53 #endif
54 
55 #define IA64_FIRST_STACKED_GR	32
56 #define IA64_FIRST_ROTATING_FR	32
57 #define SIGN_EXT9		0xffffffffffffff00ul
58 
59 /*
60  *  sysctl settable hook which tells the kernel whether to honor the
61  *  IA64_THREAD_UAC_NOPRINT prctl.  Because this is user settable, we want
62  *  to allow the super user to enable/disable this for security reasons
63  *  (i.e. don't allow attacker to fill up logs with unaligned accesses).
64  */
65 int no_unaligned_warning;
66 int unaligned_dump_stack;
67 
68 /*
69  * For M-unit:
70  *
71  *  opcode |   m  |   x6    |
72  * --------|------|---------|
73  * [40-37] | [36] | [35:30] |
74  * --------|------|---------|
75  *     4   |   1  |    6    | = 11 bits
76  * --------------------------
77  * However bits [31:30] are not directly useful to distinguish between
78  * load/store so we can use [35:32] instead, which gives the following
79  * mask ([40:32]) using 9 bits. The 'e' comes from the fact that we defer
80  * checking the m-bit until later in the load/store emulation.
81  */
82 #define IA64_OPCODE_MASK	0x1ef
83 #define IA64_OPCODE_SHIFT	32
84 
85 /*
86  * Table C-28 Integer Load/Store
87  *
88  * We ignore [35:32]= 0x6, 0x7, 0xE, 0xF
89  *
90  * ld8.fill, st8.fill  MUST be aligned because the RNATs are based on
91  * the address (bits [8:3]), so we must failed.
92  */
93 #define LD_OP            0x080
94 #define LDS_OP           0x081
95 #define LDA_OP           0x082
96 #define LDSA_OP          0x083
97 #define LDBIAS_OP        0x084
98 #define LDACQ_OP         0x085
99 /* 0x086, 0x087 are not relevant */
100 #define LDCCLR_OP        0x088
101 #define LDCNC_OP         0x089
102 #define LDCCLRACQ_OP     0x08a
103 #define ST_OP            0x08c
104 #define STREL_OP         0x08d
105 /* 0x08e,0x8f are not relevant */
106 
107 /*
108  * Table C-29 Integer Load +Reg
109  *
110  * we use the ld->m (bit [36:36]) field to determine whether or not we have
111  * a load/store of this form.
112  */
113 
114 /*
115  * Table C-30 Integer Load/Store +Imm
116  *
117  * We ignore [35:32]= 0x6, 0x7, 0xE, 0xF
118  *
119  * ld8.fill, st8.fill  must be aligned because the Nat register are based on
120  * the address, so we must fail and the program must be fixed.
121  */
122 #define LD_IMM_OP            0x0a0
123 #define LDS_IMM_OP           0x0a1
124 #define LDA_IMM_OP           0x0a2
125 #define LDSA_IMM_OP          0x0a3
126 #define LDBIAS_IMM_OP        0x0a4
127 #define LDACQ_IMM_OP         0x0a5
128 /* 0x0a6, 0xa7 are not relevant */
129 #define LDCCLR_IMM_OP        0x0a8
130 #define LDCNC_IMM_OP         0x0a9
131 #define LDCCLRACQ_IMM_OP     0x0aa
132 #define ST_IMM_OP            0x0ac
133 #define STREL_IMM_OP         0x0ad
134 /* 0x0ae,0xaf are not relevant */
135 
136 /*
137  * Table C-32 Floating-point Load/Store
138  */
139 #define LDF_OP           0x0c0
140 #define LDFS_OP          0x0c1
141 #define LDFA_OP          0x0c2
142 #define LDFSA_OP         0x0c3
143 /* 0x0c6 is irrelevant */
144 #define LDFCCLR_OP       0x0c8
145 #define LDFCNC_OP        0x0c9
146 /* 0x0cb is irrelevant  */
147 #define STF_OP           0x0cc
148 
149 /*
150  * Table C-33 Floating-point Load +Reg
151  *
152  * we use the ld->m (bit [36:36]) field to determine whether or not we have
153  * a load/store of this form.
154  */
155 
156 /*
157  * Table C-34 Floating-point Load/Store +Imm
158  */
159 #define LDF_IMM_OP       0x0e0
160 #define LDFS_IMM_OP      0x0e1
161 #define LDFA_IMM_OP      0x0e2
162 #define LDFSA_IMM_OP     0x0e3
163 /* 0x0e6 is irrelevant */
164 #define LDFCCLR_IMM_OP   0x0e8
165 #define LDFCNC_IMM_OP    0x0e9
166 #define STF_IMM_OP       0x0ec
167 
168 typedef struct {
169 	unsigned long	 qp:6;	/* [0:5]   */
170 	unsigned long    r1:7;	/* [6:12]  */
171 	unsigned long   imm:7;	/* [13:19] */
172 	unsigned long    r3:7;	/* [20:26] */
173 	unsigned long     x:1;  /* [27:27] */
174 	unsigned long  hint:2;	/* [28:29] */
175 	unsigned long x6_sz:2;	/* [30:31] */
176 	unsigned long x6_op:4;	/* [32:35], x6 = x6_sz|x6_op */
177 	unsigned long     m:1;	/* [36:36] */
178 	unsigned long    op:4;	/* [37:40] */
179 	unsigned long   pad:23; /* [41:63] */
180 } load_store_t;
181 
182 
183 typedef enum {
184 	UPD_IMMEDIATE,	/* ldXZ r1=[r3],imm(9) */
185 	UPD_REG		/* ldXZ r1=[r3],r2     */
186 } update_t;
187 
188 /*
189  * We use tables to keep track of the offsets of registers in the saved state.
190  * This way we save having big switch/case statements.
191  *
192  * We use bit 0 to indicate switch_stack or pt_regs.
193  * The offset is simply shifted by 1 bit.
194  * A 2-byte value should be enough to hold any kind of offset
195  *
196  * In case the calling convention changes (and thus pt_regs/switch_stack)
197  * simply use RSW instead of RPT or vice-versa.
