Linux Audio

Check our new training course

Loading...
v3.1
  1/*
  2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
  3 *
  4 *   This program is free software; you can redistribute it and/or
  5 *   modify it under the terms of the GNU General Public License
  6 *   as published by the Free Software Foundation, version 2.
  7 *
  8 *   This program is distributed in the hope that it will be useful, but
  9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
 10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
 11 *   NON INFRINGEMENT.  See the GNU General Public License for
 12 *   more details.
 13 *
 14 * A code-rewriter that enables instruction single-stepping.
 15 * Derived from iLib's single-stepping code.
 16 */
 17
 18#ifndef __tilegx__   /* Hardware support for single step unavailable. */
 19
 20/* These functions are only used on the TILE platform */
 21#include <linux/slab.h>
 22#include <linux/thread_info.h>
 23#include <linux/uaccess.h>
 24#include <linux/mman.h>
 25#include <linux/types.h>
 26#include <linux/err.h>
 
 27#include <asm/cacheflush.h>
 28#include <asm/opcode-tile.h>
 29#include <asm/opcode_constants.h>
 
 30#include <arch/abi.h>
 
 
 31
 32#define signExtend17(val) sign_extend((val), 17)
 33#define TILE_X1_MASK (0xffffffffULL << 31)
 34
 35int unaligned_printk;
 36
 37static int __init setup_unaligned_printk(char *str)
 38{
 39	long val;
 40	if (strict_strtol(str, 0, &val) != 0)
 41		return 0;
 42	unaligned_printk = val;
 43	pr_info("Printk for each unaligned data accesses is %s\n",
 44		unaligned_printk ? "enabled" : "disabled");
 45	return 1;
 46}
 47__setup("unaligned_printk=", setup_unaligned_printk);
 48
 49unsigned int unaligned_fixup_count;
 
 50
 51enum mem_op {
 52	MEMOP_NONE,
 53	MEMOP_LOAD,
 54	MEMOP_STORE,
 55	MEMOP_LOAD_POSTINCR,
 56	MEMOP_STORE_POSTINCR
 57};
 58
 59static inline tile_bundle_bits set_BrOff_X1(tile_bundle_bits n, s32 offset)
 
 60{
 61	tile_bundle_bits result;
 62
 63	/* mask out the old offset */
 64	tile_bundle_bits mask = create_BrOff_X1(-1);
 65	result = n & (~mask);
 66
 67	/* or in the new offset */
 68	result |= create_BrOff_X1(offset);
 69
 70	return result;
 71}
 72
 73static inline tile_bundle_bits move_X1(tile_bundle_bits n, int dest, int src)
 
 74{
 75	tile_bundle_bits result;
 76	tile_bundle_bits op;
 77
 78	result = n & (~TILE_X1_MASK);
 79
 80	op = create_Opcode_X1(SPECIAL_0_OPCODE_X1) |
 81		create_RRROpcodeExtension_X1(OR_SPECIAL_0_OPCODE_X1) |
 82		create_Dest_X1(dest) |
 83		create_SrcB_X1(TREG_ZERO) |
 84		create_SrcA_X1(src) ;
 85
 86	result |= op;
 87	return result;
 88}
 89
 90static inline tile_bundle_bits nop_X1(tile_bundle_bits n)
 91{
 92	return move_X1(n, TREG_ZERO, TREG_ZERO);
 93}
 94
 95static inline tile_bundle_bits addi_X1(
 96	tile_bundle_bits n, int dest, int src, int imm)
 97{
 98	n &= ~TILE_X1_MASK;
 99
100	n |=  (create_SrcA_X1(src) |
101	       create_Dest_X1(dest) |
102	       create_Imm8_X1(imm) |
103	       create_S_X1(0) |
104	       create_Opcode_X1(IMM_0_OPCODE_X1) |
105	       create_ImmOpcodeExtension_X1(ADDI_IMM_0_OPCODE_X1));
106
107	return n;
108}
109
110static tile_bundle_bits rewrite_load_store_unaligned(
111	struct single_step_state *state,
112	tile_bundle_bits bundle,
113	struct pt_regs *regs,
114	enum mem_op mem_op,
115	int size, int sign_ext)
116{
117	unsigned char __user *addr;
118	int val_reg, addr_reg, err, val;
 
 
 
 
 
 
 
 
 
 
 
119
120	/* Get address and value registers */
121	if (bundle & TILE_BUNDLE_Y_ENCODING_MASK) {
122		addr_reg = get_SrcA_Y2(bundle);
123		val_reg = get_SrcBDest_Y2(bundle);
124	} else if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
125		addr_reg = get_SrcA_X1(bundle);
126		val_reg  = get_Dest_X1(bundle);
127	} else {
128		addr_reg = get_SrcA_X1(bundle);
129		val_reg  = get_SrcB_X1(bundle);
130	}
131
132	/*
133	 * If registers are not GPRs, don't try to handle it.
134	 *
135	 * FIXME: we could handle non-GPR loads by getting the real value
136	 * from memory, writing it to the single step buffer, using a
137	 * temp_reg to hold a pointer to that memory, then executing that
138	 * instruction and resetting temp_reg.  For non-GPR stores, it's a
139	 * little trickier; we could use the single step buffer for that
140	 * too, but we'd have to add some more state bits so that we could
141	 * call back in here to copy that value to the real target.  For
142	 * now, we just handle the simple case.
143	 */
144	if ((val_reg >= PTREGS_NR_GPRS &&
145	     (val_reg != TREG_ZERO ||
146	      mem_op == MEMOP_LOAD ||
147	      mem_op == MEMOP_LOAD_POSTINCR)) ||
148	    addr_reg >= PTREGS_NR_GPRS)
149		return bundle;
150
151	/* If it's aligned, don't handle it specially */
152	addr = (void __user *)regs->regs[addr_reg];
153	if (((unsigned long)addr % size) == 0)
154		return bundle;
155
156#ifndef __LITTLE_ENDIAN
157# error We assume little-endian representation with copy_xx_user size 2 here
158#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
159	/* Handle unaligned load/store */
160	if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
161		unsigned short val_16;
162		switch (size) {
163		case 2:
164			err = copy_from_user(&val_16, addr, sizeof(val_16));
165			val = sign_ext ? ((short)val_16) : val_16;
166			break;
167		case 4:
168			err = copy_from_user(&val, addr, sizeof(val));
169			break;
170		default:
171			BUG();
172		}
173		if (err == 0) {
174			state->update_reg = val_reg;
175			state->update_value = val;
176			state->update = 1;
177		}
178	} else {
 
179		val = (val_reg == TREG_ZERO) ? 0 : regs->regs[val_reg];
180		err = copy_to_user(addr, &val, size);
 
 
 
 
 
 
 
 
 
