Loading...
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 <asm/system.h>
31#include <asm/stack.h>
32#include <asm/homecache.h>
33#include <asm/syscalls.h>
34#include <asm/traps.h>
35#ifdef CONFIG_HARDWALL
36#include <asm/hardwall.h>
37#endif
38#include <arch/chip.h>
39#include <arch/abi.h>
40
41
42/*
43 * Use the (x86) "idle=poll" option to prefer low latency when leaving the
44 * idle loop over low power while in the idle loop, e.g. if we have
45 * one thread per core and we want to get threads out of futex waits fast.
46 */
47static int no_idle_nap;
48static int __init idle_setup(char *str)
49{
50 if (!str)
51 return -EINVAL;
52
53 if (!strcmp(str, "poll")) {
54 pr_info("using polling idle threads.\n");
55 no_idle_nap = 1;
56 } else if (!strcmp(str, "halt"))
57 no_idle_nap = 0;
58 else
59 return -1;
60
61 return 0;
62}
63early_param("idle", idle_setup);
64
65/*
66 * The idle thread. There's no useful work to be
67 * done, so just try to conserve power and have a
68 * low exit latency (ie sit in a loop waiting for
69 * somebody to say that they'd like to reschedule)
70 */
71void cpu_idle(void)
72{
73 int cpu = smp_processor_id();
74
75
76 current_thread_info()->status |= TS_POLLING;
77
78 if (no_idle_nap) {
79 while (1) {
80 while (!need_resched())
81 cpu_relax();
82 schedule();
83 }
84 }
85
86 /* endless idle loop with no priority at all */
87 while (1) {
88 tick_nohz_stop_sched_tick(1);
89 while (!need_resched()) {
90 if (cpu_is_offline(cpu))
91 BUG(); /* no HOTPLUG_CPU */
92
93 local_irq_disable();
94 __get_cpu_var(irq_stat).idle_timestamp = jiffies;
95 current_thread_info()->status &= ~TS_POLLING;
96 /*
97 * TS_POLLING-cleared state must be visible before we
98 * test NEED_RESCHED:
99 */
100 smp_mb();
101
102 if (!need_resched())
103 _cpu_idle();
104 else
105 local_irq_enable();
106 current_thread_info()->status |= TS_POLLING;
107 }
108 tick_nohz_restart_sched_tick();
109 preempt_enable_no_resched();
110 schedule();
111 preempt_disable();
112 }
113}
114
115struct thread_info *alloc_thread_info_node(struct task_struct *task, int node)
116{
117 struct page *page;
118 gfp_t flags = GFP_KERNEL;
119
120#ifdef CONFIG_DEBUG_STACK_USAGE
121 flags |= __GFP_ZERO;
122#endif
123
124 page = alloc_pages_node(node, flags, THREAD_SIZE_ORDER);
125 if (!page)
126 return NULL;
127
128 return (struct thread_info *)page_address(page);
129}
130
131/*
132 * Free a thread_info node, and all of its derivative
133 * data structures.
134 */
135void free_thread_info(struct thread_info *info)
136{
137 struct single_step_state *step_state = info->step_state;
138
139#ifdef CONFIG_HARDWALL
140 /*
141 * We free a thread_info from the context of the task that has
142 * been scheduled next, so the original task is already dead.
143 * Calling deactivate here just frees up the data structures.
144 * If the task we're freeing held the last reference to a
145 * hardwall fd, it would have been released prior to this point
146 * anyway via exit_files(), and "hardwall" would be NULL by now.
147 */
148 if (info->task->thread.hardwall)
149 hardwall_deactivate(info->task);
150#endif
151
152 if (step_state) {
153
154 /*
155 * FIXME: we don't munmap step_state->buffer
156 * because the mm_struct for this process (info->task->mm)
157 * has already been zeroed in exit_mm(). Keeping a
158 * reference to it here seems like a bad move, so this
159 * means we can't munmap() the buffer, and therefore if we
160 * ptrace multiple threads in a process, we will slowly
161 * leak user memory. (Note that as soon as the last
162 * thread in a process dies, we will reclaim all user
163 * memory including single-step buffers in the usual way.)
164 * We should either assign a kernel VA to this buffer
165 * somehow, or we should associate the buffer(s) with the
166 * mm itself so we can clean them up that way.
