2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
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.
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
15 #include <linux/sched.h>
16 #include <linux/preempt.h>
17 #include <linux/module.h>
19 #include <linux/kprobes.h>
20 #include <linux/elfcore.h>
21 #include <linux/tick.h>
22 #include <linux/init.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>
42 #include <arch/chip.h>
44 #include <arch/sim_def.h>
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.
51 static int __init idle_setup(char *str)
56 if (!strcmp(str, "poll")) {
57 pr_info("using polling idle threads\n");
58 cpu_idle_poll_ctrl(true);
60 } else if (!strcmp(str, "halt")) {
65 early_param("idle", idle_setup);
67 void arch_cpu_idle(void)
69 __this_cpu_write(irq_stat.idle_timestamp, jiffies);
74 * Release a thread_info structure
76 void arch_release_thread_info(struct thread_info *info)
78 struct single_step_state *step_state = info->step_state;
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.
100 static void save_arch_state(struct thread_struct *t);
102 int copy_thread(unsigned long clone_flags, unsigned long sp,
103 unsigned long arg, struct task_struct *p)
105 struct pt_regs *childregs = task_pt_regs(p);
107 unsigned long *callee_regs;
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.
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;
126 /* Record the pid of the task that created this one. */
127 p->thread.creator_pid = current->pid;
129 if (unlikely(p->flags & PF_KTHREAD)) {
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 childregs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
137 p->thread.pc = (unsigned long) ret_from_kernel_thread;
142 * Start new thread in ret_from_fork so it schedules properly
143 * and then return from interrupt like the parent.
145 p->thread.pc = (unsigned long) ret_from_fork;
148 * Do not clone step state from the parent; each thread
149 * must make its own lazily.
151 task_thread_info(p)->step_state = NULL;
155 * Do not clone unalign jit fixup from the parent; each thread
156 * must allocate its own on demand.
158 task_thread_info(p)->unalign_jit_base = NULL;
162 * Copy the registers onto the kernel stack so the
163 * return-from-interrupt code will reload it into registers.
165 *childregs = *current_pt_regs();
166 childregs->regs[0] = 0; /* return value is zero */
168 childregs->sp = sp; /* override with new user stack pointer */
169 memcpy(callee_regs, &childregs->regs[CALLEE_SAVED_FIRST_REG],
170 CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long));
172 /* Save user stack top pointer so we can ID the stack vm area later. */
173 p->thread.usp0 = childregs->sp;
176 * If CLONE_SETTLS is set, set "tp" in the new task to "r4",
177 * which is passed in as arg #5 to sys_clone().
179 if (clone_flags & CLONE_SETTLS)
180 childregs->tp = childregs->regs[4];
183 #if CHIP_HAS_TILE_DMA()
185 * No DMA in the new thread. We model this on the fact that
186 * fork() clears the pending signals, alarms, and aio for the child.
188 memset(&p->thread.tile_dma_state, 0, sizeof(struct tile_dma_state));
189 memset(&p->thread.dma_async_tlb, 0, sizeof(struct async_tlb));
192 /* New thread has its miscellaneous processor state bits clear. */
193 p->thread.proc_status = 0;
195 #ifdef CONFIG_HARDWALL
196 /* New thread does not own any networks. */
197 memset(&p->thread.hardwall[0], 0,
198 sizeof(struct hardwall_task) * HARDWALL_TYPES);
203 * Start the new thread with the current architecture state
204 * (user interrupt masks, etc.).
206 save_arch_state(&p->thread);
211 int set_unalign_ctl(struct task_struct *tsk, unsigned int val)
213 task_thread_info(tsk)->align_ctl = val;
217 int get_unalign_ctl(struct task_struct *tsk, unsigned long adr)
219 return put_user(task_thread_info(tsk)->align_ctl,
220 (unsigned int __user *)adr);
223 static struct task_struct corrupt_current = { .comm = "<corrupt>" };
226 * Return "current" if it looks plausible, or else a pointer to a dummy.