198  */
199 
200 #define RPO(x)	((size_t) &((struct pt_regs *)0)->x)
201 #define RSO(x)	((size_t) &((struct switch_stack *)0)->x)
202 
203 #define RPT(x)		(RPO(x) << 1)
204 #define RSW(x)		(1| RSO(x)<<1)
205 
206 #define GR_OFFS(x)	(gr_info[x]>>1)
207 #define GR_IN_SW(x)	(gr_info[x] & 0x1)
208 
209 #define FR_OFFS(x)	(fr_info[x]>>1)
210 #define FR_IN_SW(x)	(fr_info[x] & 0x1)
211 
212 static u16 gr_info[32]={
213 	0,			/* r0 is read-only : WE SHOULD NEVER GET THIS */
214 
215 	RPT(r1), RPT(r2), RPT(r3),
216 
217 	RSW(r4), RSW(r5), RSW(r6), RSW(r7),
218 
219 	RPT(r8), RPT(r9), RPT(r10), RPT(r11),
220 	RPT(r12), RPT(r13), RPT(r14), RPT(r15),
221 
222 	RPT(r16), RPT(r17), RPT(r18), RPT(r19),
223 	RPT(r20), RPT(r21), RPT(r22), RPT(r23),
224 	RPT(r24), RPT(r25), RPT(r26), RPT(r27),
225 	RPT(r28), RPT(r29), RPT(r30), RPT(r31)
226 };
227 
228 static u16 fr_info[32]={
229 	0,			/* constant : WE SHOULD NEVER GET THIS */
230 	0,			/* constant : WE SHOULD NEVER GET THIS */
231 
232 	RSW(f2), RSW(f3), RSW(f4), RSW(f5),
233 
234 	RPT(f6), RPT(f7), RPT(f8), RPT(f9),
235 	RPT(f10), RPT(f11),
236 
237 	RSW(f12), RSW(f13), RSW(f14),
238 	RSW(f15), RSW(f16), RSW(f17), RSW(f18), RSW(f19),
239 	RSW(f20), RSW(f21), RSW(f22), RSW(f23), RSW(f24),
240 	RSW(f25), RSW(f26), RSW(f27), RSW(f28), RSW(f29),
241 	RSW(f30), RSW(f31)
242 };
243 
244 /* Invalidate ALAT entry for integer register REGNO.  */
245 static void
246 invala_gr (int regno)
247 {
248 #	define F(reg)	case reg: ia64_invala_gr(reg); break
249 
250 	switch (regno) {
251 		F(  0); F(  1); F(  2); F(  3); F(  4); F(  5); F(  6); F(  7);
252 		F(  8); F(  9); F( 10); F( 11); F( 12); F( 13); F( 14); F( 15);
253 		F( 16); F( 17); F( 18); F( 19); F( 20); F( 21); F( 22); F( 23);
254 		F( 24); F( 25); F( 26); F( 27); F( 28); F( 29); F( 30); F( 31);
255 		F( 32); F( 33); F( 34); F( 35); F( 36); F( 37); F( 38); F( 39);
256 		F( 40); F( 41); F( 42); F( 43); F( 44); F( 45); F( 46); F( 47);
257 		F( 48); F( 49); F( 50); F( 51); F( 52); F( 53); F( 54); F( 55);
258 		F( 56); F( 57); F( 58); F( 59); F( 60); F( 61); F( 62); F( 63);
259 		F( 64); F( 65); F( 66); F( 67); F( 68); F( 69); F( 70); F( 71);
260 		F( 72); F( 73); F( 74); F( 75); F( 76); F( 77); F( 78); F( 79);
261 		F( 80); F( 81); F( 82); F( 83); F( 84); F( 85); F( 86); F( 87);
262 		F( 88); F( 89); F( 90); F( 91); F( 92); F( 93); F( 94); F( 95);
263 		F( 96); F( 97); F( 98); F( 99); F(100); F(101); F(102); F(103);
264 		F(104); F(105); F(106); F(107); F(108); F(109); F(110); F(111);
265 		F(112); F(113); F(114); F(115); F(116); F(117); F(118); F(119);
266 		F(120); F(121); F(122); F(123); F(124); F(125); F(126); F(127);
267 	}
268 #	undef F
269 }
270 
271 /* Invalidate ALAT entry for floating-point register REGNO.  */
272 static void
273 invala_fr (int regno)
274 {
275 #	define F(reg)	case reg: ia64_invala_fr(reg); break
276 
277 	switch (regno) {
278 		F(  0); F(  1); F(  2); F(  3); F(  4); F(  5); F(  6); F(  7);
279 		F(  8); F(  9); F( 10); F( 11); F( 12); F( 13); F( 14); F( 15);
280 		F( 16); F( 17); F( 18); F( 19); F( 20); F( 21); F( 22); F( 23);
281 		F( 24); F( 25); F( 26); F( 27); F( 28); F( 29); F( 30); F( 31);
282 		F( 32); F( 33); F( 34); F( 35); F( 36); F( 37); F( 38); F( 39);
283 		F( 40); F( 41); F( 42); F( 43); F( 44); F( 45); F( 46); F( 47);
284 		F( 48); F( 49); F( 50); F( 51); F( 52); F( 53); F( 54); F( 55);
285 		F( 56); F( 57); F( 58); F( 59); F( 60); F( 61); F( 62); F( 63);
286 		F( 64); F( 65); F( 66); F( 67); F( 68); F( 69); F( 70); F( 71);
287 		F( 72); F( 73); F( 74); F( 75); F( 76); F( 77); F( 78); F( 79);
288 		F( 80); F( 81); F( 82); F( 83); F( 84); F( 85); F( 86); F( 87);
289 		F( 88); F( 89); F( 90); F( 91); F( 92); F( 93); F( 94); F( 95);
290 		F( 96); F( 97); F( 98); F( 99); F(100); F(101); F(102); F(103);
291 		F(104); F(105); F(106); F(107); F(108); F(109); F(110); F(111);
292 		F(112); F(113); F(114); F(115); F(116); F(117); F(118); F(119);
293 		F(120); F(121); F(122); F(123); F(124); F(125); F(126); F(127);
294 	}
295 #	undef F
296 }
297 
298 static inline unsigned long
299 rotate_reg (unsigned long sor, unsigned long rrb, unsigned long reg)
300 {
301 	reg += rrb;
302 	if (reg >= sor)
303 		reg -= sor;
304 	return reg;
305 }
306 
307 static void
308 set_rse_reg (struct pt_regs *regs, unsigned long r1, unsigned long val, int nat)
309 {
310 	struct switch_stack *sw = (struct switch_stack *) regs - 1;
311 	unsigned long *bsp, *bspstore, *addr, *rnat_addr, *ubs_end;
312 	unsigned long *kbs = (void *) current + IA64_RBS_OFFSET;
313 	unsigned long rnats, nat_mask;
314 	unsigned long on_kbs;
315 	long sof = (regs->cr_ifs) & 0x7f;
316 	long sor = 8 * ((regs->cr_ifs >> 14) & 0xf);
317 	long rrb_gr = (regs->cr_ifs >> 18) & 0x7f;
318 	long ridx = r1 - 32;
319 
320 	if (ridx >= sof) {
321 		/* this should never happen, as the "rsvd register fault" has higher priority */
322 		DPRINT("ignoring write to r%lu; only %lu registers are allocated!\n", r1, sof);
323 		return;
324 	}
325 
326 	if (ridx < sor)
327 		ridx = rotate_reg(sor, rrb_gr, ridx);
328 
329 	DPRINT("r%lu, sw.bspstore=%lx pt.bspstore=%lx sof=%ld sol=%ld ridx=%ld\n",
330 	       r1, sw->ar_bspstore, regs->ar_bspstore, sof, (regs->cr_ifs >> 7) & 0x7f, ridx);
331 
332 	on_kbs = ia64_rse_num_regs(kbs, (unsigned long *) sw->ar_bspstore);
333 	addr = ia64_rse_skip_regs((unsigned long *) sw->ar_bspstore, -sof + ridx);
334 	if (addr >= kbs) {
335 		/* the register is on the kernel backing store: easy... */
336 		rnat_addr = ia64_rse_rnat_addr(addr);
337 		if ((unsigned long) rnat_addr >= sw->ar_bspstore)
338 			rnat_addr = &sw->ar_rnat;
339 		nat_mask = 1UL << ia64_rse_slot_num(addr);
340 
341 		*addr = val;
342 		if (nat)
343 			*rnat_addr |=  nat_mask;
344 		else
345 			*rnat_addr &= ~nat_mask;
346 		return;
347 	}
348 
349 	if (!user_stack(current, regs)) {
350 		DPRINT("ignoring kernel write to r%lu; register isn't on the kernel RBS!", r1);
351 		return;
352 	}
353 
354 	bspstore = (unsigned long *)regs->ar_bspstore;
355 	ubs_end = ia64_rse_skip_regs(bspstore, on_kbs);
356 	bsp     = ia64_rse_skip_regs(ubs_end, -sof);
357 	addr    = ia64_rse_skip_regs(bsp, ridx);