167 */
168 kfree(step_state);
169 }
170
171 free_pages((unsigned long)info, THREAD_SIZE_ORDER);
172}
173
174static void save_arch_state(struct thread_struct *t);
175
176int copy_thread(unsigned long clone_flags, unsigned long sp,
177 unsigned long stack_size,
178 struct task_struct *p, struct pt_regs *regs)
179{
180 struct pt_regs *childregs;
181 unsigned long ksp;
182
183 /*
184 * When creating a new kernel thread we pass sp as zero.
185 * Assign it to a reasonable value now that we have the stack.
186 */
187 if (sp == 0 && regs->ex1 == PL_ICS_EX1(KERNEL_PL, 0))
188 sp = KSTK_TOP(p);
189
190 /*
191 * Do not clone step state from the parent; each thread
192 * must make its own lazily.
193 */
194 task_thread_info(p)->step_state = NULL;
195
196 /*
197 * Start new thread in ret_from_fork so it schedules properly
198 * and then return from interrupt like the parent.
199 */
200 p->thread.pc = (unsigned long) ret_from_fork;
201
202 /* Save user stack top pointer so we can ID the stack vm area later. */
203 p->thread.usp0 = sp;
204
205 /* Record the pid of the process that created this one. */
206 p->thread.creator_pid = current->pid;
207
208 /*
209 * Copy the registers onto the kernel stack so the
210 * return-from-interrupt code will reload it into registers.
211 */
212 childregs = task_pt_regs(p);
213 *childregs = *regs;
214 childregs->regs[0] = 0; /* return value is zero */
215 childregs->sp = sp; /* override with new user stack pointer */
216
217 /*
218 * If CLONE_SETTLS is set, set "tp" in the new task to "r4",
219 * which is passed in as arg #5 to sys_clone().
220 */
221 if (clone_flags & CLONE_SETTLS)
222 childregs->tp = regs->regs[4];
223
224 /*
225 * Copy the callee-saved registers from the passed pt_regs struct
226 * into the context-switch callee-saved registers area.
227 * This way when we start the interrupt-return sequence, the
228 * callee-save registers will be correctly in registers, which
229 * is how we assume the compiler leaves them as we start doing
230 * the normal return-from-interrupt path after calling C code.
231 * Zero out the C ABI save area to mark the top of the stack.
232 */
233 ksp = (unsigned long) childregs;
234 ksp -= C_ABI_SAVE_AREA_SIZE; /* interrupt-entry save area */
235 ((long *)ksp)[0] = ((long *)ksp)[1] = 0;
236 ksp -= CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long);
237 memcpy((void *)ksp, ®s->regs[CALLEE_SAVED_FIRST_REG],
238 CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long));
239 ksp -= C_ABI_SAVE_AREA_SIZE; /* __switch_to() save area */
240 ((long *)ksp)[0] = ((long *)ksp)[1] = 0;
241 p->thread.ksp = ksp;
242
243#if CHIP_HAS_TILE_DMA()
244 /*
245 * No DMA in the new thread. We model this on the fact that
246 * fork() clears the pending signals, alarms, and aio for the child.
247 */
248 memset(&p->thread.tile_dma_state, 0, sizeof(struct tile_dma_state));
249 memset(&p->thread.dma_async_tlb, 0, sizeof(struct async_tlb));
250#endif
251
252#if CHIP_HAS_SN_PROC()
253 /* Likewise, the new thread is not running static processor code. */
254 p->thread.sn_proc_running = 0;
255 memset(&p->thread.sn_async_tlb, 0, sizeof(struct async_tlb));
256#endif
257
258#if CHIP_HAS_PROC_STATUS_SPR()
259 /* New thread has its miscellaneous processor state bits clear. */
260 p->thread.proc_status = 0;
261#endif
262
263#ifdef CONFIG_HARDWALL
264 /* New thread does not own any networks. */
265 p->thread.hardwall = NULL;
266#endif
267
268
269 /*
270 * Start the new thread with the current architecture state
271 * (user interrupt masks, etc.).
272 */
273 save_arch_state(&p->thread);
274
275 return 0;
276}
277
278/*
279 * Return "current" if it looks plausible, or else a pointer to a dummy.
280 * This can be helpful if we are just trying to emit a clean panic.
