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 <asm/system.h>
29 #include <asm/stack.h>
30 #include <asm/homecache.h>
31 #include <asm/syscalls.h>
32 #ifdef CONFIG_HARDWALL
33 #include <asm/hardwall.h>
35 #include <arch/chip.h>
40 * Use the (x86) "idle=poll" option to prefer low latency when leaving the
41 * idle loop over low power while in the idle loop, e.g. if we have
42 * one thread per core and we want to get threads out of futex waits fast.
44 static int no_idle_nap;
45 static int __init idle_setup(char *str)
50 if (!strcmp(str, "poll")) {
51 pr_info("using polling idle threads.\n");
53 } else if (!strcmp(str, "halt"))
60 early_param("idle", idle_setup);
63 * The idle thread. There's no useful work to be
64 * done, so just try to conserve power and have a
65 * low exit latency (ie sit in a loop waiting for
66 * somebody to say that they'd like to reschedule)
70 int cpu = smp_processor_id();
73 current_thread_info()->status |= TS_POLLING;
77 while (!need_resched())
83 /* endless idle loop with no priority at all */
85 tick_nohz_stop_sched_tick(1);
86 while (!need_resched()) {
87 if (cpu_is_offline(cpu))
88 BUG(); /* no HOTPLUG_CPU */
91 __get_cpu_var(irq_stat).idle_timestamp = jiffies;
92 current_thread_info()->status &= ~TS_POLLING;
94 * TS_POLLING-cleared state must be visible before we
103 current_thread_info()->status |= TS_POLLING;
105 tick_nohz_restart_sched_tick();
106 preempt_enable_no_resched();
112 struct thread_info *alloc_thread_info(struct task_struct *task)
115 gfp_t flags = GFP_KERNEL;
117 #ifdef CONFIG_DEBUG_STACK_USAGE
121 page = alloc_pages(flags, THREAD_SIZE_ORDER);
125 return (struct thread_info *)page_address(page);
129 * Free a thread_info node, and all of its derivative
132 void free_thread_info(struct thread_info *info)
134 struct single_step_state *step_state = info->step_state;
136 #ifdef CONFIG_HARDWALL
138 * We free a thread_info from the context of the task that has
139 * been scheduled next, so the original task is already dead.
140 * Calling deactivate here just frees up the data structures.
141 * If the task we're freeing held the last reference to a
142 * hardwall fd, it would have been released prior to this point
143 * anyway via exit_files(), and "hardwall" would be NULL by now.
145 if (info->task->thread.hardwall)
146 hardwall_deactivate(info->task);
152 * FIXME: we don't munmap step_state->buffer
153 * because the mm_struct for this process (info->task->mm)
154 * has already been zeroed in exit_mm(). Keeping a
155 * reference to it here seems like a bad move, so this
156 * means we can't munmap() the buffer, and therefore if we
157 * ptrace multiple threads in a process, we will slowly
158 * leak user memory. (Note that as soon as the last
159 * thread in a process dies, we will reclaim all user
160 * memory including single-step buffers in the usual way.)
161 * We should either assign a kernel VA to this buffer
162 * somehow, or we should associate the buffer(s) with the
163 * mm itself so we can clean them up that way.
168 free_page((unsigned long)info);
171 static void save_arch_state(struct thread_struct *t);
173 int copy_thread(unsigned long clone_flags, unsigned long sp,
174 unsigned long stack_size,
175 struct task_struct *p, struct pt_regs *regs)
177 struct pt_regs *childregs;
181 * When creating a new kernel thread we pass sp as zero.
182 * Assign it to a reasonable value now that we have the stack.
184 if (sp == 0 && regs->ex1 == PL_ICS_EX1(KERNEL_PL, 0))
188 * Do not clone step state from the parent; each thread
189 * must make its own lazily.
191 task_thread_info(p)->step_state = NULL;
194 * Start new thread in ret_from_fork so it schedules properly
195 * and then return from interrupt like the parent.
197 p->thread.pc = (unsigned long) ret_from_fork;
199 /* Save user stack top pointer so we can ID the stack vm area later. */
202 /* Record the pid of the process that created this one. */
203 p->thread.creator_pid = current->pid;
206 * Copy the registers onto the kernel stack so the
207 * return-from-interrupt code will reload it into registers.
209 childregs = task_pt_regs(p);
211 childregs->regs[0] = 0; /* return value is zero */
212 childregs->sp = sp; /* override with new user stack pointer */
215 * If CLONE_SETTLS is set, set "tp" in the new task to "r4",
216 * which is passed in as arg #5 to sys_clone().
218 if (clone_flags & CLONE_SETTLS)
219 childregs->tp = regs->regs[4];
222 * Copy the callee-saved registers from the passed pt_regs struct
223 * into the context-switch callee-saved registers area.