 
181	}
182
183	if (err) {
184		siginfo_t info = {
185			.si_signo = SIGSEGV,
186			.si_code = SEGV_MAPERR,
187			.si_addr = addr
188		};
189		trace_unhandled_signal("segfault", regs,
190				       (unsigned long)addr, SIGSEGV);
191		force_sig_info(info.si_signo, &info, current);
192		return (tile_bundle_bits) 0;
193	}
194
195	if (unaligned_fixup == 0) {
196		siginfo_t info = {
197			.si_signo = SIGBUS,
198			.si_code = BUS_ADRALN,
199			.si_addr = addr
200		};
201		trace_unhandled_signal("unaligned trap", regs,
202				       (unsigned long)addr, SIGBUS);
203		force_sig_info(info.si_signo, &info, current);
204		return (tile_bundle_bits) 0;
205	}
206
207	if (unaligned_printk || unaligned_fixup_count == 0) {
208		pr_info("Process %d/%s: PC %#lx: Fixup of"
209			" unaligned %s at %#lx.\n",
210			current->pid, current->comm, regs->pc,
211			(mem_op == MEMOP_LOAD ||
212			 mem_op == MEMOP_LOAD_POSTINCR) ?
213			"load" : "store",
214			(unsigned long)addr);
215		if (!unaligned_printk) {
216#define P pr_info
217P("\n");
218P("Unaligned fixups in the kernel will slow your application considerably.\n");
219P("To find them, write a \"1\" to /proc/sys/tile/unaligned_fixup/printk,\n");
220P("which requests the kernel show all unaligned fixups, or write a \"0\"\n");
221P("to /proc/sys/tile/unaligned_fixup/enabled, in which case each unaligned\n");
222P("access will become a SIGBUS you can debug. No further warnings will be\n");
223P("shown so as to avoid additional slowdown, but you can track the number\n");
224P("of fixups performed via /proc/sys/tile/unaligned_fixup/count.\n");
225P("Use the tile-addr2line command (see \"info addr2line\") to decode PCs.\n");
226P("\n");
227#undef P
228		}
229	}
230	++unaligned_fixup_count;
231
232	if (bundle & TILE_BUNDLE_Y_ENCODING_MASK) {
233		/* Convert the Y2 instruction to a prefetch. */
234		bundle &= ~(create_SrcBDest_Y2(-1) |
235			    create_Opcode_Y2(-1));
236		bundle |= (create_SrcBDest_Y2(TREG_ZERO) |
237			   create_Opcode_Y2(LW_OPCODE_Y2));
238	/* Replace the load postincr with an addi */
239	} else if (mem_op == MEMOP_LOAD_POSTINCR) {
240		bundle = addi_X1(bundle, addr_reg, addr_reg,
241				 get_Imm8_X1(bundle));
242	/* Replace the store postincr with an addi */
243	} else if (mem_op == MEMOP_STORE_POSTINCR) {
244		bundle = addi_X1(bundle, addr_reg, addr_reg,
245				 get_Dest_Imm8_X1(bundle));
246	} else {
247		/* Convert the X1 instruction to a nop. */
248		bundle &= ~(create_Opcode_X1(-1) |
249			    create_UnShOpcodeExtension_X1(-1) |
250			    create_UnOpcodeExtension_X1(-1));
251		bundle |= (create_Opcode_X1(SHUN_0_OPCODE_X1) |
252			   create_UnShOpcodeExtension_X1(
253				   UN_0_SHUN_0_OPCODE_X1) |
254			   create_UnOpcodeExtension_X1(
255				   NOP_UN_0_SHUN_0_OPCODE_X1));
256	}
257
258	return bundle;
259}
260
261/*
262 * Called after execve() has started the new image.  This allows us
263 * to reset the info state.  Note that the the mmap'ed memory, if there
264 * was any, has already been unmapped by the exec.
265 */
266void single_step_execve(void)
267{
268	struct thread_info *ti = current_thread_info();
269	kfree(ti->step_state);
270	ti->step_state = NULL;
271}
272
273/**
274 * single_step_once() - entry point when single stepping has been triggered.
275 * @regs: The machine register state
276 *
277 *  When we arrive at this routine via a trampoline, the single step
278 *  engine copies the executing bundle to the single step buffer.
279 *  If the instruction is a condition branch, then the target is
280 *  reset to one past the next instruction. If the instruction
281 *  sets the lr, then that is noted. If the instruction is a jump
282 *  or call, then the new target pc is preserved and the current
283 *  bundle instruction set to null.
284 *
285 *  The necessary post-single-step rewriting information is stored in
286 *  single_step_state->  We use data segment values because the
287 *  stack will be rewound when we run the rewritten single-stepped
288 *  instruction.
289 */
290void single_step_once(struct pt_regs *regs)
291{
292	extern tile_bundle_bits __single_step_ill_insn;
293	extern tile_bundle_bits __single_step_j_insn;
294	extern tile_bundle_bits __single_step_addli_insn;
295	extern tile_bundle_bits __single_step_auli_insn;
296	struct thread_info *info = (void *)current_thread_info();
297	struct single_step_state *state = info->step_state;
298	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
299	tile_bundle_bits __user *buffer, *pc;
300	tile_bundle_bits bundle;
301	int temp_reg;
302	int target_reg = TREG_LR;
303	int err;
304	enum mem_op mem_op = MEMOP_NONE;
305	int size = 0, sign_ext = 0;  /* happy compiler */
 
 
 
 
 
 
 
 
 
 
 