281 */
282struct task_struct *validate_current(void)
283{
284 static struct task_struct corrupt = { .comm = "<corrupt>" };
285 struct task_struct *tsk = current;
286 if (unlikely((unsigned long)tsk < PAGE_OFFSET ||
287 (void *)tsk > high_memory ||
288 ((unsigned long)tsk & (__alignof__(*tsk) - 1)) != 0)) {
289 pr_err("Corrupt 'current' %p (sp %#lx)\n", tsk, stack_pointer);
290 tsk = &corrupt;
291 }
292 return tsk;
293}
294
295/* Take and return the pointer to the previous task, for schedule_tail(). */
296struct task_struct *sim_notify_fork(struct task_struct *prev)
297{
298 struct task_struct *tsk = current;
299 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK_PARENT |
300 (tsk->thread.creator_pid << _SIM_CONTROL_OPERATOR_BITS));
301 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK |
302 (tsk->pid << _SIM_CONTROL_OPERATOR_BITS));
303 return prev;
304}
305
306int dump_task_regs(struct task_struct *tsk, elf_gregset_t *regs)
307{
308 struct pt_regs *ptregs = task_pt_regs(tsk);
309 elf_core_copy_regs(regs, ptregs);
310 return 1;
311}
312
313#if CHIP_HAS_TILE_DMA()
314
315/* Allow user processes to access the DMA SPRs */
316void grant_dma_mpls(void)
317{
318#if CONFIG_KERNEL_PL == 2
319 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
320 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
321#else
322 __insn_mtspr(SPR_MPL_DMA_CPL_SET_0, 1);
323 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_0, 1);
324#endif
325}
326
327/* Forbid user processes from accessing the DMA SPRs */
328void restrict_dma_mpls(void)
329{
330#if CONFIG_KERNEL_PL == 2
331 __insn_mtspr(SPR_MPL_DMA_CPL_SET_2, 1);
332 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_2, 1);
333#else
334 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
335 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
336#endif
337}
338
339/* Pause the DMA engine, then save off its state registers. */
340static void save_tile_dma_state(struct tile_dma_state *dma)
341{
342 unsigned long state = __insn_mfspr(SPR_DMA_USER_STATUS);
343 unsigned long post_suspend_state;
344
345 /* If we're running, suspend the engine. */
346 if ((state & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK)
347 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
348
349 /*
350 * Wait for the engine to idle, then save regs. Note that we
351 * want to record the "running" bit from before suspension,
352 * and the "done" bit from after, so that we can properly
353 * distinguish a case where the user suspended the engine from
354 * the case where the kernel suspended as part of the context
355 * swap.
356 */
357 do {
358 post_suspend_state = __insn_mfspr(SPR_DMA_USER_STATUS);
359 } while (post_suspend_state & SPR_DMA_STATUS__BUSY_MASK);
360
361 dma->src = __insn_mfspr(SPR_DMA_SRC_ADDR);
362 dma->src_chunk = __insn_mfspr(SPR_DMA_SRC_CHUNK_ADDR);
363 dma->dest = __insn_mfspr(SPR_DMA_DST_ADDR);
364 dma->dest_chunk = __insn_mfspr(SPR_DMA_DST_CHUNK_ADDR);
365 dma->strides = __insn_mfspr(SPR_DMA_STRIDE);
366 dma->chunk_size = __insn_mfspr(SPR_DMA_CHUNK_SIZE);
367 dma->byte = __insn_mfspr(SPR_DMA_BYTE);
368 dma->status = (state & SPR_DMA_STATUS__RUNNING_MASK) |
369 (post_suspend_state & SPR_DMA_STATUS__DONE_MASK);
370}
371
372/* Restart a DMA that was running before we were context-switched out. */
373static void restore_tile_dma_state(struct thread_struct *t)
374{
375 const struct tile_dma_state *dma = &t->tile_dma_state;
376
377 /*
378 * The only way to restore the done bit is to run a zero
379 * length transaction.
380 */
381 if ((dma->status & SPR_DMA_STATUS__DONE_MASK) &&
382 !(__insn_mfspr(SPR_DMA_USER_STATUS) & SPR_DMA_STATUS__DONE_MASK)) {
383 __insn_mtspr(SPR_DMA_BYTE, 0);
384 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
385 while (__insn_mfspr(SPR_DMA_USER_STATUS) &
386 SPR_DMA_STATUS__BUSY_MASK)
387 ;
388 }
389
390 __insn_mtspr(SPR_DMA_SRC_ADDR, dma->src);
391 __insn_mtspr(SPR_DMA_SRC_CHUNK_ADDR, dma->src_chunk);
392 __insn_mtspr(SPR_DMA_DST_ADDR, dma->dest);
393 __insn_mtspr(SPR_DMA_DST_CHUNK_ADDR, dma->dest_chunk);
394 __insn_mtspr(SPR_DMA_STRIDE, dma->strides);
395 __insn_mtspr(SPR_DMA_CHUNK_SIZE, dma->chunk_size);
396 __insn_mtspr(SPR_DMA_BYTE, dma->byte);
397
398 /*
399 * Restart the engine if we were running and not done.
