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 <asm/stack.h>
31 #include <asm/switch_to.h>
32 #include <asm/homecache.h>
33 #include <asm/syscalls.h>
34 #include <asm/traps.h>
35 #include <asm/setup.h>
36 #ifdef CONFIG_HARDWALL
37 #include <asm/hardwall.h>
39 #include <arch/chip.h>
41 #include <arch/sim_def.h>
45 * Use the (x86) "idle=poll" option to prefer low latency when leaving the
46 * idle loop over low power while in the idle loop, e.g. if we have
47 * one thread per core and we want to get threads out of futex waits fast.
49 static int no_idle_nap;
50 static int __init idle_setup(char *str)
55 if (!strcmp(str, "poll")) {
56 pr_info("using polling idle threads.\n");
58 } else if (!strcmp(str, "halt"))
65 early_param("idle", idle_setup);
68 * The idle thread. There's no useful work to be
69 * done, so just try to conserve power and have a
70 * low exit latency (ie sit in a loop waiting for
71 * somebody to say that they'd like to reschedule)
75 int cpu = smp_processor_id();
78 current_thread_info()->status |= TS_POLLING;
82 while (!need_resched())
88 /* endless idle loop with no priority at all */
90 tick_nohz_idle_enter();
92 while (!need_resched()) {
93 if (cpu_is_offline(cpu))
94 BUG(); /* no HOTPLUG_CPU */
97 __get_cpu_var(irq_stat).idle_timestamp = jiffies;
98 current_thread_info()->status &= ~TS_POLLING;
100 * TS_POLLING-cleared state must be visible before we
109 current_thread_info()->status |= TS_POLLING;
112 tick_nohz_idle_exit();
113 schedule_preempt_disabled();
118 * Release a thread_info structure
120 void arch_release_thread_info(struct thread_info *info)
122 struct single_step_state *step_state = info->step_state;
124 #ifdef CONFIG_HARDWALL
126 * We free a thread_info from the context of the task that has
127 * been scheduled next, so the original task is already dead.
128 * Calling deactivate here just frees up the data structures.
129 * If the task we're freeing held the last reference to a
130 * hardwall fd, it would have been released prior to this point
131 * anyway via exit_files(), and the hardwall_task.info pointers
132 * would be NULL by now.
134 hardwall_deactivate_all(info->task);
140 * FIXME: we don't munmap step_state->buffer
141 * because the mm_struct for this process (info->task->mm)
142 * has already been zeroed in exit_mm(). Keeping a
143 * reference to it here seems like a bad move, so this
144 * means we can't munmap() the buffer, and therefore if we
145 * ptrace multiple threads in a process, we will slowly
146 * leak user memory. (Note that as soon as the last
147 * thread in a process dies, we will reclaim all user
148 * memory including single-step buffers in the usual way.)
149 * We should either assign a kernel VA to this buffer
150 * somehow, or we should associate the buffer(s) with the
151 * mm itself so we can clean them up that way.
157 static void save_arch_state(struct thread_struct *t);
159 int copy_thread(unsigned long clone_flags, unsigned long sp,
161 struct task_struct *p, struct pt_regs *unused)
163 struct pt_regs *childregs = task_pt_regs(p);
165 unsigned long *callee_regs;
168 * Set up the stack and stack pointer appropriately for the
169 * new child to find itself woken up in __switch_to().
170 * The callee-saved registers must be on the stack to be read;
171 * the new task will then jump to assembly support to handle
172 * calling schedule_tail(), etc., and (for userspace tasks)
173 * returning to the context set up in the pt_regs.
175 ksp = (unsigned long) childregs;
176 ksp -= C_ABI_SAVE_AREA_SIZE; /* interrupt-entry save area */
177 ((long *)ksp)[0] = ((long *)ksp)[1] = 0;
178 ksp -= CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long);
179 callee_regs = (unsigned long *)ksp;
180 ksp -= C_ABI_SAVE_AREA_SIZE; /* __switch_to() save area */
181 ((long *)ksp)[0] = ((long *)ksp)[1] = 0;
184 /* Record the pid of the task that created this one. */
185 p->thread.creator_pid = current->pid;
187 if (unlikely(p->flags & PF_KTHREAD)) {
189 memset(childregs, 0, sizeof(struct pt_regs));
190 memset(&callee_regs[2], 0,
191 (CALLEE_SAVED_REGS_COUNT - 2) * sizeof(unsigned long));
192 callee_regs[0] = sp; /* r30 = function */
193 callee_regs[1] = arg; /* r31 = arg */
194 childregs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
195 p->thread.pc = (unsigned long) ret_from_kernel_thread;
200 * Start new thread in ret_from_fork so it schedules properly
201 * and then return from interrupt like the parent.