306
307	asm(
308"    .pushsection .rodata.single_step\n"
309"    .align 8\n"
310"    .globl    __single_step_ill_insn\n"
311"__single_step_ill_insn:\n"
312"    ill\n"
313"    .globl    __single_step_addli_insn\n"
314"__single_step_addli_insn:\n"
315"    { nop; addli r0, zero, 0 }\n"
316"    .globl    __single_step_auli_insn\n"
317"__single_step_auli_insn:\n"
318"    { nop; auli r0, r0, 0 }\n"
319"    .globl    __single_step_j_insn\n"
320"__single_step_j_insn:\n"
321"    j .\n"
322"    .popsection\n"
323	);
324
325	/*
326	 * Enable interrupts here to allow touching userspace and the like.
327	 * The callers expect this: do_trap() already has interrupts
328	 * enabled, and do_work_pending() handles functions that enable
329	 * interrupts internally.
330	 */
331	local_irq_enable();
332
333	if (state == NULL) {
334		/* allocate a page of writable, executable memory */
335		state = kmalloc(sizeof(struct single_step_state), GFP_KERNEL);
336		if (state == NULL) {
337			pr_err("Out of kernel memory trying to single-step\n");
338			return;
339		}
340
341		/* allocate a cache line of writable, executable memory */
342		down_write(&current->mm->mmap_sem);
343		buffer = (void __user *) do_mmap(NULL, 0, 64,
344					  PROT_EXEC | PROT_READ | PROT_WRITE,
345					  MAP_PRIVATE | MAP_ANONYMOUS,
346					  0);
347		up_write(&current->mm->mmap_sem);
348
349		if (IS_ERR((void __force *)buffer)) {
350			kfree(state);
351			pr_err("Out of kernel pages trying to single-step\n");
352			return;
353		}
354
355		state->buffer = buffer;
356		state->is_enabled = 0;
357
358		info->step_state = state;
359
360		/* Validate our stored instruction patterns */
361		BUG_ON(get_Opcode_X1(__single_step_addli_insn) !=
362		       ADDLI_OPCODE_X1);
363		BUG_ON(get_Opcode_X1(__single_step_auli_insn) !=
364		       AULI_OPCODE_X1);
365		BUG_ON(get_SrcA_X1(__single_step_addli_insn) != TREG_ZERO);
366		BUG_ON(get_Dest_X1(__single_step_addli_insn) != 0);
367		BUG_ON(get_JOffLong_X1(__single_step_j_insn) != 0);
368	}
369
370	/*
371	 * If we are returning from a syscall, we still haven't hit the
372	 * "ill" for the swint1 instruction.  So back the PC up to be
373	 * pointing at the swint1, but we'll actually return directly
374	 * back to the "ill" so we come back in via SIGILL as if we
375	 * had "executed" the swint1 without ever being in kernel space.
376	 */
377	if (regs->faultnum == INT_SWINT_1)
378		regs->pc -= 8;
379
380	pc = (tile_bundle_bits __user *)(regs->pc);
381	if (get_user(bundle, pc) != 0) {
382		pr_err("Couldn't read instruction at %p trying to step\n", pc);
383		return;
384	}
385
386	/* We'll follow the instruction with 2 ill op bundles */
387	state->orig_pc = (unsigned long)pc;
388	state->next_pc = (unsigned long)(pc + 1);
389	state->branch_next_pc = 0;
390	state->update = 0;
391
392	if (!(bundle & TILE_BUNDLE_Y_ENCODING_MASK)) {
393		/* two wide, check for control flow */
394		int opcode = get_Opcode_X1(bundle);
395
396		switch (opcode) {
397		/* branches */
398		case BRANCH_OPCODE_X1:
399		{
400			s32 offset = signExtend17(get_BrOff_X1(bundle));
401
402			/*
403			 * For branches, we use a rewriting trick to let the
404			 * hardware evaluate whether the branch is taken or
405			 * untaken.  We record the target offset and then
406			 * rewrite the branch instruction to target 1 insn
407			 * ahead if the branch is taken.  We then follow the
408			 * rewritten branch with two bundles, each containing
409			 * an "ill" instruction. The supervisor examines the
410			 * pc after the single step code is executed, and if
411			 * the pc is the first ill instruction, then the
412			 * branch (if any) was not taken.  If the pc is the
413			 * second ill instruction, then the branch was
414			 * taken. The new pc is computed for these cases, and
415			 * inserted into the registers for the thread.  If
416			 * the pc is the start of the single step code, then
417			 * an exception or interrupt was taken before the
418			 * code started processing, and the same "original"
419			 * pc is restored.  This change, different from the
420			 * original implementation, has the advantage of
421			 * executing a single user instruction.
422			 */
423			state->branch_next_pc = (unsigned long)(pc + offset);
424
425			/* rewrite branch offset to go forward one bundle */
426			bundle = set_BrOff_X1(bundle, 2);
427		}
428		break;
429
430		/* jumps */
431		case JALB_OPCODE_X1:
432		case JALF_OPCODE_X1:
433			state->update = 1;
434			state->next_pc =
435				(unsigned long) (pc + get_JOffLong_X1(bundle));
436			break;
437
438		case JB_OPCODE_X1:
439		case JF_OPCODE_X1:
440			state->next_pc =
441				(unsigned long) (pc + get_JOffLong_X1(bundle));
442			bundle = nop_X1(bundle);
443			break;
444
445		case SPECIAL_0_OPCODE_X1:
446			switch (get_RRROpcodeExtension_X1(bundle)) {
447			/* jump-register */
448			case JALRP_SPECIAL_0_OPCODE_X1:
449			case JALR_SPECIAL_0_OPCODE_X1:
450				state->update = 1;
451				state->next_pc =
452					regs->regs[get_SrcA_X1(bundle)];
453				break;
454
455			case JRP_SPECIAL_0_OPCODE_X1:
456			case JR_SPECIAL_0_OPCODE_X1:
457				state->next_pc =
458					regs->regs[get_SrcA_X1(bundle)];
459				bundle = nop_X1(bundle);
460				break;
461
462			case LNK_SPECIAL_0_OPCODE_X1:
463				state->update = 1;
464				target_reg = get_Dest_X1(bundle);
465				break;
466
467			/* stores */
468			case SH_SPECIAL_0_OPCODE_X1:
469				mem_op = MEMOP_STORE;
470				size = 2;
471				break;
472
473			case SW_SPECIAL_0_OPCODE_X1:
474				mem_op = MEMOP_STORE;
475				size = 4;
476				break;
477			}
478			break;
479
480		/* loads and iret */
481		case SHUN_0_OPCODE_X1:
482			if (get_UnShOpcodeExtension_X1(bundle) ==
483			    UN_0_SHUN_0_OPCODE_X1) {
484				switch (get_UnOpcodeExtension_X1(bundle)) {
485				case LH_UN_0_SHUN_0_OPCODE_X1:
486					mem_op = MEMOP_LOAD;
487					size = 2;
488					sign_ext = 1;
489					break;
490
491				case LH_U_UN_0_SHUN_0_OPCODE_X1:
492					mem_op = MEMOP_LOAD;
493					size = 2;
494					sign_ext = 0;
495					break;
496
497				case LW_UN_0_SHUN_0_OPCODE_X1:
498					mem_op = MEMOP_LOAD;
499					size = 4;
500					break;
501
502				case IRET_UN_0_SHUN_0_OPCODE_X1:
503				{
504					unsigned long ex0_0 = __insn_mfspr(
505						SPR_EX_CONTEXT_0_0);
506					unsigned long ex0_1 = __insn_mfspr(
507						SPR_EX_CONTEXT_0_1);
508					/*
509					 * Special-case it if we're iret'ing
510					 * to PL0 again.  Otherwise just let
511					 * it run and it will generate SIGILL.
512					 */
513					if (EX1_PL(ex0_1) == USER_PL) {
514						state->next_pc = ex0_0;
515						regs->ex1 = ex0_1;
516						bundle = nop_X1(bundle);
517					}