400 * Clear a pending async DMA fault that we were waiting on return
401 * to user space to execute, since we expect the DMA engine
402 * to regenerate those faults for us now. Note that we don't
403 * try to clear the TIF_ASYNC_TLB flag, since it's relatively
404 * harmless if set, and it covers both DMA and the SN processor.
405 */
406 if ((dma->status & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK) {
407 t->dma_async_tlb.fault_num = 0;
408 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
409 }
410}
411
412#endif
413
414static void save_arch_state(struct thread_struct *t)
415{
416#if CHIP_HAS_SPLIT_INTR_MASK()
417 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0_0) |
418 ((u64)__insn_mfspr(SPR_INTERRUPT_MASK_0_1) << 32);
419#else
420 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0);
421#endif
422 t->ex_context[0] = __insn_mfspr(SPR_EX_CONTEXT_0_0);
423 t->ex_context[1] = __insn_mfspr(SPR_EX_CONTEXT_0_1);
424 t->system_save[0] = __insn_mfspr(SPR_SYSTEM_SAVE_0_0);
425 t->system_save[1] = __insn_mfspr(SPR_SYSTEM_SAVE_0_1);
426 t->system_save[2] = __insn_mfspr(SPR_SYSTEM_SAVE_0_2);
427 t->system_save[3] = __insn_mfspr(SPR_SYSTEM_SAVE_0_3);
428 t->intctrl_0 = __insn_mfspr(SPR_INTCTRL_0_STATUS);
429#if CHIP_HAS_PROC_STATUS_SPR()
430 t->proc_status = __insn_mfspr(SPR_PROC_STATUS);
431#endif
432#if !CHIP_HAS_FIXED_INTVEC_BASE()
433 t->interrupt_vector_base = __insn_mfspr(SPR_INTERRUPT_VECTOR_BASE_0);
434#endif
435#if CHIP_HAS_TILE_RTF_HWM()
436 t->tile_rtf_hwm = __insn_mfspr(SPR_TILE_RTF_HWM);
437#endif
438#if CHIP_HAS_DSTREAM_PF()
439 t->dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
440#endif
441}
442
443static void restore_arch_state(const struct thread_struct *t)
444{
445#if CHIP_HAS_SPLIT_INTR_MASK()
446 __insn_mtspr(SPR_INTERRUPT_MASK_0_0, (u32) t->interrupt_mask);
447 __insn_mtspr(SPR_INTERRUPT_MASK_0_1, t->interrupt_mask >> 32);
448#else
449 __insn_mtspr(SPR_INTERRUPT_MASK_0, t->interrupt_mask);
450#endif
451 __insn_mtspr(SPR_EX_CONTEXT_0_0, t->ex_context[0]);
452 __insn_mtspr(SPR_EX_CONTEXT_0_1, t->ex_context[1]);
453 __insn_mtspr(SPR_SYSTEM_SAVE_0_0, t->system_save[0]);
454 __insn_mtspr(SPR_SYSTEM_SAVE_0_1, t->system_save[1]);
455 __insn_mtspr(SPR_SYSTEM_SAVE_0_2, t->system_save[2]);
456 __insn_mtspr(SPR_SYSTEM_SAVE_0_3, t->system_save[3]);
457 __insn_mtspr(SPR_INTCTRL_0_STATUS, t->intctrl_0);
458#if CHIP_HAS_PROC_STATUS_SPR()
459 __insn_mtspr(SPR_PROC_STATUS, t->proc_status);
460#endif
461#if !CHIP_HAS_FIXED_INTVEC_BASE()
462 __insn_mtspr(SPR_INTERRUPT_VECTOR_BASE_0, t->interrupt_vector_base);
463#endif
464#if CHIP_HAS_TILE_RTF_HWM()
465 __insn_mtspr(SPR_TILE_RTF_HWM, t->tile_rtf_hwm);
466#endif
467#if CHIP_HAS_DSTREAM_PF()
468 __insn_mtspr(SPR_DSTREAM_PF, t->dstream_pf);
469#endif
470}
471
472
473void _prepare_arch_switch(struct task_struct *next)
474{
475#if CHIP_HAS_SN_PROC()
476 int snctl;
477#endif
478#if CHIP_HAS_TILE_DMA()
479 struct tile_dma_state *dma = ¤t->thread.tile_dma_state;
480 if (dma->enabled)
481 save_tile_dma_state(dma);
482#endif
483#if CHIP_HAS_SN_PROC()
484 /*
485 * Suspend the static network processor if it was running.
