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
15#include <linux/sched.h>
16#include <linux/preempt.h>
17#include <linux/module.h>
18#include <linux/fs.h>
19#include <linux/kprobes.h>
20#include <linux/elfcore.h>
21#include <linux/tick.h>
22#include <linux/init.h>
23#include <linux/mm.h>
24#include <linux/compat.h>
25#include <linux/hardirq.h>
26#include <linux/syscalls.h>
27#include <linux/kernel.h>
28#include <linux/tracehook.h>
29#include <linux/signal.h>
30#include <linux/delay.h>
31#include <linux/context_tracking.h>
32#include <asm/stack.h>
33#include <asm/switch_to.h>
34#include <asm/homecache.h>
35#include <asm/syscalls.h>
36#include <asm/traps.h>
37#include <asm/setup.h>
38#include <asm/uaccess.h>
39#ifdef CONFIG_HARDWALL
40#include <asm/hardwall.h>
41#endif
42#include <arch/chip.h>
43#include <arch/abi.h>
44#include <arch/sim_def.h>
45
46/*
47 * Use the (x86) "idle=poll" option to prefer low latency when leaving the
48 * idle loop over low power while in the idle loop, e.g. if we have
49 * one thread per core and we want to get threads out of futex waits fast.
50 */
51static int __init idle_setup(char *str)
52{
53	if (!str)
54		return -EINVAL;
55
56	if (!strcmp(str, "poll")) {
57		pr_info("using polling idle threads\n");
58		cpu_idle_poll_ctrl(true);
59		return 0;
60	} else if (!strcmp(str, "halt")) {
61		return 0;
62	}
63	return -1;
64}
65early_param("idle", idle_setup);
66
67void arch_cpu_idle(void)
68{
69	__this_cpu_write(irq_stat.idle_timestamp, jiffies);
70	_cpu_idle();
71}
72
73/*
74 * Release a thread_info structure
75 */
76void arch_release_thread_info(struct thread_info *info)
77{
78	struct single_step_state *step_state = info->step_state;
79
80	if (step_state) {
81
82		/*
83		 * FIXME: we don't munmap step_state->buffer
84		 * because the mm_struct for this process (info->task->mm)
85		 * has already been zeroed in exit_mm().  Keeping a
86		 * reference to it here seems like a bad move, so this
87		 * means we can't munmap() the buffer, and therefore if we
88		 * ptrace multiple threads in a process, we will slowly
89		 * leak user memory.  (Note that as soon as the last
90		 * thread in a process dies, we will reclaim all user
91		 * memory including single-step buffers in the usual way.)
92		 * We should either assign a kernel VA to this buffer
93		 * somehow, or we should associate the buffer(s) with the
94		 * mm itself so we can clean them up that way.
95		 */
96		kfree(step_state);
97	}
98}
99
100static void save_arch_state(struct thread_struct *t);
101
102int copy_thread(unsigned long clone_flags, unsigned long sp,
103		unsigned long arg, struct task_struct *p)
104{
105	struct pt_regs *childregs = task_pt_regs(p);
106	unsigned long ksp;
107	unsigned long *callee_regs;
108
109	/*
110	 * Set up the stack and stack pointer appropriately for the
111	 * new child to find itself woken up in __switch_to().
112	 * The callee-saved registers must be on the stack to be read;
113	 * the new task will then jump to assembly support to handle
114	 * calling schedule_tail(), etc., and (for userspace tasks)
115	 * returning to the context set up in the pt_regs.
116	 */
117	ksp = (unsigned long) childregs;
118	ksp -= C_ABI_SAVE_AREA_SIZE;   /* interrupt-entry save area */
119	((long *)ksp)[0] = ((long *)ksp)[1] = 0;
120	ksp -= CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long);
121	callee_regs = (unsigned long *)ksp;
122	ksp -= C_ABI_SAVE_AREA_SIZE;   /* __switch_to() save area */
123	((long *)ksp)[0] = ((long *)ksp)[1] = 0;
124	p->thread.ksp = ksp;
125
126	/* Record the pid of the task that created this one. */
127	p->thread.creator_pid = current->pid;
128
129	if (unlikely(p->flags & PF_KTHREAD)) {
130		/* kernel thread */
131		memset(childregs, 0, sizeof(struct pt_regs));
132		memset(&callee_regs[2], 0,
133		       (CALLEE_SAVED_REGS_COUNT - 2) * sizeof(unsigned long));
134		callee_regs[0] = sp;   /* r30 = function */
135		callee_regs[1] = arg;  /* r31 = arg */
136		p->thread.pc = (unsigned long) ret_from_kernel_thread;
137		return 0;
138	}
139
140	/*
141	 * Start new thread in ret_from_fork so it schedules properly
142	 * and then return from interrupt like the parent.