203 p->thread.pc = (unsigned long) ret_from_fork;
206 * Do not clone step state from the parent; each thread
207 * must make its own lazily.
209 task_thread_info(p)->step_state = NULL;
212 * Copy the registers onto the kernel stack so the
213 * return-from-interrupt code will reload it into registers.
215 *childregs = *current_pt_regs();
216 childregs->regs[0] = 0; /* return value is zero */
218 childregs->sp = sp; /* override with new user stack pointer */
219 memcpy(callee_regs, &childregs->regs[CALLEE_SAVED_FIRST_REG],
220 CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long));
222 /* Save user stack top pointer so we can ID the stack vm area later. */
223 p->thread.usp0 = childregs->sp;
226 * If CLONE_SETTLS is set, set "tp" in the new task to "r4",
227 * which is passed in as arg #5 to sys_clone().
229 if (clone_flags & CLONE_SETTLS)
230 childregs->tp = childregs->regs[4];
233 #if CHIP_HAS_TILE_DMA()
235 * No DMA in the new thread. We model this on the fact that
236 * fork() clears the pending signals, alarms, and aio for the child.
238 memset(&p->thread.tile_dma_state, 0, sizeof(struct tile_dma_state));
239 memset(&p->thread.dma_async_tlb, 0, sizeof(struct async_tlb));
242 #if CHIP_HAS_SN_PROC()
243 /* Likewise, the new thread is not running static processor code. */
244 p->thread.sn_proc_running = 0;
245 memset(&p->thread.sn_async_tlb, 0, sizeof(struct async_tlb));
248 #if CHIP_HAS_PROC_STATUS_SPR()
249 /* New thread has its miscellaneous processor state bits clear. */
250 p->thread.proc_status = 0;
253 #ifdef CONFIG_HARDWALL
254 /* New thread does not own any networks. */
255 memset(&p->thread.hardwall[0], 0,
256 sizeof(struct hardwall_task) * HARDWALL_TYPES);
261 * Start the new thread with the current architecture state
262 * (user interrupt masks, etc.).
264 save_arch_state(&p->thread);
270 * Return "current" if it looks plausible, or else a pointer to a dummy.
271 * This can be helpful if we are just trying to emit a clean panic.
273 struct task_struct *validate_current(void)
275 static struct task_struct corrupt = { .comm = "<corrupt>" };
276 struct task_struct *tsk = current;
277 if (unlikely((unsigned long)tsk < PAGE_OFFSET ||
278 (high_memory && (void *)tsk > high_memory) ||
279 ((unsigned long)tsk & (__alignof__(*tsk) - 1)) != 0)) {
280 pr_err("Corrupt 'current' %p (sp %#lx)\n", tsk, stack_pointer);
286 /* Take and return the pointer to the previous task, for schedule_tail(). */
287 struct task_struct *sim_notify_fork(struct task_struct *prev)
289 struct task_struct *tsk = current;
290 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK_PARENT |
291 (tsk->thread.creator_pid << _SIM_CONTROL_OPERATOR_BITS));
292 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK |
293 (tsk->pid << _SIM_CONTROL_OPERATOR_BITS));
297 int dump_task_regs(struct task_struct *tsk, elf_gregset_t *regs)
299 struct pt_regs *ptregs = task_pt_regs(tsk);
300 elf_core_copy_regs(regs, ptregs);
304 #if CHIP_HAS_TILE_DMA()
306 /* Allow user processes to access the DMA SPRs */
307 void grant_dma_mpls(void)
309 #if CONFIG_KERNEL_PL == 2
310 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