486 * We do not suspend the fabric itself, just like we don't
487 * try to suspend the UDN.
488 */
489 snctl = __insn_mfspr(SPR_SNCTL);
490 current->thread.sn_proc_running =
491 (snctl & SPR_SNCTL__FRZPROC_MASK) == 0;
492 if (current->thread.sn_proc_running)
493 __insn_mtspr(SPR_SNCTL, snctl | SPR_SNCTL__FRZPROC_MASK);
494#endif
495}
496
497
498struct task_struct *__sched _switch_to(struct task_struct *prev,
499 struct task_struct *next)
500{
501 /* DMA state is already saved; save off other arch state. */
502 save_arch_state(&prev->thread);
503
504#if CHIP_HAS_TILE_DMA()
505 /*
506 * Restore DMA in new task if desired.
507 * Note that it is only safe to restart here since interrupts
508 * are disabled, so we can't take any DMATLB miss or access
509 * interrupts before we have finished switching stacks.
510 */
511 if (next->thread.tile_dma_state.enabled) {
512 restore_tile_dma_state(&next->thread);
513 grant_dma_mpls();
514 } else {
515 restrict_dma_mpls();
516 }
517#endif
518
519 /* Restore other arch state. */
520 restore_arch_state(&next->thread);
521
522#if CHIP_HAS_SN_PROC()
523 /*
524 * Restart static network processor in the new process
525 * if it was running before.
526 */
527 if (next->thread.sn_proc_running) {
528 int snctl = __insn_mfspr(SPR_SNCTL);
529 __insn_mtspr(SPR_SNCTL, snctl & ~SPR_SNCTL__FRZPROC_MASK);
530 }
531#endif
532
533#ifdef CONFIG_HARDWALL
534 /* Enable or disable access to the network registers appropriately. */
535 if (prev->thread.hardwall != NULL) {
536 if (next->thread.hardwall == NULL)
537 restrict_network_mpls();
538 } else if (next->thread.hardwall != NULL) {
539 grant_network_mpls();
540 }
541#endif
542
543 /*
544 * Switch kernel SP, PC, and callee-saved registers.
545 * In the context of the new task, return the old task pointer
546 * (i.e. the task that actually called __switch_to).
547 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp.
548 */
549 return __switch_to(prev, next, next_current_ksp0(next));
550}
551
552/*
553 * This routine is called on return from interrupt if any of the
554 * TIF_WORK_MASK flags are set in thread_info->flags. It is
555 * entered with interrupts disabled so we don't miss an event
556 * that modified the thread_info flags. If any flag is set, we
557 * handle it and return, and the calling assembly code will
558 * re-disable interrupts, reload the thread flags, and call back
559 * if more flags need to be handled.
560 *
561 * We return whether we need to check the thread_info flags again
562 * or not. Note that we don't clear TIF_SINGLESTEP here, so it's
563 * important that it be tested last, and then claim that we don't
564 * need to recheck the flags.
565 */
566int do_work_pending(struct pt_regs *regs, u32 thread_info_flags)
567{
568 if (thread_info_flags & _TIF_NEED_RESCHED) {
569 schedule();
570 return 1;
571 }
572#if CHIP_HAS_TILE_DMA() || CHIP_HAS_SN_PROC()
573 if (thread_info_flags & _TIF_ASYNC_TLB) {
574 do_async_page_fault(regs);
575 return 1;
576 }
577#endif
578 if (thread_info_flags & _TIF_SIGPENDING) {
579 do_signal(regs);
580 return 1;
581 }
582 if (thread_info_flags & _TIF_NOTIFY_RESUME) {
583 clear_thread_flag(TIF_NOTIFY_RESUME);
584 tracehook_notify_resume(regs);
585 if (current->replacement_session_keyring)
586 key_replace_session_keyring();
587 return 1;
588 }
589 if (thread_info_flags & _TIF_SINGLESTEP) {
590 if ((regs->ex1 & SPR_EX_CONTEXT_1_1__PL_MASK) == 0)
591 single_step_once(regs);
592 return 0;
593 }
594 panic("work_pending: bad flags %#x\n", thread_info_flags);
595}
596
597/* Note there is an implicit fifth argument if (clone_flags & CLONE_SETTLS). */
598SYSCALL_DEFINE5(clone, unsigned long, clone_flags, unsigned long, newsp,
599 void __user *, parent_tidptr, void __user *, child_tidptr,
600 struct pt_regs *, regs)
601{
602 if (!newsp)
603 newsp = regs->sp;
604 return do_fork(clone_flags, newsp, regs, 0,
605 parent_tidptr, child_tidptr);
606}
607
608/*
609 * sys_execve() executes a new program.
