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,
160 unsigned long arg, struct task_struct *p)
162 struct pt_regs *childregs = task_pt_regs(p), *regs = current_pt_regs();
164 unsigned long *callee_regs;
167 * Set up the stack and stack pointer appropriately for the
168 * new child to find itself woken up in __switch_to().
169 * The callee-saved registers must be on the stack to be read;
170 * the new task will then jump to assembly support to handle
171 * calling schedule_tail(), etc., and (for userspace tasks)
172 * returning to the context set up in the pt_regs.
174 ksp = (unsigned long) childregs;
175 ksp -= C_ABI_SAVE_AREA_SIZE; /* interrupt-entry save area */
176 ((long *)ksp)[0] = ((long *)ksp)[1] = 0;
177 ksp -= CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long);
178 callee_regs = (unsigned long *)ksp;
179 ksp -= C_ABI_SAVE_AREA_SIZE; /* __switch_to() save area */
180 ((long *)ksp)[0] = ((long *)ksp)[1] = 0;
183 /* Record the pid of the task that created this one. */
184 p->thread.creator_pid = current->pid;
186 if (unlikely(p->flags & PF_KTHREAD)) {
188 memset(childregs, 0, sizeof(struct pt_regs));
189 memset(&callee_regs[2], 0,
190 (CALLEE_SAVED_REGS_COUNT - 2) * sizeof(unsigned long));
191 callee_regs[0] = sp; /* r30 = function */
192 callee_regs[1] = arg; /* r31 = arg */
193 childregs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
194 p->thread.pc = (unsigned long) ret_from_kernel_thread;
199 * Start new thread in ret_from_fork so it schedules properly
200 * and then return from interrupt like the parent.
202 p->thread.pc = (unsigned long) ret_from_fork;
205 * Do not clone step state from the parent; each thread
206 * must make its own lazily.
208 task_thread_info(p)->step_state = NULL;
211 * Copy the registers onto the kernel stack so the
212 * return-from-interrupt code will reload it into registers.
214 *childregs = *current_pt_regs();
215 childregs->regs[0] = 0; /* return value is zero */
217 childregs->sp = sp; /* override with new user stack pointer */
218 memcpy(callee_regs, &childregs->regs[CALLEE_SAVED_FIRST_REG],
219 CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long));
221 /* Save user stack top pointer so we can ID the stack vm area later. */
222 p->thread.usp0 = childregs->sp;
225 * If CLONE_SETTLS is set, set "tp" in the new task to "r4",
226 * which is passed in as arg #5 to sys_clone().
228 if (clone_flags & CLONE_SETTLS)
229 childregs->tp = childregs->regs[4];
232 #if CHIP_HAS_TILE_DMA()
234 * No DMA in the new thread. We model this on the fact that
235 * fork() clears the pending signals, alarms, and aio for the child.
237 memset(&p->thread.tile_dma_state, 0, sizeof(struct tile_dma_state));
238 memset(&p->thread.dma_async_tlb, 0, sizeof(struct async_tlb));
241 #if CHIP_HAS_SN_PROC()
242 /* Likewise, the new thread is not running static processor code. */
243 p->thread.sn_proc_running = 0;
244 memset(&p->thread.sn_async_tlb, 0, sizeof(struct async_tlb));
247 #if CHIP_HAS_PROC_STATUS_SPR()
248 /* New thread has its miscellaneous processor state bits clear. */
249 p->thread.proc_status = 0;
252 #ifdef CONFIG_HARDWALL
253 /* New thread does not own any networks. */
254 memset(&p->thread.hardwall[0], 0,
255 sizeof(struct hardwall_task) * HARDWALL_TYPES);
260 * Start the new thread with the current architecture state
261 * (user interrupt masks, etc.).
263 save_arch_state(&p->thread);
269 * Return "current" if it looks plausible, or else a pointer to a dummy.
270 * This can be helpful if we are just trying to emit a clean panic.