358 
359 	DPRINT("ubs_end=%p bsp=%p addr=%p\n", (void *) ubs_end, (void *) bsp, (void *) addr);
360 
361 	ia64_poke(current, sw, (unsigned long) ubs_end, (unsigned long) addr, val);
362 
363 	rnat_addr = ia64_rse_rnat_addr(addr);
364 
365 	ia64_peek(current, sw, (unsigned long) ubs_end, (unsigned long) rnat_addr, &rnats);
366 	DPRINT("rnat @%p = 0x%lx nat=%d old nat=%ld\n",
367 	       (void *) rnat_addr, rnats, nat, (rnats >> ia64_rse_slot_num(addr)) & 1);
368 
369 	nat_mask = 1UL << ia64_rse_slot_num(addr);
370 	if (nat)
371 		rnats |=  nat_mask;
372 	else
373 		rnats &= ~nat_mask;
374 	ia64_poke(current, sw, (unsigned long) ubs_end, (unsigned long) rnat_addr, rnats);
375 
376 	DPRINT("rnat changed to @%p = 0x%lx\n", (void *) rnat_addr, rnats);
377 }
378 
379 
380 static void
381 get_rse_reg (struct pt_regs *regs, unsigned long r1, unsigned long *val, int *nat)
382 {
383 	struct switch_stack *sw = (struct switch_stack *) regs - 1;
384 	unsigned long *bsp, *addr, *rnat_addr, *ubs_end, *bspstore;
385 	unsigned long *kbs = (void *) current + IA64_RBS_OFFSET;
386 	unsigned long rnats, nat_mask;
387 	unsigned long on_kbs;
388 	long sof = (regs->cr_ifs) & 0x7f;
389 	long sor = 8 * ((regs->cr_ifs >> 14) & 0xf);
390 	long rrb_gr = (regs->cr_ifs >> 18) & 0x7f;
391 	long ridx = r1 - 32;
392 
393 	if (ridx >= sof) {
394 		/* read of out-of-frame register returns an undefined value; 0 in our case.  */
395 		DPRINT("ignoring read from r%lu; only %lu registers are allocated!\n", r1, sof);
396 		goto fail;
397 	}
398 
399 	if (ridx < sor)
400 		ridx = rotate_reg(sor, rrb_gr, ridx);
401 
402 	DPRINT("r%lu, sw.bspstore=%lx pt.bspstore=%lx sof=%ld sol=%ld ridx=%ld\n",
403 	       r1, sw->ar_bspstore, regs->ar_bspstore, sof, (regs->cr_ifs >> 7) & 0x7f, ridx);
404 
405 	on_kbs = ia64_rse_num_regs(kbs, (unsigned long *) sw->ar_bspstore);
406 	addr = ia64_rse_skip_regs((unsigned long *) sw->ar_bspstore, -sof + ridx);
407 	if (addr >= kbs) {
408 		/* the register is on the kernel backing store: easy... */
409 		*val = *addr;
410 		if (nat) {
411 			rnat_addr = ia64_rse_rnat_addr(addr);
412 			if ((unsigned long) rnat_addr >= sw->ar_bspstore)
413 				rnat_addr = &sw->ar_rnat;
414 			nat_mask = 1UL << ia64_rse_slot_num(addr);
415 			*nat = (*rnat_addr & nat_mask) != 0;
416 		}
417 		return;
418 	}
419 
420 	if (!user_stack(current, regs)) {
421 		DPRINT("ignoring kernel read of r%lu; register isn't on the RBS!", r1);
422 		goto fail;
423 	}
424 
425 	bspstore = (unsigned long *)regs->ar_bspstore;
426 	ubs_end = ia64_rse_skip_regs(bspstore, on_kbs);
427 	bsp     = ia64_rse_skip_regs(ubs_end, -sof);
428 	addr    = ia64_rse_skip_regs(bsp, ridx);
429 
430 	DPRINT("ubs_end=%p bsp=%p addr=%p\n", (void *) ubs_end, (void *) bsp, (void *) addr);
431 
432 	ia64_peek(current, sw, (unsigned long) ubs_end, (unsigned long) addr, val);
433 
434 	if (nat) {
435 		rnat_addr = ia64_rse_rnat_addr(addr);
436 		nat_mask = 1UL << ia64_rse_slot_num(addr);
437 
438 		DPRINT("rnat @%p = 0x%lx\n", (void *) rnat_addr, rnats);
439 
440 		ia64_peek(current, sw, (unsigned long) ubs_end, (unsigned long) rnat_addr, &rnats);
441 		*nat = (rnats & nat_mask) != 0;
442 	}
443 	return;
444 
445   fail:
446 	*val = 0;
447 	if (nat)
448 		*nat = 0;
449 	return;
450 }
451 
452 
453 static void
454 setreg (unsigned long regnum, unsigned long val, int nat, struct pt_regs *regs)
455 {
456 	struct switch_stack *sw = (struct switch_stack *) regs - 1;
457 	unsigned long addr;
458 	unsigned long bitmask;
459 	unsigned long *unat;
460 
461 	/*
462 	 * First takes care of stacked registers
463 	 */
464 	if (regnum >= IA64_FIRST_STACKED_GR) {
465 		set_rse_reg(regs, regnum, val, nat);
466 		return;
467 	}
468 
469 	/*
470 	 * Using r0 as a target raises a General Exception fault which has higher priority
471 	 * than the Unaligned Reference fault.
472 	 */
473 
474 	/*
475 	 * Now look at registers in [0-31] range and init correct UNAT
476 	 */
477 	if (GR_IN_SW(regnum)) {
478 		addr = (unsigned long)sw;
479 		unat = &sw->ar_unat;
480 	} else {
481 		addr = (unsigned long)regs;
482 		unat = &sw->caller_unat;
483 	}
484 	DPRINT("tmp_base=%lx switch_stack=%s offset=%d\n",
485 	       addr, unat==&sw->ar_unat ? "yes":"no", GR_OFFS(regnum));
486 	/*
487 	 * add offset from base of struct
488 	 * and do it !
489 	 */
490 	addr += GR_OFFS(regnum);
491 
492 	*(unsigned long *)addr = val;
493 
494 	/*
495 	 * We need to clear the corresponding UNAT bit to fully emulate the load
496 	 * UNAT bit_pos = GR[r3]{8:3} form EAS-2.4
497 	 */
498 	bitmask   = 1UL << (addr >> 3 & 0x3f);
499 	DPRINT("*0x%lx=0x%lx NaT=%d prev_unat @%p=%lx\n", addr, val, nat, (void *) unat, *unat);
500 	if (nat) {
501 		*unat |= bitmask;
502 	} else {
503 		*unat &= ~bitmask;
504 	}
505 	DPRINT("*0x%lx=0x%lx NaT=%d new unat: %p=%lx\n", addr, val, nat, (void *) unat,*unat);
506 }
507 
508 /*
509  * Return the (rotated) index for floating point register REGNUM (REGNUM must be in the
510  * range from 32-127, result is in the range from 0-95.
511  */
512 static inline unsigned long
513 fph_index (struct pt_regs *regs, long regnum)
514 {
515 	unsigned long rrb_fr = (regs->cr_ifs >> 25) & 0x7f;
516 	return rotate_reg(96, rrb_fr, (regnum - IA64_FIRST_ROTATING_FR));
517 }
518 
519 static void
520 setfpreg (unsigned long regnum, struct ia64_fpreg *fpval, struct pt_regs *regs)
521 {
522 	struct switch_stack *sw = (struct switch_stack *)regs - 1;
523 	unsigned long addr;
524 
525 	/*
526 	 * From EAS-2.5: FPDisableFault has higher priority than Unaligned
527 	 * Fault. Thus, when we get here, we know the partition is enabled.
528 	 * To update f32-f127, there are three choices:
529 	 *
530 	 *	(1) save f32-f127 to thread.fph and update the values there
531 	 *	(2) use a gigantic switch statement to directly access the registers
532 	 *	(3) generate code on the fly to update the desired register
533 	 *
534 	 * For now, we are using approach (1).
535 	 */
536 	if (regnum >= IA64_FIRST_ROTATING_FR) {
537 		ia64_sync_fph(current);
538 		current->thread.fph[fph_index(regs, regnum)] = *fpval;
539 	} else {
540 		/*
541 		 * pt_regs or switch_stack ?