610 */
611SYSCALL_DEFINE4(execve, const char __user *, path,
612 const char __user *const __user *, argv,
613 const char __user *const __user *, envp,
614 struct pt_regs *, regs)
615{
616 long error;
617 char *filename;
618
619 filename = getname(path);
620 error = PTR_ERR(filename);
621 if (IS_ERR(filename))
622 goto out;
623 error = do_execve(filename, argv, envp, regs);
624 putname(filename);
625 if (error == 0)
626 single_step_execve();
627out:
628 return error;
629}
630
631#ifdef CONFIG_COMPAT
632long compat_sys_execve(const char __user *path,
633 compat_uptr_t __user *argv,
634 compat_uptr_t __user *envp,
635 struct pt_regs *regs)
636{
637 long error;
638 char *filename;
639
640 filename = getname(path);
641 error = PTR_ERR(filename);
642 if (IS_ERR(filename))
643 goto out;
644 error = compat_do_execve(filename, argv, envp, regs);
645 putname(filename);
646 if (error == 0)
647 single_step_execve();
648out:
649 return error;
650}
651#endif
652
653unsigned long get_wchan(struct task_struct *p)
654{
655 struct KBacktraceIterator kbt;
656
657 if (!p || p == current || p->state == TASK_RUNNING)
658 return 0;
659
660 for (KBacktraceIterator_init(&kbt, p, NULL);
661 !KBacktraceIterator_end(&kbt);
662 KBacktraceIterator_next(&kbt)) {
663 if (!in_sched_functions(kbt.it.pc))
664 return kbt.it.pc;
665 }
666
667 return 0;
668}
669
670/*
671 * We pass in lr as zero (cleared in kernel_thread) and the caller
672 * part of the backtrace ABI on the stack also zeroed (in copy_thread)
673 * so that backtraces will stop with this function.
674 * Note that we don't use r0, since copy_thread() clears it.
675 */
676static void start_kernel_thread(int dummy, int (*fn)(int), int arg)
677{
678 do_exit(fn(arg));
679}
680
681/*
682 * Create a kernel thread
683 */
684int kernel_thread(int (*fn)(void *), void * arg, unsigned long flags)
685{
686 struct pt_regs regs;
687
688 memset(®s, 0, sizeof(regs));
689 regs.ex1 = PL_ICS_EX1(KERNEL_PL, 0); /* run at kernel PL, no ICS */
690 regs.pc = (long) start_kernel_thread;
691 regs.flags = PT_FLAGS_CALLER_SAVES; /* need to restore r1 and r2 */
692 regs.regs[1] = (long) fn; /* function pointer */
693 regs.regs[2] = (long) arg; /* parameter register */
694
695 /* Ok, create the new process.. */
696 return do_fork(flags | CLONE_VM | CLONE_UNTRACED, 0, ®s,
697 0, NULL, NULL);
698}
699EXPORT_SYMBOL(kernel_thread);
700
701/* Flush thread state. */
702void flush_thread(void)
703{
704 /* Nothing */
705}
706
707/*
708 * Free current thread data structures etc..
709 */
710void exit_thread(void)
711{
712 /* Nothing */
713}
714
715void show_regs(struct pt_regs *regs)
716{
717 struct task_struct *tsk = validate_current();
718 int i;
719
720 pr_err("\n");
721 pr_err(" Pid: %d, comm: %20s, CPU: %d\n",
722 tsk->pid, tsk->comm, smp_processor_id());
723#ifdef __tilegx__
724 for (i = 0; i < 51; i += 3)
725 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
726 i, regs->regs[i], i+1, regs->regs[i+1],
727 i+2, regs->regs[i+2]);
728 pr_err(" r51: "REGFMT" r52: "REGFMT" tp : "REGFMT"\n",
729 regs->regs[51], regs->regs[52], regs->tp);
730 pr_err(" sp : "REGFMT" lr : "REGFMT"\n", regs->sp, regs->lr);
731#else
732 for (i = 0; i < 52; i += 4)
733 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT
734 " r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
735 i, regs->regs[i], i+1, regs->regs[i+1],
736 i+2, regs->regs[i+2], i+3, regs->regs[i+3]);
737 pr_err(" r52: "REGFMT" tp : "REGFMT" sp : "REGFMT" lr : "REGFMT"\n",
738 regs->regs[52], regs->tp, regs->sp, regs->lr);
739#endif
740 pr_err(" pc : "REGFMT" ex1: %ld faultnum: %ld\n",
741 regs->pc, regs->ex1, regs->faultnum);
742
743 dump_stack_regs(regs);
744}
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 = ¤t->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_ALLWORK_MASK flags are set in thread_info->flags. It is
466 * entered with interrupts disabled so we don't miss an event that
467 * modified the thread_info flags. We loop until all the tested flags
468 * are clear. Note that the function is called on certain conditions
469 * that are not listed in the loop condition here (e.g. SINGLESTEP)
470 * which guarantees we will do those things once, and redo them if any
471 * of the other work items is re-done, but won't continue looping if
472 * all the other work is done.