272 struct task_struct *validate_current(void)
274 static struct task_struct corrupt = { .comm = "<corrupt>" };
275 struct task_struct *tsk = current;
276 if (unlikely((unsigned long)tsk < PAGE_OFFSET ||
277 (high_memory && (void *)tsk > high_memory) ||
278 ((unsigned long)tsk & (__alignof__(*tsk) - 1)) != 0)) {
279 pr_err("Corrupt 'current' %p (sp %#lx)\n", tsk, stack_pointer);
285 /* Take and return the pointer to the previous task, for schedule_tail(). */
286 struct task_struct *sim_notify_fork(struct task_struct *prev)
288 struct task_struct *tsk = current;
289 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK_PARENT |
290 (tsk->thread.creator_pid << _SIM_CONTROL_OPERATOR_BITS));
291 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK |
292 (tsk->pid << _SIM_CONTROL_OPERATOR_BITS));
296 int dump_task_regs(struct task_struct *tsk, elf_gregset_t *regs)
298 struct pt_regs *ptregs = task_pt_regs(tsk);
299 elf_core_copy_regs(regs, ptregs);
303 #if CHIP_HAS_TILE_DMA()
305 /* Allow user processes to access the DMA SPRs */
306 void grant_dma_mpls(void)
308 #if CONFIG_KERNEL_PL == 2
309 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
310 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
312 __insn_mtspr(SPR_MPL_DMA_CPL_SET_0, 1);
313 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_0, 1);
317 /* Forbid user processes from accessing the DMA SPRs */
318 void restrict_dma_mpls(void)
320 #if CONFIG_KERNEL_PL == 2
321 __insn_mtspr(SPR_MPL_DMA_CPL_SET_2, 1);
322 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_2, 1);
324 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
325 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
329 /* Pause the DMA engine, then save off its state registers. */
330 static void save_tile_dma_state(struct tile_dma_state *dma)
332 unsigned long state = __insn_mfspr(SPR_DMA_USER_STATUS);
333 unsigned long post_suspend_state;
335 /* If we're running, suspend the engine. */
336 if ((state & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK)
337 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
340 * Wait for the engine to idle, then save regs. Note that we
341 * want to record the "running" bit from before suspension,
342 * and the "done" bit from after, so that we can properly
343 * distinguish a case where the user suspended the engine from
344 * the case where the kernel suspended as part of the context
348 post_suspend_state = __insn_mfspr(SPR_DMA_USER_STATUS);
349 } while (post_suspend_state & SPR_DMA_STATUS__BUSY_MASK);
351 dma->src = __insn_mfspr(SPR_DMA_SRC_ADDR);
352 dma->src_chunk = __insn_mfspr(SPR_DMA_SRC_CHUNK_ADDR);
353 dma->dest = __insn_mfspr(SPR_DMA_DST_ADDR);
354 dma->dest_chunk = __insn_mfspr(SPR_DMA_DST_CHUNK_ADDR);
355 dma->strides = __insn_mfspr(SPR_DMA_STRIDE);
356 dma->chunk_size = __insn_mfspr(SPR_DMA_CHUNK_SIZE);
357 dma->byte = __insn_mfspr(SPR_DMA_BYTE);
358 dma->status = (state & SPR_DMA_STATUS__RUNNING_MASK) |
359 (post_suspend_state & SPR_DMA_STATUS__DONE_MASK);
362 /* Restart a DMA that was running before we were context-switched out. */
363 static void restore_tile_dma_state(struct thread_struct *t)
365 const struct tile_dma_state *dma = &t->tile_dma_state;
368 * The only way to restore the done bit is to run a zero
369 * length transaction.
371 if ((dma->status & SPR_DMA_STATUS__DONE_MASK) &&
372 !(__insn_mfspr(SPR_DMA_USER_STATUS) & SPR_DMA_STATUS__DONE_MASK)) {
373 __insn_mtspr(SPR_DMA_BYTE, 0);
374 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
375 while (__insn_mfspr(SPR_DMA_USER_STATUS) &
376 SPR_DMA_STATUS__BUSY_MASK)
380 __insn_mtspr(SPR_DMA_SRC_ADDR, dma->src);
381 __insn_mtspr(SPR_DMA_SRC_CHUNK_ADDR, dma->src_chunk);
382 __insn_mtspr(SPR_DMA_DST_ADDR, dma->dest);
383 __insn_mtspr(SPR_DMA_DST_CHUNK_ADDR, dma->dest_chunk);
384 __insn_mtspr(SPR_DMA_STRIDE, dma->strides);
385 __insn_mtspr(SPR_DMA_CHUNK_SIZE, dma->chunk_size);
386 __insn_mtspr(SPR_DMA_BYTE, dma->byte);
389 * Restart the engine if we were running and not done.
390 * Clear a pending async DMA fault that we were waiting on return
391 * to user space to execute, since we expect the DMA engine
392 * to regenerate those faults for us now. Note that we don't
393 * try to clear the TIF_ASYNC_TLB flag, since it's relatively
394 * harmless if set, and it covers both DMA and the SN processor.