542 		 */
543 		if (FR_IN_SW(regnum)) {
544 			addr = (unsigned long)sw;
545 		} else {
546 			addr = (unsigned long)regs;
547 		}
548 
549 		DPRINT("tmp_base=%lx offset=%d\n", addr, FR_OFFS(regnum));
550 
551 		addr += FR_OFFS(regnum);
552 		*(struct ia64_fpreg *)addr = *fpval;
553 
554 		/*
555 		 * mark the low partition as being used now
556 		 *
557 		 * It is highly unlikely that this bit is not already set, but
558 		 * let's do it for safety.
559 		 */
560 		regs->cr_ipsr |= IA64_PSR_MFL;
561 	}
562 }
563 
564 /*
565  * Those 2 inline functions generate the spilled versions of the constant floating point
566  * registers which can be used with stfX
567  */
568 static inline void
569 float_spill_f0 (struct ia64_fpreg *final)
570 {
571 	ia64_stf_spill(final, 0);
572 }
573 
574 static inline void
575 float_spill_f1 (struct ia64_fpreg *final)
576 {
577 	ia64_stf_spill(final, 1);
578 }
579 
580 static void
581 getfpreg (unsigned long regnum, struct ia64_fpreg *fpval, struct pt_regs *regs)
582 {
583 	struct switch_stack *sw = (struct switch_stack *) regs - 1;
584 	unsigned long addr;
585 
586 	/*
587 	 * From EAS-2.5: FPDisableFault has higher priority than
588 	 * Unaligned Fault. Thus, when we get here, we know the partition is
589 	 * enabled.
590 	 *
591 	 * When regnum > 31, the register is still live and we need to force a save
592 	 * to current->thread.fph to get access to it.  See discussion in setfpreg()
593 	 * for reasons and other ways of doing this.
594 	 */
595 	if (regnum >= IA64_FIRST_ROTATING_FR) {
596 		ia64_flush_fph(current);
597 		*fpval = current->thread.fph[fph_index(regs, regnum)];
598 	} else {
599 		/*
600 		 * f0 = 0.0, f1= 1.0. Those registers are constant and are thus
601 		 * not saved, we must generate their spilled form on the fly
602 		 */
603 		switch(regnum) {
604 		case 0:
605 			float_spill_f0(fpval);
606 			break;
607 		case 1:
608 			float_spill_f1(fpval);
609 			break;
610 		default:
611 			/*
612 			 * pt_regs or switch_stack ?
613 			 */
614 			addr =  FR_IN_SW(regnum) ? (unsigned long)sw
615 						 : (unsigned long)regs;
616 
617 			DPRINT("is_sw=%d tmp_base=%lx offset=0x%x\n",
618 			       FR_IN_SW(regnum), addr, FR_OFFS(regnum));
619 
620 			addr  += FR_OFFS(regnum);
621 			*fpval = *(struct ia64_fpreg *)addr;
622 		}
623 	}
624 }
625 
626 
627 static void
628 getreg (unsigned long regnum, unsigned long *val, int *nat, struct pt_regs *regs)
629 {
630 	struct switch_stack *sw = (struct switch_stack *) regs - 1;
631 	unsigned long addr, *unat;
632 
633 	if (regnum >= IA64_FIRST_STACKED_GR) {
634 		get_rse_reg(regs, regnum, val, nat);
635 		return;
636 	}
637 
638 	/*
639 	 * take care of r0 (read-only always evaluate to 0)
640 	 */
641 	if (regnum == 0) {
642 		*val = 0;
643 		if (nat)
644 			*nat = 0;
645 		return;
646 	}
647 
648 	/*
649 	 * Now look at registers in [0-31] range and init correct UNAT
650 	 */
651 	if (GR_IN_SW(regnum)) {
652 		addr = (unsigned long)sw;
653 		unat = &sw->ar_unat;
654 	} else {
655 		addr = (unsigned long)regs;
656 		unat = &sw->caller_unat;
657 	}
658 
659 	DPRINT("addr_base=%lx offset=0x%x\n", addr,  GR_OFFS(regnum));
660 
661 	addr += GR_OFFS(regnum);
662 
663 	*val  = *(unsigned long *)addr;
664 
665 	/*
666 	 * do it only when requested
667 	 */
668 	if (nat)
669 		*nat  = (*unat >> (addr >> 3 & 0x3f)) & 0x1UL;
670 }
671 
672 static void
673 emulate_load_updates (update_t type, load_store_t ld, struct pt_regs *regs, unsigned long ifa)
674 {
675 	/*
676 	 * IMPORTANT:
677 	 * Given the way we handle unaligned speculative loads, we should
678 	 * not get to this point in the code but we keep this sanity check,
679 	 * just in case.
680 	 */
681 	if (ld.x6_op == 1 || ld.x6_op == 3) {
682 		printk(KERN_ERR "%s: register update on speculative load, error\n", __func__);
683 		if (die_if_kernel("unaligned reference on speculative load with register update\n",
684 				  regs, 30))
685 			return;
686 	}
687 
688 
689 	/*
690 	 * at this point, we know that the base register to update is valid i.e.,
691 	 * it's not r0
692 	 */
693 	if (type == UPD_IMMEDIATE) {
694 		unsigned long imm;
695 
696 		/*
697 		 * Load +Imm: ldXZ r1=[r3],imm(9)
698 		 *
699 		 *
700 		 * form imm9: [13:19] contain the first 7 bits
701 		 */
702 		imm = ld.x << 7 | ld.imm;
703 
704 		/*
705 		 * sign extend (1+8bits) if m set
706 		 */
707 		if (ld.m) imm |= SIGN_EXT9;
708 
709 		/*
710 		 * ifa == r3 and we know that the NaT bit on r3 was clear so
711 		 * we can directly use ifa.
712 		 */
713 		ifa += imm;
714 
715 		setreg(ld.r3, ifa, 0, regs);
716 
717 		DPRINT("ld.x=%d ld.m=%d imm=%ld r3=0x%lx\n", ld.x, ld.m, imm, ifa);
718 
719 	} else if (ld.m) {
720 		unsigned long r2;
721 		int nat_r2;
722 
723 		/*
724 		 * Load +Reg Opcode: ldXZ r1=[r3],r2
725 		 *
726 		 * Note: that we update r3 even in the case of ldfX.a
727 		 * (where the load does not happen)
728 		 *
729 		 * The way the load algorithm works, we know that r3 does not
730 		 * have its NaT bit set (would have gotten NaT consumption
731 		 * before getting the unaligned fault). So we can use ifa
732 		 * which equals r3 at this point.
733 		 *
734 		 * IMPORTANT:
735 		 * The above statement holds ONLY because we know that we
736 		 * never reach this code when trying to do a ldX.s.
737 		 * If we ever make it to here on an ldfX.s then
738 		 */
739 		getreg(ld.imm, &r2, &nat_r2, regs);
740 
741 		ifa += r2;
742 
743 		/*
744 		 * propagate Nat r2 -> r3
745 		 */
746 		setreg(ld.r3, ifa, nat_r2, regs);
747 
748 		DPRINT("imm=%d r2=%ld r3=0x%lx nat_r2=%d\n",ld.imm, r2, ifa, nat_r2);
749 	}
750 }
751 
752 static int emulate_store(unsigned long ifa, void *val, int len, bool kernel_mode)
753 {
754 	if (kernel_mode)
755 		return copy_to_kernel_nofault((void *)ifa, val, len);
756 
757 	return copy_to_user((void __user *)ifa, val, len);
758 }
759 
760 static int emulate_load(void *val, unsigned long ifa, int len, bool kernel_mode)
761 {
762 	if (kernel_mode)
763 	       return copy_from_kernel_nofault(val, (void *)ifa, len);
764 
765 	return copy_from_user(val, (void __user *)ifa, len);
766 }
767 
768 static int
769 emulate_load_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs,
770 		  bool kernel_mode)
771 {
772 	unsigned int len = 1 << ld.x6_sz;
773 	unsigned long val = 0;
774 
775 	/*
776 	 * r0, as target, doesn't need to be checked because Illegal Instruction
777 	 * faults have higher priority than unaligned faults.