473 */
474void prepare_exit_to_usermode(struct pt_regs *regs, u32 thread_info_flags)
475{
476 if (WARN_ON(!user_mode(regs)))
477 return;
478
479 do {
480 local_irq_enable();
481
482 if (thread_info_flags & _TIF_NEED_RESCHED)
483 schedule();
484
485#if CHIP_HAS_TILE_DMA()
486 if (thread_info_flags & _TIF_ASYNC_TLB)
487 do_async_page_fault(regs);
488#endif
489
490 if (thread_info_flags & _TIF_SIGPENDING)
491 do_signal(regs);
492
493 if (thread_info_flags & _TIF_NOTIFY_RESUME) {
494 clear_thread_flag(TIF_NOTIFY_RESUME);
495 tracehook_notify_resume(regs);
496 }
497
498 local_irq_disable();
499 thread_info_flags = READ_ONCE(current_thread_info()->flags);
500
501 } while (thread_info_flags & _TIF_WORK_MASK);
502
503 if (thread_info_flags & _TIF_SINGLESTEP) {
504 single_step_once(regs);
505#ifndef __tilegx__
506 /*
507 * FIXME: on tilepro, since we enable interrupts in
508 * this routine, it's possible that we miss a signal
509 * or other asynchronous event.
510 */
511 local_irq_disable();
512#endif
513 }
514
515 user_enter();
516}
517
518unsigned long get_wchan(struct task_struct *p)
519{
520 struct KBacktraceIterator kbt;
521
522 if (!p || p == current || p->state == TASK_RUNNING)
523 return 0;
524
525 for (KBacktraceIterator_init(&kbt, p, NULL);
526 !KBacktraceIterator_end(&kbt);
527 KBacktraceIterator_next(&kbt)) {
528 if (!in_sched_functions(kbt.it.pc))
529 return kbt.it.pc;
530 }
531
532 return 0;
533}
534
535/* Flush thread state. */
536void flush_thread(void)
537{
538 /* Nothing */
539}
540
541/*
542 * Free current thread data structures etc..
543 */
544void exit_thread(void)
545{
546#ifdef CONFIG_HARDWALL
547 /*
548 * Remove the task from the list of tasks that are associated
549 * with any live hardwalls. (If the task that is exiting held
550 * the last reference to a hardwall fd, it would already have
551 * been released and deactivated at this point.)