311 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
313 __insn_mtspr(SPR_MPL_DMA_CPL_SET_0, 1);
314 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_0, 1);
318 /* Forbid user processes from accessing the DMA SPRs */
319 void restrict_dma_mpls(void)
321 #if CONFIG_KERNEL_PL == 2
322 __insn_mtspr(SPR_MPL_DMA_CPL_SET_2, 1);
323 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_2, 1);
325 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
326 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
330 /* Pause the DMA engine, then save off its state registers. */
331 static void save_tile_dma_state(struct tile_dma_state *dma)
333 unsigned long state = __insn_mfspr(SPR_DMA_USER_STATUS);
334 unsigned long post_suspend_state;
336 /* If we're running, suspend the engine. */
337 if ((state & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK)
338 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
341 * Wait for the engine to idle, then save regs. Note that we
342 * want to record the "running" bit from before suspension,
343 * and the "done" bit from after, so that we can properly
344 * distinguish a case where the user suspended the engine from
345 * the case where the kernel suspended as part of the context
349 post_suspend_state = __insn_mfspr(SPR_DMA_USER_STATUS);
350 } while (post_suspend_state & SPR_DMA_STATUS__BUSY_MASK);
352 dma->src = __insn_mfspr(SPR_DMA_SRC_ADDR);
353 dma->src_chunk = __insn_mfspr(SPR_DMA_SRC_CHUNK_ADDR);
354 dma->dest = __insn_mfspr(SPR_DMA_DST_ADDR);
355 dma->dest_chunk = __insn_mfspr(SPR_DMA_DST_CHUNK_ADDR);
356 dma->strides = __insn_mfspr(SPR_DMA_STRIDE);
357 dma->chunk_size = __insn_mfspr(SPR_DMA_CHUNK_SIZE);
358 dma->byte = __insn_mfspr(SPR_DMA_BYTE);
359 dma->status = (state & SPR_DMA_STATUS__RUNNING_MASK) |
360 (post_suspend_state & SPR_DMA_STATUS__DONE_MASK);
363 /* Restart a DMA that was running before we were context-switched out. */
364 static void restore_tile_dma_state(struct thread_struct *t)
366 const struct tile_dma_state *dma = &t->tile_dma_state;
369 * The only way to restore the done bit is to run a zero
370 * length transaction.
372 if ((dma->status & SPR_DMA_STATUS__DONE_MASK) &&
373 !(__insn_mfspr(SPR_DMA_USER_STATUS) & SPR_DMA_STATUS__DONE_MASK)) {
374 __insn_mtspr(SPR_DMA_BYTE, 0);
375 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
376 while (__insn_mfspr(SPR_DMA_USER_STATUS) &
377 SPR_DMA_STATUS__BUSY_MASK)
381 __insn_mtspr(SPR_DMA_SRC_ADDR, dma->src);
382 __insn_mtspr(SPR_DMA_SRC_CHUNK_ADDR, dma->src_chunk);
383 __insn_mtspr(SPR_DMA_DST_ADDR, dma->dest);
384 __insn_mtspr(SPR_DMA_DST_CHUNK_ADDR, dma->dest_chunk);
385 __insn_mtspr(SPR_DMA_STRIDE, dma->strides);
386 __insn_mtspr(SPR_DMA_CHUNK_SIZE, dma->chunk_size);
387 __insn_mtspr(SPR_DMA_BYTE, dma->byte);
390 * Restart the engine if we were running and not done.
391 * Clear a pending async DMA fault that we were waiting on return
392 * to user space to execute, since we expect the DMA engine
393 * to regenerate those faults for us now. Note that we don't
394 * try to clear the TIF_ASYNC_TLB flag, since it's relatively
395 * harmless if set, and it covers both DMA and the SN processor.