778 	 *
779 	 * r0 cannot be found as the base as it would never generate an
780 	 * unaligned reference.
781 	 */
782 
783 	/*
784 	 * ldX.a we will emulate load and also invalidate the ALAT entry.
785 	 * See comment below for explanation on how we handle ldX.a
786 	 */
787 
788 	if (len != 2 && len != 4 && len != 8) {
789 		DPRINT("unknown size: x6=%d\n", ld.x6_sz);
790 		return -1;
791 	}
792 	/* this assumes little-endian byte-order: */
793 	if (emulate_load(&val, ifa, len, kernel_mode))
794 		return -1;
795 	setreg(ld.r1, val, 0, regs);
796 
797 	/*
798 	 * check for updates on any kind of loads
799 	 */
800 	if (ld.op == 0x5 || ld.m)
801 		emulate_load_updates(ld.op == 0x5 ? UPD_IMMEDIATE: UPD_REG, ld, regs, ifa);
802 
803 	/*
804 	 * handling of various loads (based on EAS2.4):
805 	 *
806 	 * ldX.acq (ordered load):
807 	 *	- acquire semantics would have been used, so force fence instead.
808 	 *
809 	 * ldX.c.clr (check load and clear):
810 	 *	- if we get to this handler, it's because the entry was not in the ALAT.
811 	 *	  Therefore the operation reverts to a normal load
812 	 *
813 	 * ldX.c.nc (check load no clear):
814 	 *	- same as previous one
815 	 *
816 	 * ldX.c.clr.acq (ordered check load and clear):
817 	 *	- same as above for c.clr part. The load needs to have acquire semantics. So
818 	 *	  we use the fence semantics which is stronger and thus ensures correctness.
819 	 *
820 	 * ldX.a (advanced load):
821 	 *	- suppose ldX.a r1=[r3]. If we get to the unaligned trap it's because the
822 	 *	  address doesn't match requested size alignment. This means that we would
823 	 *	  possibly need more than one load to get the result.
824 	 *
825 	 *	  The load part can be handled just like a normal load, however the difficult
826 	 *	  part is to get the right thing into the ALAT. The critical piece of information
827 	 *	  in the base address of the load & size. To do that, a ld.a must be executed,
828 	 *	  clearly any address can be pushed into the table by using ld1.a r1=[r3]. Now
829 	 *	  if we use the same target register, we will be okay for the check.a instruction.
830 	 *	  If we look at the store, basically a stX [r3]=r1 checks the ALAT  for any entry
831 	 *	  which would overlap within [r3,r3+X] (the size of the load was store in the
832 	 *	  ALAT). If such an entry is found the entry is invalidated. But this is not good
833 	 *	  enough, take the following example:
834 	 *		r3=3
835 	 *		ld4.a r1=[r3]
836 	 *
837 	 *	  Could be emulated by doing:
838 	 *		ld1.a r1=[r3],1
839 	 *		store to temporary;
840 	 *		ld1.a r1=[r3],1
841 	 *		store & shift to temporary;
842 	 *		ld1.a r1=[r3],1
843 	 *		store & shift to temporary;
844 	 *		ld1.a r1=[r3]
845 	 *		store & shift to temporary;
846 	 *		r1=temporary
847 	 *
848 	 *	  So in this case, you would get the right value is r1 but the wrong info in
849 	 *	  the ALAT.  Notice that you could do it in reverse to finish with address 3
850 	 *	  but you would still get the size wrong.  To get the size right, one needs to
851 	 *	  execute exactly the same kind of load. You could do it from a aligned
852 	 *	  temporary location, but you would get the address wrong.
853 	 *
854 	 *	  So no matter what, it is not possible to emulate an advanced load
855 	 *	  correctly. But is that really critical ?
856 	 *
857 	 *	  We will always convert ld.a into a normal load with ALAT invalidated.  This
858 	 *	  will enable compiler to do optimization where certain code path after ld.a
859 	 *	  is not required to have ld.c/chk.a, e.g., code path with no intervening stores.
860 	 *
861 	 *	  If there is a store after the advanced load, one must either do a ld.c.* or
862 	 *	  chk.a.* to reuse the value stored in the ALAT. Both can "fail" (meaning no
863 	 *	  entry found in ALAT), and that's perfectly ok because:
864 	 *
865 	 *		- ld.c.*, if the entry is not present a  normal load is executed
866 	 *		- chk.a.*, if the entry is not present, execution jumps to recovery code
867 	 *
868 	 *	  In either case, the load can be potentially retried in another form.
869 	 *
870 	 *	  ALAT must be invalidated for the register (so that chk.a or ld.c don't pick
871 	 *	  up a stale entry later). The register base update MUST also be performed.
872 	 */
873 
874 	/*
875 	 * when the load has the .acq completer then
876 	 * use ordering fence.
877 	 */
878 	if (ld.x6_op == 0x5 || ld.x6_op == 0xa)
879 		mb();
880 
881 	/*
882 	 * invalidate ALAT entry in case of advanced load
883 	 */
884 	if (ld.x6_op == 0x2)
885 		invala_gr(ld.r1);
886 
887 	return 0;
888 }
889 
890 static int
891 emulate_store_int (unsigned long ifa, load_store_t ld, struct pt_regs *regs,
892 		   bool kernel_mode)
893 {
894 	unsigned long r2;
895 	unsigned int len = 1 << ld.x6_sz;
896 
897 	/*
898 	 * if we get to this handler, Nat bits on both r3 and r2 have already
899 	 * been checked. so we don't need to do it
900 	 *
901 	 * extract the value to be stored
902 	 */
903 	getreg(ld.imm, &r2, NULL, regs);
904 
905 	/*
906 	 * we rely on the macros in unaligned.h for now i.e.,
907 	 * we let the compiler figure out how to read memory gracefully.
908 	 *
909 	 * We need this switch/case because the way the inline function
910 	 * works. The code is optimized by the compiler and looks like
911 	 * a single switch/case.
912 	 */
913 	DPRINT("st%d [%lx]=%lx\n", len, ifa, r2);
914 
915 	if (len != 2 && len != 4 && len != 8) {
916 		DPRINT("unknown size: x6=%d\n", ld.x6_sz);
917 		return -1;
918 	}
919 
920 	/* this assumes little-endian byte-order: */
921 	if (emulate_store(ifa, &r2, len, kernel_mode))
922 		return -1;
923 
924 	/*
925 	 * stX [r3]=r2,imm(9)
926 	 *
927 	 * NOTE:
928 	 * ld.r3 can never be r0, because r0 would not generate an
929 	 * unaligned access.