224 * This way when we start the interrupt-return sequence, the
225 * callee-save registers will be correctly in registers, which
226 * is how we assume the compiler leaves them as we start doing
227 * the normal return-from-interrupt path after calling C code.
228 * Zero out the C ABI save area to mark the top of the stack.
230 ksp = (unsigned long) childregs;
231 ksp -= C_ABI_SAVE_AREA_SIZE; /* interrupt-entry save area */
232 ((long *)ksp)[0] = ((long *)ksp)[1] = 0;
233 ksp -= CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long);
234 memcpy((void *)ksp, ®s->regs[CALLEE_SAVED_FIRST_REG],
235 CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long));
236 ksp -= C_ABI_SAVE_AREA_SIZE; /* __switch_to() save area */
237 ((long *)ksp)[0] = ((long *)ksp)[1] = 0;
240 #if CHIP_HAS_TILE_DMA()
242 * No DMA in the new thread. We model this on the fact that
243 * fork() clears the pending signals, alarms, and aio for the child.
245 memset(&p->thread.tile_dma_state, 0, sizeof(struct tile_dma_state));
246 memset(&p->thread.dma_async_tlb, 0, sizeof(struct async_tlb));
249 #if CHIP_HAS_SN_PROC()
250 /* Likewise, the new thread is not running static processor code. */
251 p->thread.sn_proc_running = 0;
252 memset(&p->thread.sn_async_tlb, 0, sizeof(struct async_tlb));
255 #if CHIP_HAS_PROC_STATUS_SPR()
256 /* New thread has its miscellaneous processor state bits clear. */
257 p->thread.proc_status = 0;
260 #ifdef CONFIG_HARDWALL
261 /* New thread does not own any networks. */
262 p->thread.hardwall = NULL;
267 * Start the new thread with the current architecture state
268 * (user interrupt masks, etc.).
270 save_arch_state(&p->thread);
276 * Return "current" if it looks plausible, or else a pointer to a dummy.
277 * This can be helpful if we are just trying to emit a clean panic.
279 struct task_struct *validate_current(void)
281 static struct task_struct corrupt = { .comm = "<corrupt>" };
282 struct task_struct *tsk = current;
283 if (unlikely((unsigned long)tsk < PAGE_OFFSET ||
284 (void *)tsk > high_memory ||
285 ((unsigned long)tsk & (__alignof__(*tsk) - 1)) != 0)) {
286 pr_err("Corrupt 'current' %p (sp %#lx)\n", tsk, stack_pointer);
292 /* Take and return the pointer to the previous task, for schedule_tail(). */
293 struct task_struct *sim_notify_fork(struct task_struct *prev)
295 struct task_struct *tsk = current;
296 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK_PARENT |
297 (tsk->thread.creator_pid << _SIM_CONTROL_OPERATOR_BITS));
298 __insn_mtspr(SPR_SIM_CONTROL, SIM_CONTROL_OS_FORK |
299 (tsk->pid << _SIM_CONTROL_OPERATOR_BITS));
303 int dump_task_regs(struct task_struct *tsk, elf_gregset_t *regs)
305 struct pt_regs *ptregs = task_pt_regs(tsk);
306 elf_core_copy_regs(regs, ptregs);
310 #if CHIP_HAS_TILE_DMA()
312 /* Allow user processes to access the DMA SPRs */
313 void grant_dma_mpls(void)
315 #if CONFIG_KERNEL_PL == 2
316 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
317 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
319 __insn_mtspr(SPR_MPL_DMA_CPL_SET_0, 1);
320 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_0, 1);
324 /* Forbid user processes from accessing the DMA SPRs */
325 void restrict_dma_mpls(void)
327 #if CONFIG_KERNEL_PL == 2
328 __insn_mtspr(SPR_MPL_DMA_CPL_SET_2, 1);
329 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_2, 1);
331 __insn_mtspr(SPR_MPL_DMA_CPL_SET_1, 1);
332 __insn_mtspr(SPR_MPL_DMA_NOTIFY_SET_1, 1);
336 /* Pause the DMA engine, then save off its state registers. */
337 static void save_tile_dma_state(struct tile_dma_state *dma)
339 unsigned long state = __insn_mfspr(SPR_DMA_USER_STATUS);
340 unsigned long post_suspend_state;
342 /* If we're running, suspend the engine. */
343 if ((state & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK)
344 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
347 * Wait for the engine to idle, then save regs. Note that we
348 * want to record the "running" bit from before suspension,
349 * and the "done" bit from after, so that we can properly
350 * distinguish a case where the user suspended the engine from
351 * the case where the kernel suspended as part of the context
355 post_suspend_state = __insn_mfspr(SPR_DMA_USER_STATUS);
356 } while (post_suspend_state & SPR_DMA_STATUS__BUSY_MASK);
358 dma->src = __insn_mfspr(SPR_DMA_SRC_ADDR);
359 dma->src_chunk = __insn_mfspr(SPR_DMA_SRC_CHUNK_ADDR);
360 dma->dest = __insn_mfspr(SPR_DMA_DST_ADDR);
361 dma->dest_chunk = __insn_mfspr(SPR_DMA_DST_CHUNK_ADDR);
362 dma->strides = __insn_mfspr(SPR_DMA_STRIDE);
363 dma->chunk_size = __insn_mfspr(SPR_DMA_CHUNK_SIZE);
364 dma->byte = __insn_mfspr(SPR_DMA_BYTE);
365 dma->status = (state & SPR_DMA_STATUS__RUNNING_MASK) |
366 (post_suspend_state & SPR_DMA_STATUS__DONE_MASK);
369 /* Restart a DMA that was running before we were context-switched out. */
370 static void restore_tile_dma_state(struct thread_struct *t)
372 const struct tile_dma_state *dma = &t->tile_dma_state;
375 * The only way to restore the done bit is to run a zero
376 * length transaction.