227 * This can be helpful if we are just trying to emit a clean panic.
229 struct task_struct *validate_current(void)
231 struct task_struct *tsk = current;
232 if (unlikely((unsigned long)tsk < PAGE_OFFSET ||
233 (high_memory && (void *)tsk > high_memory) ||
234 ((unsigned long)tsk & (__alignof__(*tsk) - 1)) != 0)) {
235 pr_err("Corrupt 'current' %p (sp %#lx)\n", tsk, stack_pointer);
236 tsk = &corrupt_current;
241 /* Take and return the pointer to the previous task, for schedule_tail(). */
242 struct task_struct *sim_notify_fork(struct task_struct *prev)
244 struct task_struct *tsk = current;
245 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK_PARENT |
246 (tsk->thread.creator_pid << _SIM_CONTROL_OPERATOR_BITS));
247 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK |
248 (tsk->pid << _SIM_CONTROL_OPERATOR_BITS));
252 int dump_task_regs(struct task_struct *tsk, elf_gregset_t *regs)
254 struct pt_regs *ptregs = task_pt_regs(tsk);
255 elf_core_copy_regs(regs, ptregs);
259 #if CHIP_HAS_TILE_DMA()
261 /* Allow user processes to access the DMA SPRs */
262 void grant_dma_mpls(void)
264 #if CONFIG_KERNEL_PL == 2
265 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
266 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
268 __insn_mtspr(SPR_MPL_DMA_CPL_SET_0, 1);
269 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_0, 1);
273 /* Forbid user processes from accessing the DMA SPRs */
274 void restrict_dma_mpls(void)
276 #if CONFIG_KERNEL_PL == 2
277 __insn_mtspr(SPR_MPL_DMA_CPL_SET_2, 1);
278 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_2, 1);
280 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
281 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
285 /* Pause the DMA engine, then save off its state registers. */
286 static void save_tile_dma_state(struct tile_dma_state *dma)
288 unsigned long state = __insn_mfspr(SPR_DMA_USER_STATUS);
289 unsigned long post_suspend_state;
291 /* If we're running, suspend the engine. */
292 if ((state & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK)
293 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
296 * Wait for the engine to idle, then save regs. Note that we
297 * want to record the "running" bit from before suspension,
298 * and the "done" bit from after, so that we can properly
299 * distinguish a case where the user suspended the engine from
300 * the case where the kernel suspended as part of the context
304 post_suspend_state = __insn_mfspr(SPR_DMA_USER_STATUS);
305 } while (post_suspend_state & SPR_DMA_STATUS__BUSY_MASK);
307 dma->src = __insn_mfspr(SPR_DMA_SRC_ADDR);
308 dma->src_chunk = __insn_mfspr(SPR_DMA_SRC_CHUNK_ADDR);
309 dma->dest = __insn_mfspr(SPR_DMA_DST_ADDR);
310 dma->dest_chunk = __insn_mfspr(SPR_DMA_DST_CHUNK_ADDR);
311 dma->strides = __insn_mfspr(SPR_DMA_STRIDE);
312 dma->chunk_size = __insn_mfspr(SPR_DMA_CHUNK_SIZE);
313 dma->byte = __insn_mfspr(SPR_DMA_BYTE);
314 dma->status = (state & SPR_DMA_STATUS__RUNNING_MASK) |
315 (post_suspend_state & SPR_DMA_STATUS__DONE_MASK);
318 /* Restart a DMA that was running before we were context-switched out. */
319 static void restore_tile_dma_state(struct thread_struct *t)
321 const struct tile_dma_state *dma = &t->tile_dma_state;
324 * The only way to restore the done bit is to run a zero
325 * length transaction.