518				}
519				}
520			}
521			break;
522
523#if CHIP_HAS_WH64()
524		/* postincrement operations */
525		case IMM_0_OPCODE_X1:
526			switch (get_ImmOpcodeExtension_X1(bundle)) {
527			case LWADD_IMM_0_OPCODE_X1:
528				mem_op = MEMOP_LOAD_POSTINCR;
529				size = 4;
530				break;
531
532			case LHADD_IMM_0_OPCODE_X1:
533				mem_op = MEMOP_LOAD_POSTINCR;
534				size = 2;
535				sign_ext = 1;
536				break;
537
538			case LHADD_U_IMM_0_OPCODE_X1:
539				mem_op = MEMOP_LOAD_POSTINCR;
540				size = 2;
541				sign_ext = 0;
542				break;
543
544			case SWADD_IMM_0_OPCODE_X1:
545				mem_op = MEMOP_STORE_POSTINCR;
546				size = 4;
547				break;
548
549			case SHADD_IMM_0_OPCODE_X1:
550				mem_op = MEMOP_STORE_POSTINCR;
551				size = 2;
552				break;
553
554			default:
555				break;
556			}
557			break;
558#endif /* CHIP_HAS_WH64() */
559		}
560
561		if (state->update) {
562			/*
563			 * Get an available register.  We start with a
564			 * bitmask with 1's for available registers.
565			 * We truncate to the low 32 registers since
566			 * we are guaranteed to have set bits in the
567			 * low 32 bits, then use ctz to pick the first.
568			 */
569			u32 mask = (u32) ~((1ULL << get_Dest_X0(bundle)) |
570					   (1ULL << get_SrcA_X0(bundle)) |
571					   (1ULL << get_SrcB_X0(bundle)) |
572					   (1ULL << target_reg));
573			temp_reg = __builtin_ctz(mask);
574			state->update_reg = temp_reg;
575			state->update_value = regs->regs[temp_reg];
576			regs->regs[temp_reg] = (unsigned long) (pc+1);
577			regs->flags |= PT_FLAGS_RESTORE_REGS;
578			bundle = move_X1(bundle, target_reg, temp_reg);
579		}
580	} else {
581		int opcode = get_Opcode_Y2(bundle);
582
583		switch (opcode) {
584		/* loads */
585		case LH_OPCODE_Y2:
586			mem_op = MEMOP_LOAD;
587			size = 2;
588			sign_ext = 1;
589			break;
590
591		case LH_U_OPCODE_Y2:
592			mem_op = MEMOP_LOAD;
593			size = 2;
594			sign_ext = 0;
595			break;
596
597		case LW_OPCODE_Y2:
598			mem_op = MEMOP_LOAD;
599			size = 4;
600			break;
601
602		/* stores */
603		case SH_OPCODE_Y2:
604			mem_op = MEMOP_STORE;
605			size = 2;
606			break;
607
608		case SW_OPCODE_Y2:
609			mem_op = MEMOP_STORE;
610			size = 4;
611			break;
612		}
613	}
614
615	/*
616	 * Check if we need to rewrite an unaligned load/store.
617	 * Returning zero is a special value meaning we need to SIGSEGV.
618	 */
619	if (mem_op != MEMOP_NONE && unaligned_fixup >= 0) {
620		bundle = rewrite_load_store_unaligned(state, bundle, regs,
621						      mem_op, size, sign_ext);
622		if (bundle == 0)
623			return;
624	}
625
626	/* write the bundle to our execution area */
627	buffer = state->buffer;
628	err = __put_user(bundle, buffer++);
629
630	/*
631	 * If we're really single-stepping, we take an INT_ILL after.
632	 * If we're just handling an unaligned access, we can just
633	 * jump directly back to where we were in user code.
634	 */
635	if (is_single_step) {
636		err |= __put_user(__single_step_ill_insn, buffer++);
637		err |= __put_user(__single_step_ill_insn, buffer++);
638	} else {
639		long delta;
640
641		if (state->update) {
642			/* We have some state to update; do it inline */
643			int ha16;
644			bundle = __single_step_addli_insn;
645			bundle |= create_Dest_X1(state->update_reg);
646			bundle |= create_Imm16_X1(state->update_value);
647			err |= __put_user(bundle, buffer++);
648			bundle = __single_step_auli_insn;
649			bundle |= create_Dest_X1(state->update_reg);
650			bundle |= create_SrcA_X1(state->update_reg);
651			ha16 = (state->update_value + 0x8000) >> 16;
652			bundle |= create_Imm16_X1(ha16);
653			err |= __put_user(bundle, buffer++);
654			state->update = 0;
655		}
656
657		/* End with a jump back to the next instruction */
658		delta = ((regs->pc + TILE_BUNDLE_SIZE_IN_BYTES) -
659			(unsigned long)buffer) >>
660			TILE_LOG2_BUNDLE_ALIGNMENT_IN_BYTES;
661		bundle = __single_step_j_insn;
662		bundle |= create_JOffLong_X1(delta);
663		err |= __put_user(bundle, buffer++);
664	}
665
666	if (err) {
667		pr_err("Fault when writing to single-step buffer\n");
668		return;
669	}
670
671	/*
672	 * Flush the buffer.
673	 * We do a local flush only, since this is a thread-specific buffer.
674	 */
675	__flush_icache_range((unsigned long)state->buffer,
676			     (unsigned long)buffer);
677
678	/* Indicate enabled */
679	state->is_enabled = is_single_step;
680	regs->pc = (unsigned long)state->buffer;
681
682	/* Fault immediately if we are coming back from a syscall. */
683	if (regs->faultnum == INT_SWINT_1)
684		regs->pc += 8;
685}
686
687#else
688#include <linux/smp.h>
689#include <linux/ptrace.h>
690#include <arch/spr_def.h>
691
692static DEFINE_PER_CPU(unsigned long, ss_saved_pc);
693
694
695/*
696 * Called directly on the occasion of an interrupt.
697 *
698 * If the process doesn't have single step set, then we use this as an
699 * opportunity to turn single step off.
700 *
701 * It has been mentioned that we could conditionally turn off single stepping
702 * on each entry into the kernel and rely on single_step_once to turn it
703 * on for the processes that matter (as we already do), but this
704 * implementation is somewhat more efficient in that we muck with registers
705 * once on a bum interrupt rather than on every entry into the kernel.
706 *
707 * If SINGLE_STEP_CONTROL_K has CANCELED set, then an interrupt occurred,
708 * so we have to run through this process again before we can say that an
709 * instruction has executed.
710 *
711 * swint will set CANCELED, but it's a legitimate instruction.  Fortunately
712 * it changes the PC.  If it hasn't changed, then we know that the interrupt
713 * wasn't generated by swint and we'll need to run this process again before
714 * we can say an instruction has executed.
715 *
716 * If either CANCELED == 0 or the PC's changed, we send out SIGTRAPs and get
717 * on with our lives.
718 */
719
720void gx_singlestep_handle(struct pt_regs *regs, int fault_num)
721{
722	unsigned long *ss_pc = &__get_cpu_var(ss_saved_pc);
723	struct thread_info *info = (void *)current_thread_info();
724	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
725	unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
726
727	if (is_single_step == 0) {
728		__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 0);
729
730	} else if ((*ss_pc != regs->pc) ||
731		   (!(control & SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK))) {
732
733		ptrace_notify(SIGTRAP);
734		control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
735		control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
736		__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
 