396 if ((dma->status & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK) {
397 t->dma_async_tlb.fault_num = 0;
398 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
404 static void save_arch_state(struct thread_struct *t)
406 #if CHIP_HAS_SPLIT_INTR_MASK()
407 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0_0) |
408 ((u64)__insn_mfspr(SPR_INTERRUPT_MASK_0_1) << 32);
410 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0);
412 t->ex_context[0] = __insn_mfspr(SPR_EX_CONTEXT_0_0);
413 t->ex_context[1] = __insn_mfspr(SPR_EX_CONTEXT_0_1);
414 t->system_save[0] = __insn_mfspr(SPR_SYSTEM_SAVE_0_0);
415 t->system_save[1] = __insn_mfspr(SPR_SYSTEM_SAVE_0_1);
416 t->system_save[2] = __insn_mfspr(SPR_SYSTEM_SAVE_0_2);
417 t->system_save[3] = __insn_mfspr(SPR_SYSTEM_SAVE_0_3);
418 t->intctrl_0 = __insn_mfspr(SPR_INTCTRL_0_STATUS);
419 #if CHIP_HAS_PROC_STATUS_SPR()
420 t->proc_status = __insn_mfspr(SPR_PROC_STATUS);
422 #if !CHIP_HAS_FIXED_INTVEC_BASE()
423 t->interrupt_vector_base = __insn_mfspr(SPR_INTERRUPT_VECTOR_BASE_0);
425 #if CHIP_HAS_TILE_RTF_HWM()
426 t->tile_rtf_hwm = __insn_mfspr(SPR_TILE_RTF_HWM);
428 #if CHIP_HAS_DSTREAM_PF()
429 t->dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
433 static void restore_arch_state(const struct thread_struct *t)
435 #if CHIP_HAS_SPLIT_INTR_MASK()
436 __insn_mtspr(SPR_INTERRUPT_MASK_0_0, (u32) t->interrupt_mask);
437 __insn_mtspr(SPR_INTERRUPT_MASK_0_1, t->interrupt_mask >> 32);
439 __insn_mtspr(SPR_INTERRUPT_MASK_0, t->interrupt_mask);
441 __insn_mtspr(SPR_EX_CONTEXT_0_0, t->ex_context[0]);
442 __insn_mtspr(SPR_EX_CONTEXT_0_1, t->ex_context[1]);
443 __insn_mtspr(SPR_SYSTEM_SAVE_0_0, t->system_save[0]);
444 __insn_mtspr(SPR_SYSTEM_SAVE_0_1, t->system_save[1]);
445 __insn_mtspr(SPR_SYSTEM_SAVE_0_2, t->system_save[2]);
446 __insn_mtspr(SPR_SYSTEM_SAVE_0_3, t->system_save[3]);
447 __insn_mtspr(SPR_INTCTRL_0_STATUS, t->intctrl_0);
448 #if CHIP_HAS_PROC_STATUS_SPR()
449 __insn_mtspr(SPR_PROC_STATUS, t->proc_status);
451 #if !CHIP_HAS_FIXED_INTVEC_BASE()
452 __insn_mtspr(SPR_INTERRUPT_VECTOR_BASE_0, t->interrupt_vector_base);
454 #if CHIP_HAS_TILE_RTF_HWM()
455 __insn_mtspr(SPR_TILE_RTF_HWM, t->tile_rtf_hwm);
457 #if CHIP_HAS_DSTREAM_PF()
458 __insn_mtspr(SPR_DSTREAM_PF, t->dstream_pf);
463 void _prepare_arch_switch(struct task_struct *next)
465 #if CHIP_HAS_SN_PROC()
468 #if CHIP_HAS_TILE_DMA()
469 struct tile_dma_state *dma = ¤t->thread.tile_dma_state;
471 save_tile_dma_state(dma);
473 #if CHIP_HAS_SN_PROC()
475 * Suspend the static network processor if it was running.
476 * We do not suspend the fabric itself, just like we don't
477 * try to suspend the UDN.
479 snctl = __insn_mfspr(SPR_SNCTL);
480 current->thread.sn_proc_running =
481 (snctl & SPR_SNCTL__FRZPROC_MASK) == 0;
482 if (current->thread.sn_proc_running)
483 __insn_mtspr(SPR_SNCTL, snctl | SPR_SNCTL__FRZPROC_MASK);
488 struct task_struct *__sched _switch_to(struct task_struct *prev,
489 struct task_struct *next)
491 /* DMA state is already saved; save off other arch state. */
492 save_arch_state(&prev->thread);
494 #if CHIP_HAS_TILE_DMA()
496 * Restore DMA in new task if desired.