397 if ((dma->status & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK) {
398 t->dma_async_tlb.fault_num = 0;
399 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
405 static void save_arch_state(struct thread_struct *t)
407 #if CHIP_HAS_SPLIT_INTR_MASK()
408 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0_0) |
409 ((u64)__insn_mfspr(SPR_INTERRUPT_MASK_0_1) << 32);
411 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0);
413 t->ex_context[0] = __insn_mfspr(SPR_EX_CONTEXT_0_0);
414 t->ex_context[1] = __insn_mfspr(SPR_EX_CONTEXT_0_1);
415 t->system_save[0] = __insn_mfspr(SPR_SYSTEM_SAVE_0_0);
416 t->system_save[1] = __insn_mfspr(SPR_SYSTEM_SAVE_0_1);
417 t->system_save[2] = __insn_mfspr(SPR_SYSTEM_SAVE_0_2);
418 t->system_save[3] = __insn_mfspr(SPR_SYSTEM_SAVE_0_3);
419 t->intctrl_0 = __insn_mfspr(SPR_INTCTRL_0_STATUS);
420 #if CHIP_HAS_PROC_STATUS_SPR()
421 t->proc_status = __insn_mfspr(SPR_PROC_STATUS);
423 #if !CHIP_HAS_FIXED_INTVEC_BASE()
424 t->interrupt_vector_base = __insn_mfspr(SPR_INTERRUPT_VECTOR_BASE_0);
426 #if CHIP_HAS_TILE_RTF_HWM()
427 t->tile_rtf_hwm = __insn_mfspr(SPR_TILE_RTF_HWM);
429 #if CHIP_HAS_DSTREAM_PF()
430 t->dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
434 static void restore_arch_state(const struct thread_struct *t)
436 #if CHIP_HAS_SPLIT_INTR_MASK()
437 __insn_mtspr(SPR_INTERRUPT_MASK_0_0, (u32) t->interrupt_mask);
438 __insn_mtspr(SPR_INTERRUPT_MASK_0_1, t->interrupt_mask >> 32);
440 __insn_mtspr(SPR_INTERRUPT_MASK_0, t->interrupt_mask);
442 __insn_mtspr(SPR_EX_CONTEXT_0_0, t->ex_context[0]);
443 __insn_mtspr(SPR_EX_CONTEXT_0_1, t->ex_context[1]);
444 __insn_mtspr(SPR_SYSTEM_SAVE_0_0, t->system_save[0]);
445 __insn_mtspr(SPR_SYSTEM_SAVE_0_1, t->system_save[1]);
446 __insn_mtspr(SPR_SYSTEM_SAVE_0_2, t->system_save[2]);
447 __insn_mtspr(SPR_SYSTEM_SAVE_0_3, t->system_save[3]);
448 __insn_mtspr(SPR_INTCTRL_0_STATUS, t->intctrl_0);
449 #if CHIP_HAS_PROC_STATUS_SPR()
450 __insn_mtspr(SPR_PROC_STATUS, t->proc_status);
452 #if !CHIP_HAS_FIXED_INTVEC_BASE()
453 __insn_mtspr(SPR_INTERRUPT_VECTOR_BASE_0, t->interrupt_vector_base);
455 #if CHIP_HAS_TILE_RTF_HWM()
456 __insn_mtspr(SPR_TILE_RTF_HWM, t->tile_rtf_hwm);
458 #if CHIP_HAS_DSTREAM_PF()
459 __insn_mtspr(SPR_DSTREAM_PF, t->dstream_pf);
464 void _prepare_arch_switch(struct task_struct *next)
466 #if CHIP_HAS_SN_PROC()
469 #if CHIP_HAS_TILE_DMA()
470 struct tile_dma_state *dma = ¤t->thread.tile_dma_state;
472 save_tile_dma_state(dma);
474 #if CHIP_HAS_SN_PROC()
476 * Suspend the static network processor if it was running.
477 * We do not suspend the fabric itself, just like we don't
478 * try to suspend the UDN.
480 snctl = __insn_mfspr(SPR_SNCTL);
481 current->thread.sn_proc_running =
482 (snctl & SPR_SNCTL__FRZPROC_MASK) == 0;
483 if (current->thread.sn_proc_running)
484 __insn_mtspr(SPR_SNCTL, snctl | SPR_SNCTL__FRZPROC_MASK);
489 struct task_struct *__sched _switch_to(struct task_struct *prev,
490 struct task_struct *next)
492 /* DMA state is already saved; save off other arch state. */
493 save_arch_state(&prev->thread);
495 #if CHIP_HAS_TILE_DMA()
497 * Restore DMA in new task if desired.
498 * Note that it is only safe to restart here since interrupts
499 * are disabled, so we can't take any DMATLB miss or access
500 * interrupts before we have finished switching stacks.
502 if (next->thread.tile_dma_state.enabled) {
503 restore_tile_dma_state(&next->thread);
510 /* Restore other arch state. */
511 restore_arch_state(&next->thread);
513 #if CHIP_HAS_SN_PROC()
515 * Restart static network processor in the new process
516 * if it was running before.