143	 */
144	p->thread.pc = (unsigned long) ret_from_fork;
145
146	/*
147	 * Do not clone step state from the parent; each thread
148	 * must make its own lazily.
149	 */
150	task_thread_info(p)->step_state = NULL;
151
152#ifdef __tilegx__
153	/*
154	 * Do not clone unalign jit fixup from the parent; each thread
155	 * must allocate its own on demand.
156	 */
157	task_thread_info(p)->unalign_jit_base = NULL;
158#endif
159
160	/*
161	 * Copy the registers onto the kernel stack so the
162	 * return-from-interrupt code will reload it into registers.
163	 */
164	*childregs = *current_pt_regs();
165	childregs->regs[0] = 0;         /* return value is zero */
166	if (sp)
167		childregs->sp = sp;  /* override with new user stack pointer */
168	memcpy(callee_regs, &childregs->regs[CALLEE_SAVED_FIRST_REG],
169	       CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long));
170
171	/* Save user stack top pointer so we can ID the stack vm area later. */
172	p->thread.usp0 = childregs->sp;
173
174	/*
175	 * If CLONE_SETTLS is set, set "tp" in the new task to "r4",
176	 * which is passed in as arg #5 to sys_clone().
177	 */
178	if (clone_flags & CLONE_SETTLS)
179		childregs->tp = childregs->regs[4];
180
181
182#if CHIP_HAS_TILE_DMA()
183	/*
184	 * No DMA in the new thread.  We model this on the fact that
185	 * fork() clears the pending signals, alarms, and aio for the child.
186	 */
187	memset(&p->thread.tile_dma_state, 0, sizeof(struct tile_dma_state));
188	memset(&p->thread.dma_async_tlb, 0, sizeof(struct async_tlb));
189#endif
190
191	/* New thread has its miscellaneous processor state bits clear. */
192	p->thread.proc_status = 0;
193
194#ifdef CONFIG_HARDWALL
195	/* New thread does not own any networks. */
196	memset(&p->thread.hardwall[0], 0,
197	       sizeof(struct hardwall_task) * HARDWALL_TYPES);
198#endif
199
200
201	/*
202	 * Start the new thread with the current architecture state
203	 * (user interrupt masks, etc.).
204	 */
205	save_arch_state(&p->thread);
206
207	return 0;
208}
209
210int set_unalign_ctl(struct task_struct *tsk, unsigned int val)
211{
212	task_thread_info(tsk)->align_ctl = val;
213	return 0;
214}
215
216int get_unalign_ctl(struct task_struct *tsk, unsigned long adr)
217{
218	return put_user(task_thread_info(tsk)->align_ctl,
219			(unsigned int __user *)adr);
220}
221
222static struct task_struct corrupt_current = { .comm = "<corrupt>" };
223
224/*
225 * Return "current" if it looks plausible, or else a pointer to a dummy.
226 * This can be helpful if we are just trying to emit a clean panic.