378 if ((dma->status & SPR_DMA_STATUS__DONE_MASK) &&
379 !(__insn_mfspr(SPR_DMA_USER_STATUS) & SPR_DMA_STATUS__DONE_MASK)) {
380 __insn_mtspr(SPR_DMA_BYTE, 0);
381 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
382 while (__insn_mfspr(SPR_DMA_USER_STATUS) &
383 SPR_DMA_STATUS__BUSY_MASK)
387 __insn_mtspr(SPR_DMA_SRC_ADDR, dma->src);
388 __insn_mtspr(SPR_DMA_SRC_CHUNK_ADDR, dma->src_chunk);
389 __insn_mtspr(SPR_DMA_DST_ADDR, dma->dest);
390 __insn_mtspr(SPR_DMA_DST_CHUNK_ADDR, dma->dest_chunk);
391 __insn_mtspr(SPR_DMA_STRIDE, dma->strides);
392 __insn_mtspr(SPR_DMA_CHUNK_SIZE, dma->chunk_size);
393 __insn_mtspr(SPR_DMA_BYTE, dma->byte);
396 * Restart the engine if we were running and not done.
397 * Clear a pending async DMA fault that we were waiting on return
398 * to user space to execute, since we expect the DMA engine
399 * to regenerate those faults for us now. Note that we don't
400 * try to clear the TIF_ASYNC_TLB flag, since it's relatively
401 * harmless if set, and it covers both DMA and the SN processor.
403 if ((dma->status & DMA_STATUS_MASK) == SPR_DMA_STATUS__RUNNING_MASK) {
404 t->dma_async_tlb.fault_num = 0;
405 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
411 static void save_arch_state(struct thread_struct *t)
413 #if CHIP_HAS_SPLIT_INTR_MASK()
414 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0_0) |
415 ((u64)__insn_mfspr(SPR_INTERRUPT_MASK_0_1) << 32);
417 t->interrupt_mask = __insn_mfspr(SPR_INTERRUPT_MASK_0);
419 t->ex_context[0] = __insn_mfspr(SPR_EX_CONTEXT_0_0);
420 t->ex_context[1] = __insn_mfspr(SPR_EX_CONTEXT_0_1);
421 t->system_save[0] = __insn_mfspr(SPR_SYSTEM_SAVE_0_0);
422 t->system_save[1] = __insn_mfspr(SPR_SYSTEM_SAVE_0_1);
423 t->system_save[2] = __insn_mfspr(SPR_SYSTEM_SAVE_0_2);
424 t->system_save[3] = __insn_mfspr(SPR_SYSTEM_SAVE_0_3);
425 t->intctrl_0 = __insn_mfspr(SPR_INTCTRL_0_STATUS);
426 #if CHIP_HAS_PROC_STATUS_SPR()
427 t->proc_status = __insn_mfspr(SPR_PROC_STATUS);
429 #if !CHIP_HAS_FIXED_INTVEC_BASE()
430 t->interrupt_vector_base = __insn_mfspr(SPR_INTERRUPT_VECTOR_BASE_0);
432 #if CHIP_HAS_TILE_RTF_HWM()
433 t->tile_rtf_hwm = __insn_mfspr(SPR_TILE_RTF_HWM);
435 #if CHIP_HAS_DSTREAM_PF()
436 t->dstream_pf = __insn_mfspr(SPR_DSTREAM_PF);
440 static void restore_arch_state(const struct thread_struct *t)
442 #if CHIP_HAS_SPLIT_INTR_MASK()
443 __insn_mtspr(SPR_INTERRUPT_MASK_0_0, (u32) t->interrupt_mask);
444 __insn_mtspr(SPR_INTERRUPT_MASK_0_1, t->interrupt_mask >> 32);
446 __insn_mtspr(SPR_INTERRUPT_MASK_0, t->interrupt_mask);
448 __insn_mtspr(SPR_EX_CONTEXT_0_0, t->ex_context[0]);
449 __insn_mtspr(SPR_EX_CONTEXT_0_1, t->ex_context[1]);
450 __insn_mtspr(SPR_SYSTEM_SAVE_0_0, t->system_save[0]);
451 __insn_mtspr(SPR_SYSTEM_SAVE_0_1, t->system_save[1]);
452 __insn_mtspr(SPR_SYSTEM_SAVE_0_2, t->system_save[2]);
453 __insn_mtspr(SPR_SYSTEM_SAVE_0_3, t->system_save[3]);
454 __insn_mtspr(SPR_INTCTRL_0_STATUS, t->intctrl_0);
455 #if CHIP_HAS_PROC_STATUS_SPR()
456 __insn_mtspr(SPR_PROC_STATUS, t->proc_status);