737	}
738}
739
740
741/*
742 * Called from need_singlestep.  Set up the control registers and the enable
743 * register, then return back.
744 */
745
746void single_step_once(struct pt_regs *regs)
747{
748	unsigned long *ss_pc = &__get_cpu_var(ss_saved_pc);
749	unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
750
751	*ss_pc = regs->pc;
752	control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
753	control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
754	__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
755	__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 1 << USER_PL);
756}
757
758void single_step_execve(void)
759{
760	/* Nothing */
761}
762
763#endif /* !__tilegx__ */
v3.15
  1/*
  2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
  3 *
  4 *   This program is free software; you can redistribute it and/or
  5 *   modify it under the terms of the GNU General Public License
  6 *   as published by the Free Software Foundation, version 2.
  7 *
  8 *   This program is distributed in the hope that it will be useful, but
  9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
 10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
 11 *   NON INFRINGEMENT.  See the GNU General Public License for
 12 *   more details.
 13 *
 14 * A code-rewriter that enables instruction single-stepping.
 
 15 */
 16
 17#include <linux/smp.h>
 18#include <linux/ptrace.h>
 
 19#include <linux/slab.h>
 20#include <linux/thread_info.h>
 21#include <linux/uaccess.h>
 22#include <linux/mman.h>
 23#include <linux/types.h>
 24#include <linux/err.h>
 25#include <linux/prctl.h>
 26#include <asm/cacheflush.h>
 27#include <asm/traps.h>
 28#include <asm/uaccess.h>
 29#include <asm/unaligned.h>
 30#include <arch/abi.h>
 31#include <arch/spr_def.h>
 32#include <arch/opcode.h>
 33
 
 
 
 
 34
 35#ifndef __tilegx__   /* Hardware support for single step unavailable. */
 
 
 
 
 
 
 
 
 
 
 36
 37#define signExtend17(val) sign_extend((val), 17)
 38#define TILE_X1_MASK (0xffffffffULL << 31)
 39
 40enum mem_op {
 41	MEMOP_NONE,
 42	MEMOP_LOAD,
 43	MEMOP_STORE,
 44	MEMOP_LOAD_POSTINCR,
 45	MEMOP_STORE_POSTINCR
 46};
 47
 48static inline tilepro_bundle_bits set_BrOff_X1(tilepro_bundle_bits n,
 49	s32 offset)
 50{
 51	tilepro_bundle_bits result;
 52
 53	/* mask out the old offset */
 54	tilepro_bundle_bits mask = create_BrOff_X1(-1);
 55	result = n & (~mask);
 56
 57	/* or in the new offset */
 58	result |= create_BrOff_X1(offset);
 59
 60	return result;
 61}
 62
 63static inline tilepro_bundle_bits move_X1(tilepro_bundle_bits n, int dest,
 64	int src)
 65{
 66	tilepro_bundle_bits result;
 67	tilepro_bundle_bits op;
 68
 69	result = n & (~TILE_X1_MASK);
 70
 71	op = create_Opcode_X1(SPECIAL_0_OPCODE_X1) |
 72		create_RRROpcodeExtension_X1(OR_SPECIAL_0_OPCODE_X1) |
 73		create_Dest_X1(dest) |
 74		create_SrcB_X1(TREG_ZERO) |
 75		create_SrcA_X1(src) ;
 76
 77	result |= op;
 78	return result;
 79}
 80
 81static inline tilepro_bundle_bits nop_X1(tilepro_bundle_bits n)
 82{
 83	return move_X1(n, TREG_ZERO, TREG_ZERO);
 84}
 85
 86static inline tilepro_bundle_bits addi_X1(
 87	tilepro_bundle_bits n, int dest, int src, int imm)
 88{
 89	n &= ~TILE_X1_MASK;
 90
 91	n |=  (create_SrcA_X1(src) |
 92	       create_Dest_X1(dest) |
 93	       create_Imm8_X1(imm) |
 94	       create_S_X1(0) |
 95	       create_Opcode_X1(IMM_0_OPCODE_X1) |
 96	       create_ImmOpcodeExtension_X1(ADDI_IMM_0_OPCODE_X1));
 97
 98	return n;
 99}
100
101static tilepro_bundle_bits rewrite_load_store_unaligned(
102	struct single_step_state *state,
103	tilepro_bundle_bits bundle,
104	struct pt_regs *regs,
105	enum mem_op mem_op,
106	int size, int sign_ext)
107{
108	unsigned char __user *addr;
109	int val_reg, addr_reg, err, val;
110	int align_ctl;
111
112	align_ctl = unaligned_fixup;
113	switch (task_thread_info(current)->align_ctl) {
114	case PR_UNALIGN_NOPRINT:
115		align_ctl = 1;
116		break;
117	case PR_UNALIGN_SIGBUS:
118		align_ctl = 0;
119		break;
120	}
121
122	/* Get address and value registers */
123	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
124		addr_reg = get_SrcA_Y2(bundle);
125		val_reg = get_SrcBDest_Y2(bundle);
126	} else if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
127		addr_reg = get_SrcA_X1(bundle);
128		val_reg  = get_Dest_X1(bundle);
129	} else {
130		addr_reg = get_SrcA_X1(bundle);
131		val_reg  = get_SrcB_X1(bundle);
132	}
133
134	/*
135	 * If registers are not GPRs, don't try to handle it.
136	 *
137	 * FIXME: we could handle non-GPR loads by getting the real value
138	 * from memory, writing it to the single step buffer, using a
139	 * temp_reg to hold a pointer to that memory, then executing that
140	 * instruction and resetting temp_reg.  For non-GPR stores, it's a
141	 * little trickier; we could use the single step buffer for that
142	 * too, but we'd have to add some more state bits so that we could
143	 * call back in here to copy that value to the real target.  For
144	 * now, we just handle the simple case.
145	 */
146	if ((val_reg >= PTREGS_NR_GPRS &&
147	     (val_reg != TREG_ZERO ||
148	      mem_op == MEMOP_LOAD ||
149	      mem_op == MEMOP_LOAD_POSTINCR)) ||
150	    addr_reg >= PTREGS_NR_GPRS)
151		return bundle;
152
153	/* If it's aligned, don't handle it specially */
154	addr = (void __user *)regs->regs[addr_reg];
155	if (((unsigned long)addr % size) == 0)
156		return bundle;
157
158	/*
159	 * Return SIGBUS with the unaligned address, if requested.
160	 * Note that we return SIGBUS even for completely invalid addresses
161	 * as long as they are in fact unaligned; this matches what the
162	 * tilepro hardware would be doing, if it could provide us with the
163	 * actual bad address in an SPR, which it doesn't.
164	 */
165	if (align_ctl == 0) {
166		siginfo_t info = {
167			.si_signo = SIGBUS,
168			.si_code = BUS_ADRALN,
169			.si_addr = addr
170		};
171		trace_unhandled_signal("unaligned trap", regs,
172				       (unsigned long)addr, SIGBUS);
173		force_sig_info(info.si_signo, &info, current);
174		return (tilepro_bundle_bits) 0;
175	}
176
177	/* Handle unaligned load/store */
178	if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
179		unsigned short val_16;
180		switch (size) {
181		case 2:
182			err = copy_from_user(&val_16, addr, sizeof(val_16));
183			val = sign_ext ? ((short)val_16) : val_16;
184			break;
185		case 4:
186			err = copy_from_user(&val, addr, sizeof(val));
187			break;
188		default:
189			BUG();
190		}
191		if (err == 0) {
192			state->update_reg = val_reg;
193			state->update_value = val;
194			state->update = 1;
195		}
196	} else {
197		unsigned short val_16;
198		val = (val_reg == TREG_ZERO) ? 0 : regs->regs[val_reg];
199		switch (size) {
200		case 2:
201			val_16 = val;
202			err = copy_to_user(addr, &val_16, sizeof(val_16));
203			break;
204		case 4:
205			err = copy_to_user(addr, &val, sizeof(val));
206			break;
207		default:
208			BUG();
209		}
210	}
211
212	if (err) {
213		siginfo_t info = {
 