327 if ((dma->status & SPR_DMA_STATUS__DONE_MASK) &&
328 !(__insn_mfspr(SPR_DMA_USER_STATUS) & SPR_DMA_STATUS__DONE_MASK)) {
329 __insn_mtspr(SPR_DMA_BYTE, 0);
330 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
331 while (__insn_mfspr(SPR_DMA_USER_STATUS) &
332 SPR_DMA_STATUS__BUSY_MASK)
336 __insn_mtspr(SPR_DMA_SRC_ADDR, dma->src);
337 __insn_mtspr(SPR_DMA_SRC_CHUNK_ADDR, dma->src_chunk);
338 __insn_mtspr(SPR_DMA_DST_ADDR, dma->dest);
339 __insn_mtspr(SPR_DMA_DST_CHUNK_ADDR, dma->dest_chunk);
340 __insn_mtspr(SPR_DMA_STRIDE, dma->strides);
341 __insn_mtspr(SPR_DMA_CHUNK_SIZE, dma->chunk_size);
342 __insn_mtspr(SPR_DMA_BYTE, dma->byte);
345 * Restart the engine if we were running and not done.
346 * Clear a pending async DMA fault that we were waiting on return
347 * to user space to execute, since we expect the DMA engine
348 * to regenerate those faults for us now. Note that we don't
349 * try to clear the TIF_ASYNC_TLB flag, since it's relatively
350 * harmless if set, and it covers both DMA and the SN processor.
352 if ((dma->status & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK) {
353 t->dma_async_tlb.fault_num = 0;
354 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
360 static void save_arch_state(struct thread_struct *t)
362 #if CHIP_HAS_SPLIT_INTR_MASK()
363 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0_0) |
364 ((u64)__insn_mfspr(SPR_INTERRUPT_MASK_0_1) << 32);
366 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0);
368 t->ex_context[0] = __insn_mfspr(SPR_EX_CONTEXT_0_0);
369 t->ex_context[1] = __insn_mfspr(SPR_EX_CONTEXT_0_1);
370 t->system_save[0] = __insn_mfspr(SPR_SYSTEM_SAVE_0_0);
371 t->system_save[1] = __insn_mfspr(SPR_SYSTEM_SAVE_0_1);
372 t->system_save[2] = __insn_mfspr(SPR_SYSTEM_SAVE_0_2);
373 t->system_save[3] = __insn_mfspr(SPR_SYSTEM_SAVE_0_3);
374 t->intctrl_0 = __insn_mfspr(SPR_INTCTRL_0_STATUS);
375 t->proc_status = __insn_mfspr(SPR_PROC_STATUS);
376 #if !CHIP_HAS_FIXED_INTVEC_BASE()
377 t->interrupt_vector_base = __insn_mfspr(SPR_INTERRUPT_VECTOR_BASE_0);
379 t->tile_rtf_hwm = __insn_mfspr(SPR_TILE_RTF_HWM);
380 #if CHIP_HAS_DSTREAM_PF()
381 t->dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
385 static void restore_arch_state(const struct thread_struct *t)
387 #if CHIP_HAS_SPLIT_INTR_MASK()
388 __insn_mtspr(SPR_INTERRUPT_MASK_0_0, (u32) t->interrupt_mask);
389 __insn_mtspr(SPR_INTERRUPT_MASK_0_1, t->interrupt_mask >> 32);
391 __insn_mtspr(SPR_INTERRUPT_MASK_0, t->interrupt_mask);
393 __insn_mtspr(SPR_EX_CONTEXT_0_0, t->ex_context[0]);
394 __insn_mtspr(SPR_EX_CONTEXT_0_1, t->ex_context[1]);
395 __insn_mtspr(SPR_SYSTEM_SAVE_0_0, t->system_save[0]);
396 __insn_mtspr(SPR_SYSTEM_SAVE_0_1, t->system_save[1]);
397 __insn_mtspr(SPR_SYSTEM_SAVE_0_2, t->system_save[2]);
398 __insn_mtspr(SPR_SYSTEM_SAVE_0_3, t->system_save[3]);
399 __insn_mtspr(SPR_INTCTRL_0_STATUS, t->intctrl_0);
400 __insn_mtspr(SPR_PROC_STATUS, t->proc_status);
401 #if !CHIP_HAS_FIXED_INTVEC_BASE()
402 __insn_mtspr(SPR_INTERRUPT_VECTOR_BASE_0, t->interrupt_vector_base);
404 __insn_mtspr(SPR_TILE_RTF_HWM, t->tile_rtf_hwm);
405 #if CHIP_HAS_DSTREAM_PF()
406 __insn_mtspr(SPR_DSTREAM_PF, t->dstream_pf);
411 void _prepare_arch_switch(struct task_struct *next)
413 #if CHIP_HAS_TILE_DMA()
414 struct tile_dma_state *dma = ¤t->thread.tile_dma_state;
416 save_tile_dma_state(dma);
421 struct task_struct *__sched _switch_to(struct task_struct *prev,
422 struct task_struct *next)
424 /* DMA state is already saved; save off other arch state. */
425 save_arch_state(&prev->thread);
427 #if CHIP_HAS_TILE_DMA()
429 * Restore DMA in new task if desired.