518 if (next->thread.sn_proc_running) {
519 int snctl = __insn_mfspr(SPR_SNCTL);
520 __insn_mtspr(SPR_SNCTL, snctl & ~SPR_SNCTL__FRZPROC_MASK);
524 #ifdef CONFIG_HARDWALL
525 /* Enable or disable access to the network registers appropriately. */
526 hardwall_switch_tasks(prev, next);
530 * Switch kernel SP, PC, and callee-saved registers.
531 * In the context of the new task, return the old task pointer
532 * (i.e. the task that actually called __switch_to).
533 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp.
535 return __switch_to(prev, next, next_current_ksp0(next));
539 * This routine is called on return from interrupt if any of the
540 * TIF_WORK_MASK flags are set in thread_info->flags. It is
541 * entered with interrupts disabled so we don't miss an event
542 * that modified the thread_info flags. If any flag is set, we
543 * handle it and return, and the calling assembly code will
544 * re-disable interrupts, reload the thread flags, and call back
545 * if more flags need to be handled.
547 * We return whether we need to check the thread_info flags again
548 * or not. Note that we don't clear TIF_SINGLESTEP here, so it's
549 * important that it be tested last, and then claim that we don't
550 * need to recheck the flags.
552 int do_work_pending(struct pt_regs *regs, u32 thread_info_flags)
554 /* If we enter in kernel mode, do nothing and exit the caller loop. */
555 if (!user_mode(regs))
558 /* Enable interrupts; they are disabled again on return to caller. */
561 if (thread_info_flags & _TIF_NEED_RESCHED) {
565 #if CHIP_HAS_TILE_DMA() || CHIP_HAS_SN_PROC()
566 if (thread_info_flags & _TIF_ASYNC_TLB) {
567 do_async_page_fault(regs);
571 if (thread_info_flags & _TIF_SIGPENDING) {
575 if (thread_info_flags & _TIF_NOTIFY_RESUME) {
576 clear_thread_flag(TIF_NOTIFY_RESUME);
577 tracehook_notify_resume(regs);
580 if (thread_info_flags & _TIF_SINGLESTEP) {
581 single_step_once(regs);
584 panic("work_pending: bad flags %#x\n", thread_info_flags);
587 /* Note there is an implicit fifth argument if (clone_flags & CLONE_SETTLS). */
588 SYSCALL_DEFINE4(clone, unsigned long, clone_flags, unsigned long, newsp,
589 void __user *, parent_tidptr, void __user *, child_tidptr)
591 return do_fork(clone_flags, newsp, current_pt_regs(), 0,
592 parent_tidptr, child_tidptr);
595 unsigned long get_wchan(struct task_struct *p)
597 struct KBacktraceIterator kbt;
599 if (!p || p == current || p->state == TASK_RUNNING)
602 for (KBacktraceIterator_init(&kbt, p, NULL);
603 !KBacktraceIterator_end(&kbt);
604 KBacktraceIterator_next(&kbt)) {
605 if (!in_sched_functions(kbt.it.pc))
612 /* Flush thread state. */
613 void flush_thread(void)
619 * Free current thread data structures etc..
621 void exit_thread(void)
626 void show_regs(struct pt_regs *regs)
628 struct task_struct *tsk = validate_current();
632 pr_err(" Pid: %d, comm: %20s, CPU: %d\n",
633 tsk->pid, tsk->comm, smp_processor_id());
635 for (i = 0; i < 51; i += 3)
636 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
637 i, regs->regs[i], i+1, regs->regs[i+1],
638 i+2, regs->regs[i+2]);
639 pr_err(" r51: "REGFMT" r52: "REGFMT" tp : "REGFMT"\n",
640 regs->regs[51], regs->regs[52], regs->tp);
641 pr_err(" sp : "REGFMT" lr : "REGFMT"\n", regs->sp, regs->lr);
643 for (i = 0; i < 52; i += 4)
644 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT
645 " r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
646 i, regs->regs[i], i+1, regs->regs[i+1],
647 i+2, regs->regs[i+2], i+3, regs->regs[i+3]);
648 pr_err(" r52: "REGFMT" tp : "REGFMT" sp : "REGFMT" lr : "REGFMT"\n",
649 regs->regs[52], regs->tp, regs->sp, regs->lr);
651 pr_err(" pc : "REGFMT" ex1: %ld faultnum: %ld\n",
652 regs->pc, regs->ex1, regs->faultnum);
654 dump_stack_regs(regs);