497 * Note that it is only safe to restart here since interrupts
498 * are disabled, so we can't take any DMATLB miss or access
499 * interrupts before we have finished switching stacks.
501 if (next->thread.tile_dma_state.enabled) {
502 restore_tile_dma_state(&next->thread);
509 /* Restore other arch state. */
510 restore_arch_state(&next->thread);
512 #if CHIP_HAS_SN_PROC()
514 * Restart static network processor in the new process
515 * if it was running before.
517 if (next->thread.sn_proc_running) {
518 int snctl = __insn_mfspr(SPR_SNCTL);
519 __insn_mtspr(SPR_SNCTL, snctl & ~SPR_SNCTL__FRZPROC_MASK);
523 #ifdef CONFIG_HARDWALL
524 /* Enable or disable access to the network registers appropriately. */
525 hardwall_switch_tasks(prev, next);
529 * Switch kernel SP, PC, and callee-saved registers.
530 * In the context of the new task, return the old task pointer
531 * (i.e. the task that actually called __switch_to).
532 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp.
534 return __switch_to(prev, next, next_current_ksp0(next));
538 * This routine is called on return from interrupt if any of the
539 * TIF_WORK_MASK flags are set in thread_info->flags. It is
540 * entered with interrupts disabled so we don't miss an event
541 * that modified the thread_info flags. If any flag is set, we
542 * handle it and return, and the calling assembly code will
543 * re-disable interrupts, reload the thread flags, and call back
544 * if more flags need to be handled.
546 * We return whether we need to check the thread_info flags again
547 * or not. Note that we don't clear TIF_SINGLESTEP here, so it's
548 * important that it be tested last, and then claim that we don't
549 * need to recheck the flags.
551 int do_work_pending(struct pt_regs *regs, u32 thread_info_flags)
553 /* If we enter in kernel mode, do nothing and exit the caller loop. */
554 if (!user_mode(regs))
557 /* Enable interrupts; they are disabled again on return to caller. */
560 if (thread_info_flags & _TIF_NEED_RESCHED) {
564 #if CHIP_HAS_TILE_DMA() || CHIP_HAS_SN_PROC()
565 if (thread_info_flags & _TIF_ASYNC_TLB) {
566 do_async_page_fault(regs);
570 if (thread_info_flags & _TIF_SIGPENDING) {
574 if (thread_info_flags & _TIF_NOTIFY_RESUME) {
575 clear_thread_flag(TIF_NOTIFY_RESUME);
576 tracehook_notify_resume(regs);
579 if (thread_info_flags & _TIF_SINGLESTEP) {
580 single_step_once(regs);
583 panic("work_pending: bad flags %#x\n", thread_info_flags);
586 unsigned long get_wchan(struct task_struct *p)
588 struct KBacktraceIterator kbt;
590 if (!p || p == current || p->state == TASK_RUNNING)
593 for (KBacktraceIterator_init(&kbt, p, NULL);
594 !KBacktraceIterator_end(&kbt);
595 KBacktraceIterator_next(&kbt)) {
596 if (!in_sched_functions(kbt.it.pc))
603 /* Flush thread state. */
604 void flush_thread(void)
610 * Free current thread data structures etc..
612 void exit_thread(void)
617 void show_regs(struct pt_regs *regs)
619 struct task_struct *tsk = validate_current();
623 pr_err(" Pid: %d, comm: %20s, CPU: %d\n",
624 tsk->pid, tsk->comm, smp_processor_id());
626 for (i = 0; i < 51; i += 3)
627 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
628 i, regs->regs[i], i+1, regs->regs[i+1],
629 i+2, regs->regs[i+2]);
630 pr_err(" r51: "REGFMT" r52: "REGFMT" tp : "REGFMT"\n",
631 regs->regs[51], regs->regs[52], regs->tp);
632 pr_err(" sp : "REGFMT" lr : "REGFMT"\n", regs->sp, regs->lr);
634 for (i = 0; i < 52; i += 4)
635 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT
636 " 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], i+3, regs->regs[i+3]);
639 pr_err(" r52: "REGFMT" tp : "REGFMT" sp : "REGFMT" lr : "REGFMT"\n",
640 regs->regs[52], regs->tp, regs->sp, regs->lr);
642 pr_err(" pc : "REGFMT" ex1: %ld faultnum: %ld\n",
643 regs->pc, regs->ex1, regs->faultnum);
645 dump_stack_regs(regs);