430 * Note that it is only safe to restart here since interrupts
431 * are disabled, so we can't take any DMATLB miss or access
432 * interrupts before we have finished switching stacks.
434 if (next->thread.tile_dma_state.enabled) {
435 restore_tile_dma_state(&next->thread);
442 /* Restore other arch state. */
443 restore_arch_state(&next->thread);
445 #ifdef CONFIG_HARDWALL
446 /* Enable or disable access to the network registers appropriately. */
447 hardwall_switch_tasks(prev, next);
451 * Switch kernel SP, PC, and callee-saved registers.
452 * In the context of the new task, return the old task pointer
453 * (i.e. the task that actually called __switch_to).
454 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp.
456 return __switch_to(prev, next, next_current_ksp0(next));
460 * This routine is called on return from interrupt if any of the
461 * TIF_WORK_MASK flags are set in thread_info->flags. It is
462 * entered with interrupts disabled so we don't miss an event
463 * that modified the thread_info flags. If any flag is set, we
464 * handle it and return, and the calling assembly code will
465 * re-disable interrupts, reload the thread flags, and call back
466 * if more flags need to be handled.
468 * We return whether we need to check the thread_info flags again
469 * or not. Note that we don't clear TIF_SINGLESTEP here, so it's
470 * important that it be tested last, and then claim that we don't
471 * need to recheck the flags.
473 int do_work_pending(struct pt_regs *regs, u32 thread_info_flags)
475 /* If we enter in kernel mode, do nothing and exit the caller loop. */
476 if (!user_mode(regs))
481 /* Enable interrupts; they are disabled again on return to caller. */
484 if (thread_info_flags & _TIF_NEED_RESCHED) {
488 #if CHIP_HAS_TILE_DMA()
489 if (thread_info_flags & _TIF_ASYNC_TLB) {
490 do_async_page_fault(regs);
494 if (thread_info_flags & _TIF_SIGPENDING) {
498 if (thread_info_flags & _TIF_NOTIFY_RESUME) {
499 clear_thread_flag(TIF_NOTIFY_RESUME);
500 tracehook_notify_resume(regs);
503 if (thread_info_flags & _TIF_SINGLESTEP)
504 single_step_once(regs);
511 unsigned long get_wchan(struct task_struct *p)
513 struct KBacktraceIterator kbt;
515 if (!p || p == current || p->state == TASK_RUNNING)
518 for (KBacktraceIterator_init(&kbt, p, NULL);
519 !KBacktraceIterator_end(&kbt);
520 KBacktraceIterator_next(&kbt)) {
521 if (!in_sched_functions(kbt.it.pc))
528 /* Flush thread state. */
529 void flush_thread(void)
535 * Free current thread data structures etc..