 
 
 
 
 
 
 
 
 
 
 
214			.si_signo = SIGBUS,
215			.si_code = BUS_ADRALN,
216			.si_addr = addr
217		};
218		trace_unhandled_signal("bad address for unaligned fixup", regs,
219				       (unsigned long)addr, SIGBUS);
220		force_sig_info(info.si_signo, &info, current);
221		return (tilepro_bundle_bits) 0;
222	}
223
224	if (unaligned_printk || unaligned_fixup_count == 0) {
225		pr_info("Process %d/%s: PC %#lx: Fixup of"
226			" unaligned %s at %#lx.\n",
227			current->pid, current->comm, regs->pc,
228			(mem_op == MEMOP_LOAD ||
229			 mem_op == MEMOP_LOAD_POSTINCR) ?
230			"load" : "store",
231			(unsigned long)addr);
232		if (!unaligned_printk) {
233#define P pr_info
234P("\n");
235P("Unaligned fixups in the kernel will slow your application considerably.\n");
236P("To find them, write a \"1\" to /proc/sys/tile/unaligned_fixup/printk,\n");
237P("which requests the kernel show all unaligned fixups, or write a \"0\"\n");
238P("to /proc/sys/tile/unaligned_fixup/enabled, in which case each unaligned\n");
239P("access will become a SIGBUS you can debug. No further warnings will be\n");
240P("shown so as to avoid additional slowdown, but you can track the number\n");
241P("of fixups performed via /proc/sys/tile/unaligned_fixup/count.\n");
242P("Use the tile-addr2line command (see \"info addr2line\") to decode PCs.\n");
243P("\n");
244#undef P
245		}
246	}
247	++unaligned_fixup_count;
248
249	if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
250		/* Convert the Y2 instruction to a prefetch. */
251		bundle &= ~(create_SrcBDest_Y2(-1) |
252			    create_Opcode_Y2(-1));
253		bundle |= (create_SrcBDest_Y2(TREG_ZERO) |
254			   create_Opcode_Y2(LW_OPCODE_Y2));
255	/* Replace the load postincr with an addi */
256	} else if (mem_op == MEMOP_LOAD_POSTINCR) {
257		bundle = addi_X1(bundle, addr_reg, addr_reg,
258				 get_Imm8_X1(bundle));
259	/* Replace the store postincr with an addi */
260	} else if (mem_op == MEMOP_STORE_POSTINCR) {
261		bundle = addi_X1(bundle, addr_reg, addr_reg,
262				 get_Dest_Imm8_X1(bundle));
263	} else {
264		/* Convert the X1 instruction to a nop. */
265		bundle &= ~(create_Opcode_X1(-1) |
266			    create_UnShOpcodeExtension_X1(-1) |
267			    create_UnOpcodeExtension_X1(-1));
268		bundle |= (create_Opcode_X1(SHUN_0_OPCODE_X1) |
269			   create_UnShOpcodeExtension_X1(
270				   UN_0_SHUN_0_OPCODE_X1) |
271			   create_UnOpcodeExtension_X1(
272				   NOP_UN_0_SHUN_0_OPCODE_X1));
273	}
274
275	return bundle;
276}
277
278/*
279 * Called after execve() has started the new image.  This allows us
280 * to reset the info state.  Note that the the mmap'ed memory, if there
281 * was any, has already been unmapped by the exec.
282 */
283void single_step_execve(void)
284{
285	struct thread_info *ti = current_thread_info();
286	kfree(ti->step_state);
287	ti->step_state = NULL;
288}
289
290/*
291 * single_step_once() - entry point when single stepping has been triggered.
292 * @regs: The machine register state
293 *
294 *  When we arrive at this routine via a trampoline, the single step
295 *  engine copies the executing bundle to the single step buffer.
296 *  If the instruction is a condition branch, then the target is
297 *  reset to one past the next instruction. If the instruction
298 *  sets the lr, then that is noted. If the instruction is a jump
299 *  or call, then the new target pc is preserved and the current
300 *  bundle instruction set to null.
301 *
302 *  The necessary post-single-step rewriting information is stored in
303 *  single_step_state->  We use data segment values because the
304 *  stack will be rewound when we run the rewritten single-stepped
305 *  instruction.
306 */
307void single_step_once(struct pt_regs *regs)
308{
309	extern tilepro_bundle_bits __single_step_ill_insn;
310	extern tilepro_bundle_bits __single_step_j_insn;
311	extern tilepro_bundle_bits __single_step_addli_insn;
312	extern tilepro_bundle_bits __single_step_auli_insn;
313	struct thread_info *info = (void *)current_thread_info();
314	struct single_step_state *state = info->step_state;
315	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
316	tilepro_bundle_bits __user *buffer, *pc;
317	tilepro_bundle_bits bundle;
318	int temp_reg;
319	int target_reg = TREG_LR;
320	int err;
321	enum mem_op mem_op = MEMOP_NONE;
322	int size = 0, sign_ext = 0;  /* happy compiler */
323	int align_ctl;
324
325	align_ctl = unaligned_fixup;
326	switch (task_thread_info(current)->align_ctl) {
327	case PR_UNALIGN_NOPRINT:
328		align_ctl = 1;
329		break;
330	case PR_UNALIGN_SIGBUS:
331		align_ctl = 0;
332		break;
333	}
334
335	asm(
336"    .pushsection .rodata.single_step\n"
337"    .align 8\n"
338"    .globl    __single_step_ill_insn\n"
339"__single_step_ill_insn:\n"
340"    ill\n"
341"    .globl    __single_step_addli_insn\n"
342"__single_step_addli_insn:\n"
343"    { nop; addli r0, zero, 0 }\n"
344"    .globl    __single_step_auli_insn\n"
345"__single_step_auli_insn:\n"
346"    { nop; auli r0, r0, 0 }\n"
347"    .globl    __single_step_j_insn\n"
348"__single_step_j_insn:\n"
349"    j .\n"
350"    .popsection\n"
351	);
352
353	/*
354	 * Enable interrupts here to allow touching userspace and the like.
355	 * The callers expect this: do_trap() already has interrupts
356	 * enabled, and do_work_pending() handles functions that enable
357	 * interrupts internally.
358	 */
359	local_irq_enable();
360
361	if (state == NULL) {
362		/* allocate a page of writable, executable memory */
363		state = kmalloc(sizeof(struct single_step_state), GFP_KERNEL);
364		if (state == NULL) {
365			pr_err("Out of kernel memory trying to single-step\n");
366			return;
367		}
368
369		/* allocate a cache line of writable, executable memory */
370		buffer = (void __user *) vm_mmap(NULL, 0, 64,
 