930 	 */
931 	if (ld.op == 0x5) {
932 		unsigned long imm;
933 
934 		/*
935 		 * form imm9: [12:6] contain first 7bits
936 		 */
937 		imm = ld.x << 7 | ld.r1;
938 		/*
939 		 * sign extend (8bits) if m set
940 		 */
941 		if (ld.m) imm |= SIGN_EXT9;
942 		/*
943 		 * ifa == r3 (NaT is necessarily cleared)
944 		 */
945 		ifa += imm;
946 
947 		DPRINT("imm=%lx r3=%lx\n", imm, ifa);
948 
949 		setreg(ld.r3, ifa, 0, regs);
950 	}
951 	/*
952 	 * we don't have alat_invalidate_multiple() so we need
953 	 * to do the complete flush :-<<
954 	 */
955 	ia64_invala();
956 
957 	/*
958 	 * stX.rel: use fence instead of release
959 	 */
960 	if (ld.x6_op == 0xd)
961 		mb();
962 
963 	return 0;
964 }
965 
966 /*
967  * floating point operations sizes in bytes
968  */
969 static const unsigned char float_fsz[4]={
970 	10, /* extended precision (e) */
971 	8,  /* integer (8)            */
972 	4,  /* single precision (s)   */
973 	8   /* double precision (d)   */
974 };
975 
976 static inline void
977 mem2float_extended (struct ia64_fpreg *init, struct ia64_fpreg *final)
978 {
979 	ia64_ldfe(6, init);
980 	ia64_stop();
981 	ia64_stf_spill(final, 6);
982 }
983 
984 static inline void
985 mem2float_integer (struct ia64_fpreg *init, struct ia64_fpreg *final)
986 {
987 	ia64_ldf8(6, init);
988 	ia64_stop();
989 	ia64_stf_spill(final, 6);
990 }
991 
992 static inline void
993 mem2float_single (struct ia64_fpreg *init, struct ia64_fpreg *final)
994 {
995 	ia64_ldfs(6, init);
996 	ia64_stop();
997 	ia64_stf_spill(final, 6);
998 }
999 
1000 static inline void
1001 mem2float_double (struct ia64_fpreg *init, struct ia64_fpreg *final)
1002 {
1003 	ia64_ldfd(6, init);
1004 	ia64_stop();
1005 	ia64_stf_spill(final, 6);
1006 }
1007 
1008 static inline void
1009 float2mem_extended (struct ia64_fpreg *init, struct ia64_fpreg *final)
1010 {
1011 	ia64_ldf_fill(6, init);
1012 	ia64_stop();
1013 	ia64_stfe(final, 6);
1014 }
1015 
1016 static inline void
1017 float2mem_integer (struct ia64_fpreg *init, struct ia64_fpreg *final)
1018 {
1019 	ia64_ldf_fill(6, init);
1020 	ia64_stop();
1021 	ia64_stf8(final, 6);
1022 }
1023 
1024 static inline void
1025 float2mem_single (struct ia64_fpreg *init, struct ia64_fpreg *final)
1026 {
1027 	ia64_ldf_fill(6, init);
1028 	ia64_stop();
1029 	ia64_stfs(final, 6);
1030 }
1031 
1032 static inline void
1033 float2mem_double (struct ia64_fpreg *init, struct ia64_fpreg *final)
1034 {
1035 	ia64_ldf_fill(6, init);
1036 	ia64_stop();
1037 	ia64_stfd(final, 6);
1038 }
1039 
1040 static int
1041 emulate_load_floatpair (unsigned long ifa, load_store_t ld, struct pt_regs *regs, bool kernel_mode)
1042 {
1043 	struct ia64_fpreg fpr_init[2];
1044 	struct ia64_fpreg fpr_final[2];
1045 	unsigned long len = float_fsz[ld.x6_sz];
1046 
1047 	/*
1048 	 * fr0 & fr1 don't need to be checked because Illegal Instruction faults have
1049 	 * higher priority than unaligned faults.
1050 	 *
1051 	 * r0 cannot be found as the base as it would never generate an unaligned
1052 	 * reference.
1053 	 */
1054 
1055 	/*
1056 	 * make sure we get clean buffers
1057 	 */
1058 	memset(&fpr_init, 0, sizeof(fpr_init));
1059 	memset(&fpr_final, 0, sizeof(fpr_final));
1060 
1061 	/*
1062 	 * ldfpX.a: we don't try to emulate anything but we must
1063 	 * invalidate the ALAT entry and execute updates, if any.
1064 	 */
1065 	if (ld.x6_op != 0x2) {
1066 		/*
1067 		 * This assumes little-endian byte-order.  Note that there is no "ldfpe"
1068 		 * instruction:
1069 		 */
1070 		if (emulate_load(&fpr_init[0], ifa, len, kernel_mode)
1071 		    || emulate_load(&fpr_init[1], (ifa + len), len, kernel_mode))
1072 			return -1;
1073 
1074 		DPRINT("ld.r1=%d ld.imm=%d x6_sz=%d\n", ld.r1, ld.imm, ld.x6_sz);
1075 		DDUMP("frp_init =", &fpr_init, 2*len);
1076 		/*
1077 		 * XXX fixme
1078 		 * Could optimize inlines by using ldfpX & 2 spills
1079 		 */
1080 		switch( ld.x6_sz ) {
1081 			case 0:
1082 				mem2float_extended(&fpr_init[0], &fpr_final[0]);
1083 				mem2float_extended(&fpr_init[1], &fpr_final[1]);
1084 				break;
1085 			case 1:
1086 				mem2float_integer(&fpr_init[0], &fpr_final[0]);
1087 				mem2float_integer(&fpr_init[1], &fpr_final[1]);
1088 				break;
1089 			case 2:
1090 				mem2float_single(&fpr_init[0], &fpr_final[0]);
1091 				mem2float_single(&fpr_init[1], &fpr_final[1]);
1092 				break;
1093 			case 3:
1094 				mem2float_double(&fpr_init[0], &fpr_final[0]);
1095 				mem2float_double(&fpr_init[1], &fpr_final[1]);
1096 				break;
1097 		}
1098 		DDUMP("fpr_final =", &fpr_final, 2*len);
1099 		/*
1100 		 * XXX fixme
1101 		 *
1102 		 * A possible optimization would be to drop fpr_final and directly
1103 		 * use the storage from the saved context i.e., the actual final
1104 		 * destination (pt_regs, switch_stack or thread structure).
1105 		 */
1106 		setfpreg(ld.r1, &fpr_final[0], regs);
1107 		setfpreg(ld.imm, &fpr_final[1], regs);
1108 	}
1109 
1110 	/*
1111 	 * Check for updates: only immediate updates are available for this
1112 	 * instruction.
1113 	 */
1114 	if (ld.m) {
1115 		/*
1116 		 * the immediate is implicit given the ldsz of the operation:
1117 		 * single: 8 (2x4) and for  all others it's 16 (2x8)
1118 		 */
1119 		ifa += len<<1;
1120 
1121 		/*
1122 		 * IMPORTANT:
1123 		 * the fact that we force the NaT of r3 to zero is ONLY valid
1124 		 * as long as we don't come here with a ldfpX.s.
1125 		 * For this reason we keep this sanity check
1126 		 */
1127 		if (ld.x6_op == 1 || ld.x6_op == 3)
1128 			printk(KERN_ERR "%s: register update on speculative load pair, error\n",
1129 			       __func__);
1130 
1131 		setreg(ld.r3, ifa, 0, regs);
1132 	}
1133 
1134 	/*
1135 	 * Invalidate ALAT entries, if any, for both registers.
1136 	 */
1137 	if (ld.x6_op == 0x2) {
1138 		invala_fr(ld.r1);
1139 		invala_fr(ld.imm);
1140 	}
1141 	return 0;
1142 }
1143 
1144 
1145 static int
1146 emulate_load_float (unsigned long ifa, load_store_t ld, struct pt_regs *regs,
1147 	            bool kernel_mode)
1148 {
1149 	struct ia64_fpreg fpr_init;
1150 	struct ia64_fpreg fpr_final;
1151 	unsigned long len = float_fsz[ld.x6_sz];
1152 
1153 	/*
1154 	 * fr0 & fr1 don't need to be checked because Illegal Instruction
1155 	 * faults have higher priority than unaligned faults.
1156 	 *
1157 	 * r0 cannot be found as the base as it would never generate an
1158 	 * unaligned reference.
1159 	 */
1160 
1161 	/*
1162 	 * make sure we get clean buffers
1163 	 */
1164 	memset(&fpr_init,0, sizeof(fpr_init));
1165 	memset(&fpr_final,0, sizeof(fpr_final));
1166 
1167 	/*
1168 	 * ldfX.a we don't try to emulate anything but we must
1169 	 * invalidate the ALAT entry.
1170 	 * See comments in ldX for descriptions on how the various loads are handled.