227 */
228struct task_struct *validate_current(void)
229{
230	struct task_struct *tsk = current;
231	if (unlikely((unsigned long)tsk < PAGE_OFFSET ||
232		     (high_memory && (void *)tsk > high_memory) ||
233		     ((unsigned long)tsk & (__alignof__(*tsk) - 1)) != 0)) {
234		pr_err("Corrupt 'current' %p (sp %#lx)\n", tsk, stack_pointer);
235		tsk = &corrupt_current;
236	}
237	return tsk;
238}
239
240/* Take and return the pointer to the previous task, for schedule_tail(). */
241struct task_struct *sim_notify_fork(struct task_struct *prev)
242{
243	struct task_struct *tsk = current;
244	__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK_PARENT |
245		     (tsk->thread.creator_pid << _SIM_CONTROL_OPERATOR_BITS));
246	__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK |
247		     (tsk->pid << _SIM_CONTROL_OPERATOR_BITS));
248	return prev;
249}
250
251int dump_task_regs(struct task_struct *tsk, elf_gregset_t *regs)
252{
253	struct pt_regs *ptregs = task_pt_regs(tsk);
254	elf_core_copy_regs(regs, ptregs);
255	return 1;
256}
257
258#if CHIP_HAS_TILE_DMA()
259
260/* Allow user processes to access the DMA SPRs */
261void grant_dma_mpls(void)
262{
263#if CONFIG_KERNEL_PL == 2
264	__insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
265	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
266#else
267	__insn_mtspr(SPR_MPL_DMA_CPL_SET_0, 1);
268	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_0, 1);
269#endif
270}
271
272/* Forbid user processes from accessing the DMA SPRs */
273void restrict_dma_mpls(void)
274{
275#if CONFIG_KERNEL_PL == 2
276	__insn_mtspr(SPR_MPL_DMA_CPL_SET_2, 1);
277	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_2, 1);
278#else
279	__insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
280	__insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
281#endif
282}
283
284/* Pause the DMA engine, then save off its state registers. */
285static void save_tile_dma_state(struct tile_dma_state *dma)
286{
287	unsigned long state = __insn_mfspr(SPR_DMA_USER_STATUS);
288	unsigned long post_suspend_state;
289
290	/* If we're running, suspend the engine. */
291	if ((state & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK)
292		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
293
294	/*
295	 * Wait for the engine to idle, then save regs.  Note that we
296	 * want to record the "running" bit from before suspension,
297	 * and the "done" bit from after, so that we can properly
298	 * distinguish a case where the user suspended the engine from
299	 * the case where the kernel suspended as part of the context
300	 * swap.
301	 */
302	do {
303		post_suspend_state = __insn_mfspr(SPR_DMA_USER_STATUS);
304	} while (post_suspend_state & SPR_DMA_STATUS__BUSY_MASK);
305
306	dma->src = __insn_mfspr(SPR_DMA_SRC_ADDR);
307	dma->src_chunk = __insn_mfspr(SPR_DMA_SRC_CHUNK_ADDR);
308	dma->dest = __insn_mfspr(SPR_DMA_DST_ADDR);
309	dma->dest_chunk = __insn_mfspr(SPR_DMA_DST_CHUNK_ADDR);
310	dma->strides = __insn_mfspr(SPR_DMA_STRIDE);
311	dma->chunk_size = __insn_mfspr(SPR_DMA_CHUNK_SIZE);
312	dma->byte = __insn_mfspr(SPR_DMA_BYTE);
313	dma->status = (state & SPR_DMA_STATUS__RUNNING_MASK) |
314		(post_suspend_state & SPR_DMA_STATUS__DONE_MASK);
315}
316
317/* Restart a DMA that was running before we were context-switched out. */
318static void restore_tile_dma_state(struct thread_struct *t)
319{
320	const struct tile_dma_state *dma = &t->tile_dma_state;
321
322	/*
323	 * The only way to restore the done bit is to run a zero
324	 * length transaction.
325	 */
326	if ((dma->status & SPR_DMA_STATUS__DONE_MASK) &&
327	    !(__insn_mfspr(SPR_DMA_USER_STATUS) & SPR_DMA_STATUS__DONE_MASK)) {
328		__insn_mtspr(SPR_DMA_BYTE, 0);
329		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
330		while (__insn_mfspr(SPR_DMA_USER_STATUS) &
331		       SPR_DMA_STATUS__BUSY_MASK)
332			;
333	}
334
335	__insn_mtspr(SPR_DMA_SRC_ADDR, dma->src);
336	__insn_mtspr(SPR_DMA_SRC_CHUNK_ADDR, dma->src_chunk);
337	__insn_mtspr(SPR_DMA_DST_ADDR, dma->dest);
338	__insn_mtspr(SPR_DMA_DST_CHUNK_ADDR, dma->dest_chunk);
339	__insn_mtspr(SPR_DMA_STRIDE, dma->strides);
340	__insn_mtspr(SPR_DMA_CHUNK_SIZE, dma->chunk_size);
341	__insn_mtspr(SPR_DMA_BYTE, dma->byte);
342
343	/*
344	 * Restart the engine if we were running and not done.
345	 * Clear a pending async DMA fault that we were waiting on return
346	 * to user space to execute, since we expect the DMA engine
347	 * to regenerate those faults for us now.  Note that we don't
348	 * try to clear the TIF_ASYNC_TLB flag, since it's relatively
349	 * harmless if set, and it covers both DMA and the SN processor.