371					  PROT_EXEC | PROT_READ | PROT_WRITE,
372					  MAP_PRIVATE | MAP_ANONYMOUS,
373					  0);
 
374
375		if (IS_ERR((void __force *)buffer)) {
376			kfree(state);
377			pr_err("Out of kernel pages trying to single-step\n");
378			return;
379		}
380
381		state->buffer = buffer;
382		state->is_enabled = 0;
383
384		info->step_state = state;
385
386		/* Validate our stored instruction patterns */
387		BUG_ON(get_Opcode_X1(__single_step_addli_insn) !=
388		       ADDLI_OPCODE_X1);
389		BUG_ON(get_Opcode_X1(__single_step_auli_insn) !=
390		       AULI_OPCODE_X1);
391		BUG_ON(get_SrcA_X1(__single_step_addli_insn) != TREG_ZERO);
392		BUG_ON(get_Dest_X1(__single_step_addli_insn) != 0);
393		BUG_ON(get_JOffLong_X1(__single_step_j_insn) != 0);
394	}
395
396	/*
397	 * If we are returning from a syscall, we still haven't hit the
398	 * "ill" for the swint1 instruction.  So back the PC up to be
399	 * pointing at the swint1, but we'll actually return directly
400	 * back to the "ill" so we come back in via SIGILL as if we
401	 * had "executed" the swint1 without ever being in kernel space.
402	 */
403	if (regs->faultnum == INT_SWINT_1)
404		regs->pc -= 8;
405
406	pc = (tilepro_bundle_bits __user *)(regs->pc);
407	if (get_user(bundle, pc) != 0) {
408		pr_err("Couldn't read instruction at %p trying to step\n", pc);
409		return;
410	}
411
412	/* We'll follow the instruction with 2 ill op bundles */
413	state->orig_pc = (unsigned long)pc;
414	state->next_pc = (unsigned long)(pc + 1);
415	state->branch_next_pc = 0;
416	state->update = 0;
417
418	if (!(bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK)) {
419		/* two wide, check for control flow */
420		int opcode = get_Opcode_X1(bundle);
421
422		switch (opcode) {
423		/* branches */
424		case BRANCH_OPCODE_X1:
425		{
426			s32 offset = signExtend17(get_BrOff_X1(bundle));
427
428			/*
429			 * For branches, we use a rewriting trick to let the
430			 * hardware evaluate whether the branch is taken or
431			 * untaken.  We record the target offset and then
432			 * rewrite the branch instruction to target 1 insn
433			 * ahead if the branch is taken.  We then follow the
434			 * rewritten branch with two bundles, each containing
435			 * an "ill" instruction. The supervisor examines the
436			 * pc after the single step code is executed, and if
437			 * the pc is the first ill instruction, then the
438			 * branch (if any) was not taken.  If the pc is the
439			 * second ill instruction, then the branch was
440			 * taken. The new pc is computed for these cases, and
441			 * inserted into the registers for the thread.  If
442			 * the pc is the start of the single step code, then
443			 * an exception or interrupt was taken before the
444			 * code started processing, and the same "original"
445			 * pc is restored.  This change, different from the
446			 * original implementation, has the advantage of
447			 * executing a single user instruction.
448			 */
449			state->branch_next_pc = (unsigned long)(pc + offset);
450
451			/* rewrite branch offset to go forward one bundle */
452			bundle = set_BrOff_X1(bundle, 2);
453		}
454		break;
455
456		/* jumps */
457		case JALB_OPCODE_X1:
458		case JALF_OPCODE_X1:
459			state->update = 1;
460			state->next_pc =
461				(unsigned long) (pc + get_JOffLong_X1(bundle));
462			break;
463
464		case JB_OPCODE_X1:
465		case JF_OPCODE_X1:
466			state->next_pc =
467				(unsigned long) (pc + get_JOffLong_X1(bundle));
468			bundle = nop_X1(bundle);
469			break;
470
471		case SPECIAL_0_OPCODE_X1:
472			switch (get_RRROpcodeExtension_X1(bundle)) {
473			/* jump-register */
474			case JALRP_SPECIAL_0_OPCODE_X1:
475			case JALR_SPECIAL_0_OPCODE_X1:
476				state->update = 1;
477				state->next_pc =
478					regs->regs[get_SrcA_X1(bundle)];
479				break;
480
481			case JRP_SPECIAL_0_OPCODE_X1:
482			case JR_SPECIAL_0_OPCODE_X1:
483				state->next_pc =
484					regs->regs[get_SrcA_X1(bundle)];
485				bundle = nop_X1(bundle);
486				break;
487
488			case LNK_SPECIAL_0_OPCODE_X1:
489				state->update = 1;
490				target_reg = get_Dest_X1(bundle);
491				break;
492
493			/* stores */
494			case SH_SPECIAL_0_OPCODE_X1:
495				mem_op = MEMOP_STORE;
496				size = 2;
497				break;
498
499			case SW_SPECIAL_0_OPCODE_X1:
500				mem_op = MEMOP_STORE;
501				size = 4;
502				break;
503			}
504			break;
505
506		/* loads and iret */
507		case SHUN_0_OPCODE_X1:
508			if (get_UnShOpcodeExtension_X1(bundle) ==
509			    UN_0_SHUN_0_OPCODE_X1) {
510				switch (get_UnOpcodeExtension_X1(bundle)) {
511				case LH_UN_0_SHUN_0_OPCODE_X1:
512					mem_op = MEMOP_LOAD;
513					size = 2;
514					sign_ext = 1;
515					break;
516
517				case LH_U_UN_0_SHUN_0_OPCODE_X1:
518					mem_op = MEMOP_LOAD;
519					size = 2;
520					sign_ext = 0;
521					break;
522
523				case LW_UN_0_SHUN_0_OPCODE_X1:
524					mem_op = MEMOP_LOAD;
525					size = 4;
526					break;
527
528				case IRET_UN_0_SHUN_0_OPCODE_X1:
529				{
530					unsigned long ex0_0 = __insn_mfspr(
531						SPR_EX_CONTEXT_0_0);
532					unsigned long ex0_1 = __insn_mfspr(
533						SPR_EX_CONTEXT_0_1);
534					/*
535					 * Special-case it if we're iret'ing
536					 * to PL0 again.  Otherwise just let
537					 * it run and it will generate SIGILL.
538					 */
539					if (EX1_PL(ex0_1) == USER_PL) {
540						state->next_pc = ex0_0;
541						regs->ex1 = ex0_1;
542						bundle = nop_X1(bundle);
543					}
544				}
545				}
546			}
547			break;
548
 
549		/* postincrement operations */
550		case IMM_0_OPCODE_X1:
551			switch (get_ImmOpcodeExtension_X1(bundle)) {
552			case LWADD_IMM_0_OPCODE_X1:
553				mem_op = MEMOP_LOAD_POSTINCR;
554				size = 4;
555				break;
556
557			case LHADD_IMM_0_OPCODE_X1:
558				mem_op = MEMOP_LOAD_POSTINCR;
559				size = 2;
560				sign_ext = 1;
561				break;
562
563			case LHADD_U_IMM_0_OPCODE_X1:
564				mem_op = MEMOP_LOAD_POSTINCR;
565				size = 2;
566				sign_ext = 0;
567				break;
568
569			case SWADD_IMM_0_OPCODE_X1:
570				mem_op = MEMOP_STORE_POSTINCR;
571				size = 4;
572				break;
573
574			case SHADD_IMM_0_OPCODE_X1:
575				mem_op = MEMOP_STORE_POSTINCR;
576				size = 2;
577				break;
578
579			default:
580				break;
581			}
582			break;
 