1171 	 */
1172 	if (ld.x6_op != 0x2) {
1173 		if (emulate_load(&fpr_init, ifa, len, kernel_mode))
1174 			return -1;
1175 
1176 		DPRINT("ld.r1=%d x6_sz=%d\n", ld.r1, ld.x6_sz);
1177 		DDUMP("fpr_init =", &fpr_init, len);
1178 		/*
1179 		 * we only do something for x6_op={0,8,9}
1180 		 */
1181 		switch( ld.x6_sz ) {
1182 			case 0:
1183 				mem2float_extended(&fpr_init, &fpr_final);
1184 				break;
1185 			case 1:
1186 				mem2float_integer(&fpr_init, &fpr_final);
1187 				break;
1188 			case 2:
1189 				mem2float_single(&fpr_init, &fpr_final);
1190 				break;
1191 			case 3:
1192 				mem2float_double(&fpr_init, &fpr_final);
1193 				break;
1194 		}
1195 		DDUMP("fpr_final =", &fpr_final, len);
1196 		/*
1197 		 * XXX fixme
1198 		 *
1199 		 * A possible optimization would be to drop fpr_final and directly
1200 		 * use the storage from the saved context i.e., the actual final
1201 		 * destination (pt_regs, switch_stack or thread structure).
1202 		 */
1203 		setfpreg(ld.r1, &fpr_final, regs);
1204 	}
1205 
1206 	/*
1207 	 * check for updates on any loads
1208 	 */
1209 	if (ld.op == 0x7 || ld.m)
1210 		emulate_load_updates(ld.op == 0x7 ? UPD_IMMEDIATE: UPD_REG, ld, regs, ifa);
1211 
1212 	/*
1213 	 * invalidate ALAT entry in case of advanced floating point loads
1214 	 */
1215 	if (ld.x6_op == 0x2)
1216 		invala_fr(ld.r1);
1217 
1218 	return 0;
1219 }
1220 
1221 
1222 static int
1223 emulate_store_float (unsigned long ifa, load_store_t ld, struct pt_regs *regs,
1224 		     bool kernel_mode)
1225 {
1226 	struct ia64_fpreg fpr_init;
1227 	struct ia64_fpreg fpr_final;
1228 	unsigned long len = float_fsz[ld.x6_sz];
1229 
1230 	/*
1231 	 * make sure we get clean buffers
1232 	 */
1233 	memset(&fpr_init,0, sizeof(fpr_init));
1234 	memset(&fpr_final,0, sizeof(fpr_final));
1235 
1236 	/*
1237 	 * if we get to this handler, Nat bits on both r3 and r2 have already
1238 	 * been checked. so we don't need to do it
1239 	 *
1240 	 * extract the value to be stored
1241 	 */
1242 	getfpreg(ld.imm, &fpr_init, regs);
1243 	/*
1244 	 * during this step, we extract the spilled registers from the saved
1245 	 * context i.e., we refill. Then we store (no spill) to temporary
1246 	 * aligned location
1247 	 */
1248 	switch( ld.x6_sz ) {
1249 		case 0:
1250 			float2mem_extended(&fpr_init, &fpr_final);
1251 			break;
1252 		case 1:
1253 			float2mem_integer(&fpr_init, &fpr_final);
1254 			break;
1255 		case 2:
1256 			float2mem_single(&fpr_init, &fpr_final);
1257 			break;
1258 		case 3:
1259 			float2mem_double(&fpr_init, &fpr_final);
1260 			break;
1261 	}
1262 	DPRINT("ld.r1=%d x6_sz=%d\n", ld.r1, ld.x6_sz);
1263 	DDUMP("fpr_init =", &fpr_init, len);
1264 	DDUMP("fpr_final =", &fpr_final, len);
1265 
1266 	if (emulate_store(ifa, &fpr_final, len, kernel_mode))
1267 		return -1;
1268 
1269 	/*
1270 	 * stfX [r3]=r2,imm(9)
1271 	 *
1272 	 * NOTE:
1273 	 * ld.r3 can never be r0, because r0 would not generate an
1274 	 * unaligned access.
1275 	 */
1276 	if (ld.op == 0x7) {
1277 		unsigned long imm;
1278 
1279 		/*
1280 		 * form imm9: [12:6] contain first 7bits
1281 		 */
1282 		imm = ld.x << 7 | ld.r1;
1283 		/*
1284 		 * sign extend (8bits) if m set
1285 		 */
1286 		if (ld.m)
1287 			imm |= SIGN_EXT9;
1288 		/*
1289 		 * ifa == r3 (NaT is necessarily cleared)
1290 		 */
1291 		ifa += imm;
1292 
1293 		DPRINT("imm=%lx r3=%lx\n", imm, ifa);
1294 
1295 		setreg(ld.r3, ifa, 0, regs);
1296 	}
1297 	/*
1298 	 * we don't have alat_invalidate_multiple() so we need
1299 	 * to do the complete flush :-<<
1300 	 */
1301 	ia64_invala();
1302 
1303 	return 0;
1304 }
1305 
1306 /*
1307  * Make sure we log the unaligned access, so that user/sysadmin can notice it and
1308  * eventually fix the program.  However, we don't want to do that for every access so we
1309  * pace it with jiffies.
1310  */
1311 static DEFINE_RATELIMIT_STATE(logging_rate_limit, 5 * HZ, 5);
1312 
1313 void
1314 ia64_handle_unaligned (unsigned long ifa, struct pt_regs *regs)
1315 {
1316 	struct ia64_psr *ipsr = ia64_psr(regs);
1317 	unsigned long bundle[2];
1318 	unsigned long opcode;
1319 	const struct exception_table_entry *eh = NULL;
1320 	union {
1321 		unsigned long l;
1322 		load_store_t insn;
1323 	} u;
1324 	int ret = -1;
1325 	bool kernel_mode = false;
1326 
1327 	if (ia64_psr(regs)->be) {
1328 		/* we don't support big-endian accesses */
1329 		if (die_if_kernel("big-endian unaligned accesses are not supported", regs, 0))
1330 			return;
1331 		goto force_sigbus;
1332 	}
1333 
1334 	/*
1335 	 * Treat kernel accesses for which there is an exception handler entry the same as
1336 	 * user-level unaligned accesses.  Otherwise, a clever program could trick this
1337 	 * handler into reading an arbitrary kernel addresses...
1338 	 */
1339 	if (!user_mode(regs))
1340 		eh = search_exception_tables(regs->cr_iip + ia64_psr(regs)->ri);
1341 	if (user_mode(regs) || eh) {
1342 		if ((current->thread.flags & IA64_THREAD_UAC_SIGBUS) != 0)
1343 			goto force_sigbus;
1344 
1345 		if (!no_unaligned_warning &&
1346 		    !(current->thread.flags & IA64_THREAD_UAC_NOPRINT) &&
1347 		    __ratelimit(&logging_rate_limit))
1348 		{
1349 			char buf[200];	/* comm[] is at most 16 bytes... */
1350 			size_t len;
1351 
1352 			len = sprintf(buf, "%s(%d): unaligned access to 0x%016lx, "
1353 				      "ip=0x%016lx\n\r", current->comm,
1354 				      task_pid_nr(current),
1355 				      ifa, regs->cr_iip + ipsr->ri);
1356 			/*
1357 			 * Don't call tty_write_message() if we're in the kernel; we might
1358 			 * be holding locks...