537 void exit_thread(void)
539 #ifdef CONFIG_HARDWALL
541 * Remove the task from the list of tasks that are associated
542 * with any live hardwalls. (If the task that is exiting held
543 * the last reference to a hardwall fd, it would already have
544 * been released and deactivated at this point.)
546 hardwall_deactivate_all(current);
550 void show_regs(struct pt_regs *regs)
552 struct task_struct *tsk = validate_current();
555 if (tsk != &corrupt_current)
556 show_regs_print_info(KERN_ERR);
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);
566 for (i = 0; i < 13; i++)
567 pr_err(" r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT "\n",
568 i, regs->regs[i], i+14, regs->regs[i+14],
569 i+27, regs->regs[i+27], i+40, regs->regs[i+40]);
570 pr_err(" r13: " REGFMT " tp : " REGFMT " sp : " REGFMT " lr : " REGFMT "\n",
571 regs->regs[13], regs->tp, regs->sp, regs->lr);
573 pr_err(" pc : " REGFMT " ex1: %ld faultnum: %ld\n",
574 regs->pc, regs->ex1, regs->faultnum);
576 dump_stack_regs(regs);
579 /* To ensure stack dump on tiles occurs one by one. */
580 static DEFINE_SPINLOCK(backtrace_lock);
581 /* To ensure no backtrace occurs before all of the stack dump are done. */
582 static atomic_t backtrace_cpus;
583 /* The cpu mask to avoid reentrance. */
584 static struct cpumask backtrace_mask;
586 void do_nmi_dump_stack(struct pt_regs *regs)
588 int is_idle = is_idle_task(current) && !in_interrupt();
592 cpu = smp_processor_id();
593 if (WARN_ON_ONCE(!cpumask_test_and_clear_cpu(cpu, &backtrace_mask)))
596 spin_lock(&backtrace_lock);
598 pr_info("CPU: %d idle\n", cpu);
601 spin_unlock(&backtrace_lock);
602 atomic_dec(&backtrace_cpus);
608 void arch_trigger_all_cpu_backtrace(bool self)
612 unsigned int timeout;
615 HV_NMI_Info info[NR_CPUS];
617 ongoing = atomic_cmpxchg(&backtrace_cpus, 0, num_online_cpus() - 1);
619 pr_err("Trying to do all-cpu backtrace.\n");
620 pr_err("But another all-cpu backtrace is ongoing (%d cpus left)\n",
623 pr_err("Reporting the stack on this cpu only.\n");
629 cpumask_copy(&mask, cpu_online_mask);
630 cpumask_clear_cpu(smp_processor_id(), &mask);
631 cpumask_copy(&backtrace_mask, &mask);
633 /* Backtrace for myself first. */
637 /* Tentatively dump stack on remote tiles via NMI. */
639 while (!cpumask_empty(&mask) && timeout) {
640 for_each_cpu(cpu, &mask) {
643 info[cpu] = hv_send_nmi(tile, TILE_NMI_DUMP_STACK, 0);
644 if (info[cpu].result == HV_NMI_RESULT_OK)
645 cpumask_clear_cpu(cpu, &mask);
652 /* Warn about cpus stuck in ICS and decrement their counts here. */
653 if (!cpumask_empty(&mask)) {
654 for_each_cpu(cpu, &mask) {
655 switch (info[cpu].result) {
656 case HV_NMI_RESULT_FAIL_ICS:
657 pr_warn("Skipping stack dump of cpu %d in ICS at pc %#llx\n",
660 case HV_NMI_RESULT_FAIL_HV:
661 pr_warn("Skipping stack dump of cpu %d in hypervisor\n",
665 pr_warn("Hypervisor too old to allow remote stack dumps.\n");
667 default: /* should not happen */
668 pr_warn("Skipping stack dump of cpu %d [%d,%#llx]\n",
669 cpu, info[cpu].result, info[cpu].pc);
674 atomic_sub(cpumask_weight(&mask), &backtrace_cpus);
677 #endif /* __tilegx_ */