583		}
584
585		if (state->update) {
586			/*
587			 * Get an available register.  We start with a
588			 * bitmask with 1's for available registers.
589			 * We truncate to the low 32 registers since
590			 * we are guaranteed to have set bits in the
591			 * low 32 bits, then use ctz to pick the first.
592			 */
593			u32 mask = (u32) ~((1ULL << get_Dest_X0(bundle)) |
594					   (1ULL << get_SrcA_X0(bundle)) |
595					   (1ULL << get_SrcB_X0(bundle)) |
596					   (1ULL << target_reg));
597			temp_reg = __builtin_ctz(mask);
598			state->update_reg = temp_reg;
599			state->update_value = regs->regs[temp_reg];
600			regs->regs[temp_reg] = (unsigned long) (pc+1);
601			regs->flags |= PT_FLAGS_RESTORE_REGS;
602			bundle = move_X1(bundle, target_reg, temp_reg);
603		}
604	} else {
605		int opcode = get_Opcode_Y2(bundle);
606
607		switch (opcode) {
608		/* loads */
609		case LH_OPCODE_Y2:
610			mem_op = MEMOP_LOAD;
611			size = 2;
612			sign_ext = 1;
613			break;
614
615		case LH_U_OPCODE_Y2:
616			mem_op = MEMOP_LOAD;
617			size = 2;
618			sign_ext = 0;
619			break;
620
621		case LW_OPCODE_Y2:
622			mem_op = MEMOP_LOAD;
623			size = 4;
624			break;
625
626		/* stores */
627		case SH_OPCODE_Y2:
628			mem_op = MEMOP_STORE;
629			size = 2;
630			break;
631
632		case SW_OPCODE_Y2:
633			mem_op = MEMOP_STORE;
634			size = 4;
635			break;
636		}
637	}
638
639	/*
640	 * Check if we need to rewrite an unaligned load/store.
641	 * Returning zero is a special value meaning we generated a signal.
642	 */
643	if (mem_op != MEMOP_NONE && align_ctl >= 0) {
644		bundle = rewrite_load_store_unaligned(state, bundle, regs,
645						      mem_op, size, sign_ext);
646		if (bundle == 0)
647			return;
648	}
649
650	/* write the bundle to our execution area */
651	buffer = state->buffer;
652	err = __put_user(bundle, buffer++);
653
654	/*
655	 * If we're really single-stepping, we take an INT_ILL after.
656	 * If we're just handling an unaligned access, we can just
657	 * jump directly back to where we were in user code.
658	 */
659	if (is_single_step) {
660		err |= __put_user(__single_step_ill_insn, buffer++);
661		err |= __put_user(__single_step_ill_insn, buffer++);
662	} else {
663		long delta;
664
665		if (state->update) {
666			/* We have some state to update; do it inline */
667			int ha16;
668			bundle = __single_step_addli_insn;
669			bundle |= create_Dest_X1(state->update_reg);
670			bundle |= create_Imm16_X1(state->update_value);
671			err |= __put_user(bundle, buffer++);
672			bundle = __single_step_auli_insn;
673			bundle |= create_Dest_X1(state->update_reg);
674			bundle |= create_SrcA_X1(state->update_reg);
675			ha16 = (state->update_value + 0x8000) >> 16;
676			bundle |= create_Imm16_X1(ha16);
677			err |= __put_user(bundle, buffer++);
678			state->update = 0;
679		}
680
681		/* End with a jump back to the next instruction */
682		delta = ((regs->pc + TILEPRO_BUNDLE_SIZE_IN_BYTES) -
683			(unsigned long)buffer) >>
684			TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES;
685		bundle = __single_step_j_insn;
686		bundle |= create_JOffLong_X1(delta);
687		err |= __put_user(bundle, buffer++);
688	}
689
690	if (err) {
691		pr_err("Fault when writing to single-step buffer\n");
692		return;
693	}
694
695	/*
696	 * Flush the buffer.
697	 * We do a local flush only, since this is a thread-specific buffer.
698	 */
699	__flush_icache_range((unsigned long)state->buffer,
700			     (unsigned long)buffer);
701
702	/* Indicate enabled */
703	state->is_enabled = is_single_step;
704	regs->pc = (unsigned long)state->buffer;
705
706	/* Fault immediately if we are coming back from a syscall. */
707	if (regs->faultnum == INT_SWINT_1)
708		regs->pc += 8;
709}
710
711#else
 
 
 
712
713static DEFINE_PER_CPU(unsigned long, ss_saved_pc);
714
715
716/*
717 * Called directly on the occasion of an interrupt.
718 *
719 * If the process doesn't have single step set, then we use this as an
720 * opportunity to turn single step off.
721 *
722 * It has been mentioned that we could conditionally turn off single stepping
723 * on each entry into the kernel and rely on single_step_once to turn it
724 * on for the processes that matter (as we already do), but this
725 * implementation is somewhat more efficient in that we muck with registers
726 * once on a bum interrupt rather than on every entry into the kernel.
727 *
728 * If SINGLE_STEP_CONTROL_K has CANCELED set, then an interrupt occurred,
729 * so we have to run through this process again before we can say that an
730 * instruction has executed.
731 *
732 * swint will set CANCELED, but it's a legitimate instruction.  Fortunately
733 * it changes the PC.  If it hasn't changed, then we know that the interrupt
734 * wasn't generated by swint and we'll need to run this process again before
735 * we can say an instruction has executed.
736 *
737 * If either CANCELED == 0 or the PC's changed, we send out SIGTRAPs and get
738 * on with our lives.
739 */
740
741void gx_singlestep_handle(struct pt_regs *regs, int fault_num)
742{
743	unsigned long *ss_pc = &__get_cpu_var(ss_saved_pc);
744	struct thread_info *info = (void *)current_thread_info();
745	int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
746	unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
747
748	if (is_single_step == 0) {
749		__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 0);
750
751	} else if ((*ss_pc != regs->pc) ||
752		   (!(control & SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK))) {
753
 
754		control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
755		control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
756		__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
757		send_sigtrap(current, regs);
758	}
759}
760
761
762/*
763 * Called from need_singlestep.  Set up the control registers and the enable
764 * register, then return back.
765 */
766
767void single_step_once(struct pt_regs *regs)
768{
769	unsigned long *ss_pc = &__get_cpu_var(ss_saved_pc);
770	unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
771
772	*ss_pc = regs->pc;
773	control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
774	control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
775	__insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
776	__insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 1 << USER_PL);
777}
778
779void single_step_execve(void)
780{
781	/* Nothing */
782}
783
784#endif /* !__tilegx__ */