1359 			 */
1360 			if (user_mode(regs)) {
1361 				struct tty_struct *tty = get_current_tty();
1362 				tty_write_message(tty, buf);
1363 				tty_kref_put(tty);
1364 			}
1365 			buf[len-1] = '\0';	/* drop '\r' */
1366 			/* watch for command names containing %s */
1367 			printk(KERN_WARNING "%s", buf);
1368 		} else {
1369 			if (no_unaligned_warning) {
1370 				printk_once(KERN_WARNING "%s(%d) encountered an "
1371 				       "unaligned exception which required\n"
1372 				       "kernel assistance, which degrades "
1373 				       "the performance of the application.\n"
1374 				       "Unaligned exception warnings have "
1375 				       "been disabled by the system "
1376 				       "administrator\n"
1377 				       "echo 0 > /proc/sys/kernel/ignore-"
1378 				       "unaligned-usertrap to re-enable\n",
1379 				       current->comm, task_pid_nr(current));
1380 			}
1381 		}
1382 	} else {
1383 		if (__ratelimit(&logging_rate_limit)) {
1384 			printk(KERN_WARNING "kernel unaligned access to 0x%016lx, ip=0x%016lx\n",
1385 			       ifa, regs->cr_iip + ipsr->ri);
1386 			if (unaligned_dump_stack)
1387 				dump_stack();
1388 		}
1389 		kernel_mode = true;
1390 	}
1391 
1392 	DPRINT("iip=%lx ifa=%lx isr=%lx (ei=%d, sp=%d)\n",
1393 	       regs->cr_iip, ifa, regs->cr_ipsr, ipsr->ri, ipsr->it);
1394 
1395 	if (emulate_load(bundle, regs->cr_iip, 16, kernel_mode))
1396 		goto failure;
1397 
1398 	/*
1399 	 * extract the instruction from the bundle given the slot number
1400 	 */
1401 	switch (ipsr->ri) {
1402 	      default:
1403 	      case 0: u.l = (bundle[0] >>  5); break;
1404 	      case 1: u.l = (bundle[0] >> 46) | (bundle[1] << 18); break;
1405 	      case 2: u.l = (bundle[1] >> 23); break;
1406 	}
1407 	opcode = (u.l >> IA64_OPCODE_SHIFT) & IA64_OPCODE_MASK;
1408 
1409 	DPRINT("opcode=%lx ld.qp=%d ld.r1=%d ld.imm=%d ld.r3=%d ld.x=%d ld.hint=%d "
1410 	       "ld.x6=0x%x ld.m=%d ld.op=%d\n", opcode, u.insn.qp, u.insn.r1, u.insn.imm,
1411 	       u.insn.r3, u.insn.x, u.insn.hint, u.insn.x6_sz, u.insn.m, u.insn.op);
1412 
1413 	/*
1414 	 * IMPORTANT:
1415 	 * Notice that the switch statement DOES not cover all possible instructions
1416 	 * that DO generate unaligned references. This is made on purpose because for some
1417 	 * instructions it DOES NOT make sense to try and emulate the access. Sometimes it
1418 	 * is WRONG to try and emulate. Here is a list of instruction we don't emulate i.e.,
1419 	 * the program will get a signal and die:
1420 	 *
1421 	 *	load/store:
1422 	 *		- ldX.spill
1423 	 *		- stX.spill
1424 	 *	Reason: RNATs are based on addresses
1425 	 *		- ld16
1426 	 *		- st16
1427 	 *	Reason: ld16 and st16 are supposed to occur in a single
1428 	 *		memory op
1429 	 *
1430 	 *	synchronization:
1431 	 *		- cmpxchg
1432 	 *		- fetchadd
1433 	 *		- xchg
1434 	 *	Reason: ATOMIC operations cannot be emulated properly using multiple
1435 	 *	        instructions.
1436 	 *
1437 	 *	speculative loads:
1438 	 *		- ldX.sZ
1439 	 *	Reason: side effects, code must be ready to deal with failure so simpler
1440 	 *		to let the load fail.
1441 	 * ---------------------------------------------------------------------------------
1442 	 * XXX fixme
1443 	 *
1444 	 * I would like to get rid of this switch case and do something
1445 	 * more elegant.
1446 	 */
1447 	switch (opcode) {
1448 	      case LDS_OP:
1449 	      case LDSA_OP:
1450 		if (u.insn.x)
1451 			/* oops, really a semaphore op (cmpxchg, etc) */
1452 			goto failure;
1453 		fallthrough;
1454 	      case LDS_IMM_OP:
1455 	      case LDSA_IMM_OP:
1456 	      case LDFS_OP:
1457 	      case LDFSA_OP:
1458 	      case LDFS_IMM_OP:
1459 		/*
1460 		 * The instruction will be retried with deferred exceptions turned on, and
1461 		 * we should get Nat bit installed
1462 		 *
1463 		 * IMPORTANT: When PSR_ED is set, the register & immediate update forms
1464 		 * are actually executed even though the operation failed. So we don't
1465 		 * need to take care of this.
1466 		 */
1467 		DPRINT("forcing PSR_ED\n");
1468 		regs->cr_ipsr |= IA64_PSR_ED;
1469 		goto done;
1470 
1471 	      case LD_OP:
1472 	      case LDA_OP:
1473 	      case LDBIAS_OP:
1474 	      case LDACQ_OP:
1475 	      case LDCCLR_OP:
1476 	      case LDCNC_OP:
1477 	      case LDCCLRACQ_OP:
1478 		if (u.insn.x)
1479 			/* oops, really a semaphore op (cmpxchg, etc) */
1480 			goto failure;
1481 		fallthrough;
1482 	      case LD_IMM_OP:
1483 	      case LDA_IMM_OP:
1484 	      case LDBIAS_IMM_OP:
1485 	      case LDACQ_IMM_OP:
1486 	      case LDCCLR_IMM_OP:
1487 	      case LDCNC_IMM_OP:
1488 	      case LDCCLRACQ_IMM_OP:
1489 		ret = emulate_load_int(ifa, u.insn, regs, kernel_mode);
1490 		break;
1491 
1492 	      case ST_OP:
1493 	      case STREL_OP:
1494 		if (u.insn.x)
1495 			/* oops, really a semaphore op (cmpxchg, etc) */
1496 			goto failure;
1497 		fallthrough;
1498 	      case ST_IMM_OP:
1499 	      case STREL_IMM_OP:
1500 		ret = emulate_store_int(ifa, u.insn, regs, kernel_mode);
1501 		break;
1502 
1503 	      case LDF_OP:
1504 	      case LDFA_OP:
1505 	      case LDFCCLR_OP:
1506 	      case LDFCNC_OP:
1507 		if (u.insn.x)
1508 			ret = emulate_load_floatpair(ifa, u.insn, regs, kernel_mode);
1509 		else
1510 			ret = emulate_load_float(ifa, u.insn, regs, kernel_mode);
1511 		break;
1512 
1513 	      case LDF_IMM_OP:
1514 	      case LDFA_IMM_OP:
1515 	      case LDFCCLR_IMM_OP:
1516 	      case LDFCNC_IMM_OP:
1517 		ret = emulate_load_float(ifa, u.insn, regs, kernel_mode);
1518 		break;
1519 
1520 	      case STF_OP:
1521 	      case STF_IMM_OP:
1522 		ret = emulate_store_float(ifa, u.insn, regs, kernel_mode);
1523 		break;
1524 
1525 	      default:
1526 		goto failure;
1527 	}
1528 	DPRINT("ret=%d\n", ret);
1529 	if (ret)
1530 		goto failure;
1531 
1532 	if (ipsr->ri == 2)
1533 		/*
1534 		 * given today's architecture this case is not likely to happen because a
1535 		 * memory access instruction (M) can never be in the last slot of a
1536 		 * bundle. But let's keep it for now.
1537 		 */
1538 		regs->cr_iip += 16;
1539 	ipsr->ri = (ipsr->ri + 1) & 0x3;
1540 
1541 	DPRINT("ipsr->ri=%d iip=%lx\n", ipsr->ri, regs->cr_iip);
1542   done:
1543 	return;
1544 
1545   failure:
1546 	/* something went wrong... */
1547 	if (!user_mode(regs)) {
1548 		if (eh) {
1549 			ia64_handle_exception(regs, eh);
1550 			goto done;
1551 		}
1552 		if (die_if_kernel("error during unaligned kernel access\n", regs, ret))
1553 			return;
1554 		/* NOT_REACHED */
1555 	}
1556   force_sigbus:
1557 	force_sig_fault(SIGBUS, BUS_ADRALN, (void __user *) ifa,
1558 			0, 0, 0);
1559 	goto done;
1560 }
1561