552 */
553 hardwall_deactivate_all(current);
554#endif
555}
556
557void tile_show_regs(struct pt_regs *regs)
558{
559 int i;
560#ifdef __tilegx__
561 for (i = 0; i < 17; i++)
562 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
563 i, regs->regs[i], i+18, regs->regs[i+18],
564 i+36, regs->regs[i+36]);
565 pr_err(" r17: "REGFMT" r35: "REGFMT" tp : "REGFMT"\n",
566 regs->regs[17], regs->regs[35], regs->tp);
567 pr_err(" sp : "REGFMT" lr : "REGFMT"\n", regs->sp, regs->lr);
568#else
569 for (i = 0; i < 13; i++)
570 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT
571 " r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
572 i, regs->regs[i], i+14, regs->regs[i+14],
573 i+27, regs->regs[i+27], i+40, regs->regs[i+40]);
574 pr_err(" r13: "REGFMT" tp : "REGFMT" sp : "REGFMT" lr : "REGFMT"\n",
575 regs->regs[13], regs->tp, regs->sp, regs->lr);
576#endif
577 pr_err(" pc : "REGFMT" ex1: %ld faultnum: %ld flags:%s%s%s%s\n",
578 regs->pc, regs->ex1, regs->faultnum,
579 is_compat_task() ? " compat" : "",
580 (regs->flags & PT_FLAGS_DISABLE_IRQ) ? " noirq" : "",
581 !(regs->flags & PT_FLAGS_CALLER_SAVES) ? " nocallersave" : "",
582 (regs->flags & PT_FLAGS_RESTORE_REGS) ? " restoreregs" : "");
583}
584
585void show_regs(struct pt_regs *regs)
586{
587 struct KBacktraceIterator kbt;
588
589 show_regs_print_info(KERN_DEFAULT);
590 tile_show_regs(regs);
591
592 KBacktraceIterator_init(&kbt, NULL, regs);
593 tile_show_stack(&kbt);
594}
595
596/* To ensure stack dump on tiles occurs one by one. */
597static DEFINE_SPINLOCK(backtrace_lock);
598/* To ensure no backtrace occurs before all of the stack dump are done. */
599static atomic_t backtrace_cpus;
600/* The cpu mask to avoid reentrance. */
601static struct cpumask backtrace_mask;
602
603void do_nmi_dump_stack(struct pt_regs *regs)
604{
605 int is_idle = is_idle_task(current) && !in_interrupt();
606 int cpu;
607
608 nmi_enter();
609 cpu = smp_processor_id();
610 if (WARN_ON_ONCE(!cpumask_test_and_clear_cpu(cpu, &backtrace_mask)))
611 goto done;
612
613 spin_lock(&backtrace_lock);
614 if (is_idle)
615 pr_info("CPU: %d idle\n", cpu);
616 else
617 show_regs(regs);
618 spin_unlock(&backtrace_lock);
619 atomic_dec(&backtrace_cpus);
620done:
621 nmi_exit();
622}
623
624#ifdef __tilegx__
625void arch_trigger_all_cpu_backtrace(bool self)
626{
627 struct cpumask mask;
628 HV_Coord tile;
629 unsigned int timeout;
630 int cpu;
631 int ongoing;
632 HV_NMI_Info info[NR_CPUS];
633
634 ongoing = atomic_cmpxchg(&backtrace_cpus, 0, num_online_cpus() - 1);
635 if (ongoing != 0) {
636 pr_err("Trying to do all-cpu backtrace.\n");
637 pr_err("But another all-cpu backtrace is ongoing (%d cpus left)\n",
638 ongoing);
639 if (self) {
640 pr_err("Reporting the stack on this cpu only.\n");
641 dump_stack();
642 }
643 return;
644 }
645
646 cpumask_copy(&mask, cpu_online_mask);
647 cpumask_clear_cpu(smp_processor_id(), &mask);
648 cpumask_copy(&backtrace_mask, &mask);
649
650 /* Backtrace for myself first. */
651 if (self)
652 dump_stack();
653
654 /* Tentatively dump stack on remote tiles via NMI. */
655 timeout = 100;
656 while (!cpumask_empty(&mask) && timeout) {
657 for_each_cpu(cpu, &mask) {
658 tile.x = cpu_x(cpu);
659 tile.y = cpu_y(cpu);
660 info[cpu] = hv_send_nmi(tile, TILE_NMI_DUMP_STACK, 0);
661 if (info[cpu].result == HV_NMI_RESULT_OK)
662 cpumask_clear_cpu(cpu, &mask);
663 }
664
665 mdelay(10);
666 timeout--;
667 }
668
669 /* Warn about cpus stuck in ICS and decrement their counts here. */
670 if (!cpumask_empty(&mask)) {
671 for_each_cpu(cpu, &mask) {
672 switch (info[cpu].result) {
673 case HV_NMI_RESULT_FAIL_ICS:
674 pr_warn("Skipping stack dump of cpu %d in ICS at pc %#llx\n",
675 cpu, info[cpu].pc);
676 break;
677 case HV_NMI_RESULT_FAIL_HV:
678 pr_warn("Skipping stack dump of cpu %d in hypervisor\n",
679 cpu);
680 break;
681 case HV_ENOSYS:
682 pr_warn("Hypervisor too old to allow remote stack dumps.\n");
683 goto skip_for_each;
684 default: /* should not happen */
685 pr_warn("Skipping stack dump of cpu %d [%d,%#llx]\n",
686 cpu, info[cpu].result, info[cpu].pc);
687 break;
688 }
689 }
690skip_for_each:
691 atomic_sub(cpumask_weight(&mask), &backtrace_cpus);
692 }
693}
694#endif /* __tilegx_ */