350	 */
351	if ((dma->status & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK) {
352		t->dma_async_tlb.fault_num = 0;
353		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
354	}
355}
356
357#endif
358
359static void save_arch_state(struct thread_struct *t)
360{
361#if CHIP_HAS_SPLIT_INTR_MASK()
362	t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0_0) |
363		((u64)__insn_mfspr(SPR_INTERRUPT_MASK_0_1) << 32);
364#else
365	t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0);
366#endif
367	t->ex_context[0] = __insn_mfspr(SPR_EX_CONTEXT_0_0);
368	t->ex_context[1] = __insn_mfspr(SPR_EX_CONTEXT_0_1);
369	t->system_save[0] = __insn_mfspr(SPR_SYSTEM_SAVE_0_0);
370	t->system_save[1] = __insn_mfspr(SPR_SYSTEM_SAVE_0_1);
371	t->system_save[2] = __insn_mfspr(SPR_SYSTEM_SAVE_0_2);
372	t->system_save[3] = __insn_mfspr(SPR_SYSTEM_SAVE_0_3);
373	t->intctrl_0 = __insn_mfspr(SPR_INTCTRL_0_STATUS);
374	t->proc_status = __insn_mfspr(SPR_PROC_STATUS);
375#if !CHIP_HAS_FIXED_INTVEC_BASE()
376	t->interrupt_vector_base = __insn_mfspr(SPR_INTERRUPT_VECTOR_BASE_0);
377#endif
378	t->tile_rtf_hwm = __insn_mfspr(SPR_TILE_RTF_HWM);
379#if CHIP_HAS_DSTREAM_PF()
380	t->dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
381#endif
382}
383
384static void restore_arch_state(const struct thread_struct *t)
385{
386#if CHIP_HAS_SPLIT_INTR_MASK()
387	__insn_mtspr(SPR_INTERRUPT_MASK_0_0, (u32) t->interrupt_mask);
388	__insn_mtspr(SPR_INTERRUPT_MASK_0_1, t->interrupt_mask >> 32);
389#else
390	__insn_mtspr(SPR_INTERRUPT_MASK_0, t->interrupt_mask);
391#endif
392	__insn_mtspr(SPR_EX_CONTEXT_0_0, t->ex_context[0]);
393	__insn_mtspr(SPR_EX_CONTEXT_0_1, t->ex_context[1]);
394	__insn_mtspr(SPR_SYSTEM_SAVE_0_0, t->system_save[0]);
395	__insn_mtspr(SPR_SYSTEM_SAVE_0_1, t->system_save[1]);
396	__insn_mtspr(SPR_SYSTEM_SAVE_0_2, t->system_save[2]);
397	__insn_mtspr(SPR_SYSTEM_SAVE_0_3, t->system_save[3]);
398	__insn_mtspr(SPR_INTCTRL_0_STATUS, t->intctrl_0);
399	__insn_mtspr(SPR_PROC_STATUS, t->proc_status);
400#if !CHIP_HAS_FIXED_INTVEC_BASE()
401	__insn_mtspr(SPR_INTERRUPT_VECTOR_BASE_0, t->interrupt_vector_base);
402#endif
403	__insn_mtspr(SPR_TILE_RTF_HWM, t->tile_rtf_hwm);
404#if CHIP_HAS_DSTREAM_PF()
405	__insn_mtspr(SPR_DSTREAM_PF, t->dstream_pf);
406#endif
407}
408
409
410void _prepare_arch_switch(struct task_struct *next)
411{
412#if CHIP_HAS_TILE_DMA()
413	struct tile_dma_state *dma = &current->thread.tile_dma_state;
414	if (dma->enabled)
415		save_tile_dma_state(dma);
416#endif
417}
418
419
420struct task_struct *__sched _switch_to(struct task_struct *prev,
421				       struct task_struct *next)
422{
423	/* DMA state is already saved; save off other arch state. */
424	save_arch_state(&prev->thread);
425
426#if CHIP_HAS_TILE_DMA()
427	/*
428	 * Restore DMA in new task if desired.
429	 * Note that it is only safe to restart here since interrupts
430	 * are disabled, so we can't take any DMATLB miss or access
431	 * interrupts before we have finished switching stacks.
432	 */
433	if (next->thread.tile_dma_state.enabled) {
434		restore_tile_dma_state(&next->thread);
435		grant_dma_mpls();
436	} else {
437		restrict_dma_mpls();
438	}
439#endif
440
441	/* Restore other arch state. */
442	restore_arch_state(&next->thread);
443
444#ifdef CONFIG_HARDWALL
445	/* Enable or disable access to the network registers appropriately. */
446	hardwall_switch_tasks(prev, next);
447#endif
448
449	/* Notify the simulator of task exit. */
450	if (unlikely(prev->state == TASK_DEAD))
451		__insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_EXIT |
452			     (prev->pid << _SIM_CONTROL_OPERATOR_BITS));
453
454	/*
455	 * Switch kernel SP, PC, and callee-saved registers.
456	 * In the context of the new task, return the old task pointer
457	 * (i.e. the task that actually called __switch_to).
458	 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp.
459	 */
460	return __switch_to(prev, next, next_current_ksp0(next));
461}
462
463/*
464 * This routine is called on return from interrupt if any of the
465 * TIF_WORK_MASK flags are set in thread_info->flags.  It is
466 * entered with interrupts disabled so we don't miss an event
467 * that modified the thread_info flags.  If any flag is set, we
468 * handle it and return, and the calling assembly code will
469 * re-disable interrupts, reload the thread flags, and call back
470 * if more flags need to be handled.
471 *
472 * We return whether we need to check the thread_info flags again
473 * or not.  Note that we don't clear TIF_SINGLESTEP here, so it's
474 * important that it be tested last, and then claim that we don't
475 * need to recheck the flags.
476 */
477int do_work_pending(struct pt_regs *regs, u32 thread_info_flags)
478{
479	/* If we enter in kernel mode, do nothing and exit the caller loop. */
480	if (!user_mode(regs))
481		return 0;
482
483	user_exit();
484
485	/* Enable interrupts; they are disabled again on return to caller. */
486	local_irq_enable();
487
488	if (thread_info_flags & _TIF_NEED_RESCHED) {
489		schedule();
490		return 1;
491	}
492#if CHIP_HAS_TILE_DMA()
493	if (thread_info_flags & _TIF_ASYNC_TLB) {
494		do_async_page_fault(regs);
495		return 1;
496	}
497#endif
498	if (thread_info_flags & _TIF_SIGPENDING) {
499		do_signal(regs);
500		return 1;
501	}
502	if (thread_info_flags & _TIF_NOTIFY_RESUME) {
503		clear_thread_flag(TIF_NOTIFY_RESUME);
504		tracehook_notify_resume(regs);
505		return 1;
506	}
507	if (thread_info_flags & _TIF_SINGLESTEP)
508		single_step_once(regs);
509
510	user_enter();
511
512	return 0;
513}
514
515unsigned long get_wchan(struct task_struct *p)
516{
517	struct KBacktraceIterator kbt;
518
519	if (!p || p == current || p->state == TASK_RUNNING)
520		return 0;
521
522	for (KBacktraceIterator_init(&kbt, p, NULL);
523	     !KBacktraceIterator_end(&kbt);
524	     KBacktraceIterator_next(&kbt)) {
525		if (!in_sched_functions(kbt.it.pc))
526			return kbt.it.pc;
527	}
528
529	return 0;
530}
531
532/* Flush thread state. */
533void flush_thread(void)
534{
535	/* Nothing */
536}
537
538/*
539 * Free current thread data structures etc..
540 */
541void exit_thread(void)
542{
543#ifdef CONFIG_HARDWALL
544	/*
545	 * Remove the task from the list of tasks that are associated
546	 * with any live hardwalls.  (If the task that is exiting held
547	 * the last reference to a hardwall fd, it would already have
548	 * been released and deactivated at this point.)
549	 */
550	hardwall_deactivate_all(current);
551#endif
552}
553
554void tile_show_regs(struct pt_regs *regs)
555{
556	int i;
557#ifdef __tilegx__
558	for (i = 0; i < 17; i++)
559		pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
560		       i, regs->regs[i], i+18, regs->regs[i+18],
561		       i+36, regs->regs[i+36]);
562	pr_err(" r17: "REGFMT" r35: "REGFMT" tp : "REGFMT"\n",
563	       regs->regs[17], regs->regs[35], regs->tp);
564	pr_err(" sp : "REGFMT" lr : "REGFMT"\n", regs->sp, regs->lr);
565#else
566	for (i = 0; i < 13; i++)
567		pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT
568		       " r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
569		       i, regs->regs[i], i+14, regs->regs[i+14],
570		       i+27, regs->regs[i+27], i+40, regs->regs[i+40]);
571	pr_err(" r13: "REGFMT" tp : "REGFMT" sp : "REGFMT" lr : "REGFMT"\n",
572	       regs->regs[13], regs->tp, regs->sp, regs->lr);
573#endif
574	pr_err(" pc : "REGFMT" ex1: %ld     faultnum: %ld flags:%s%s%s%s\n",
575	       regs->pc, regs->ex1, regs->faultnum,
576	       is_compat_task() ? " compat" : "",
577	       (regs->flags & PT_FLAGS_DISABLE_IRQ) ? " noirq" : "",
578	       !(regs->flags & PT_FLAGS_CALLER_SAVES) ? " nocallersave" : "",
579	       (regs->flags & PT_FLAGS_RESTORE_REGS) ? " restoreregs" : "");
580}
581
582void show_regs(struct pt_regs *regs)
583{
584	struct KBacktraceIterator kbt;
585
586	show_regs_print_info(KERN_DEFAULT);
587	tile_show_regs(regs);
588
589	KBacktraceIterator_init(&kbt, NULL, regs);
590	tile_show_stack(&kbt);
591}
592
593/* To ensure stack dump on tiles occurs one by one. */
594static DEFINE_SPINLOCK(backtrace_lock);
595/* To ensure no backtrace occurs before all of the stack dump are done. */
596static atomic_t backtrace_cpus;
597/* The cpu mask to avoid reentrance. */
598static struct cpumask backtrace_mask;
599
600void do_nmi_dump_stack(struct pt_regs *regs)
601{
602	int is_idle = is_idle_task(current) && !in_interrupt();
603	int cpu;
604
605	nmi_enter();
606	cpu = smp_processor_id();
607	if (WARN_ON_ONCE(!cpumask_test_and_clear_cpu(cpu, &backtrace_mask)))
608		goto done;
609
610	spin_lock(&backtrace_lock);
611	if (is_idle)
612		pr_info("CPU: %d idle\n", cpu);
613	else
614		show_regs(regs);
615	spin_unlock(&backtrace_lock);
616	atomic_dec(&backtrace_cpus);
617done:
618	nmi_exit();
619}
620
621#ifdef __tilegx__
622void arch_trigger_all_cpu_backtrace(bool self)
623{
624	struct cpumask mask;
625	HV_Coord tile;
626	unsigned int timeout;
627	int cpu;
628	int ongoing;
629	HV_NMI_Info info[NR_CPUS];
630
631	ongoing = atomic_cmpxchg(&backtrace_cpus, 0, num_online_cpus() - 1);
632	if (ongoing != 0) {
633		pr_err("Trying to do all-cpu backtrace.\n");
634		pr_err("But another all-cpu backtrace is ongoing (%d cpus left)\n",
635		       ongoing);
636		if (self) {
637			pr_err("Reporting the stack on this cpu only.\n");
638			dump_stack();
639		}
640		return;
641	}
642
643	cpumask_copy(&mask, cpu_online_mask);
644	cpumask_clear_cpu(smp_processor_id(), &mask);
645	cpumask_copy(&backtrace_mask, &mask);
646
647	/* Backtrace for myself first. */
648	if (self)
649		dump_stack();
650
651	/* Tentatively dump stack on remote tiles via NMI. */
652	timeout = 100;
653	while (!cpumask_empty(&mask) && timeout) {
654		for_each_cpu(cpu, &mask) {
655			tile.x = cpu_x(cpu);
656			tile.y = cpu_y(cpu);
657			info[cpu] = hv_send_nmi(tile, TILE_NMI_DUMP_STACK, 0);
658			if (info[cpu].result == HV_NMI_RESULT_OK)
659				cpumask_clear_cpu(cpu, &mask);
660		}
661
662		mdelay(10);
663		timeout--;
664	}
665
666	/* Warn about cpus stuck in ICS and decrement their counts here. */
667	if (!cpumask_empty(&mask)) {
668		for_each_cpu(cpu, &mask) {
669			switch (info[cpu].result) {
670			case HV_NMI_RESULT_FAIL_ICS:
671				pr_warn("Skipping stack dump of cpu %d in ICS at pc %#llx\n",
672					cpu, info[cpu].pc);
673				break;
674			case HV_NMI_RESULT_FAIL_HV:
675				pr_warn("Skipping stack dump of cpu %d in hypervisor\n",
676					cpu);
677				break;
678			case HV_ENOSYS:
679				pr_warn("Hypervisor too old to allow remote stack dumps.\n");
680				goto skip_for_each;
681			default:  /* should not happen */
682				pr_warn("Skipping stack dump of cpu %d [%d,%#llx]\n",
683					cpu, info[cpu].result, info[cpu].pc);
684				break;
685			}
686		}
687skip_for_each:
688		atomic_sub(cpumask_weight(&mask), &backtrace_cpus);
689	}
690}
691#endif /* __tilegx_ */
692