458 #if !CHIP_HAS_FIXED_INTVEC_BASE()
459 __insn_mtspr(SPR_INTERRUPT_VECTOR_BASE_0, t->interrupt_vector_base);
461 #if CHIP_HAS_TILE_RTF_HWM()
462 __insn_mtspr(SPR_TILE_RTF_HWM, t->tile_rtf_hwm);
464 #if CHIP_HAS_DSTREAM_PF()
465 __insn_mtspr(SPR_DSTREAM_PF, t->dstream_pf);
470 void _prepare_arch_switch(struct task_struct *next)
472 #if CHIP_HAS_SN_PROC()
475 #if CHIP_HAS_TILE_DMA()
476 struct tile_dma_state *dma = ¤t->thread.tile_dma_state;
478 save_tile_dma_state(dma);
480 #if CHIP_HAS_SN_PROC()
482 * Suspend the static network processor if it was running.
483 * We do not suspend the fabric itself, just like we don't
484 * try to suspend the UDN.
486 snctl = __insn_mfspr(SPR_SNCTL);
487 current->thread.sn_proc_running =
488 (snctl & SPR_SNCTL__FRZPROC_MASK) == 0;
489 if (current->thread.sn_proc_running)
490 __insn_mtspr(SPR_SNCTL, snctl | SPR_SNCTL__FRZPROC_MASK);
495 struct task_struct *__sched _switch_to(struct task_struct *prev,
496 struct task_struct *next)
498 /* DMA state is already saved; save off other arch state. */
499 save_arch_state(&prev->thread);
501 #if CHIP_HAS_TILE_DMA()
503 * Restore DMA in new task if desired.
504 * Note that it is only safe to restart here since interrupts
505 * are disabled, so we can't take any DMATLB miss or access
506 * interrupts before we have finished switching stacks.
508 if (next->thread.tile_dma_state.enabled) {
509 restore_tile_dma_state(&next->thread);
516 /* Restore other arch state. */
517 restore_arch_state(&next->thread);
519 #if CHIP_HAS_SN_PROC()
521 * Restart static network processor in the new process
522 * if it was running before.
524 if (next->thread.sn_proc_running) {
525 int snctl = __insn_mfspr(SPR_SNCTL);
526 __insn_mtspr(SPR_SNCTL, snctl & ~SPR_SNCTL__FRZPROC_MASK);
530 #ifdef CONFIG_HARDWALL
531 /* Enable or disable access to the network registers appropriately. */
532 if (prev->thread.hardwall != NULL) {
533 if (next->thread.hardwall == NULL)
534 restrict_network_mpls();
535 } else if (next->thread.hardwall != NULL) {
536 grant_network_mpls();
541 * Switch kernel SP, PC, and callee-saved registers.
542 * In the context of the new task, return the old task pointer
543 * (i.e. the task that actually called __switch_to).
544 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp.
546 return __switch_to(prev, next, next_current_ksp0(next));
549 /* Note there is an implicit fifth argument if (clone_flags & CLONE_SETTLS). */
550 SYSCALL_DEFINE5(clone, unsigned long, clone_flags, unsigned long, newsp,
551 void __user *, parent_tidptr, void __user *, child_tidptr,
552 struct pt_regs *, regs)
556 return do_fork(clone_flags, newsp, regs, 0,
557 parent_tidptr, child_tidptr);
561 * sys_execve() executes a new program.
563 SYSCALL_DEFINE4(execve, const char __user *, path,
564 const char __user *const __user *, argv,
565 const char __user *const __user *, envp,
566 struct pt_regs *, regs)
571 filename = getname(path);
572 error = PTR_ERR(filename);
573 if (IS_ERR(filename))
575 error = do_execve(filename, argv, envp, regs);
582 long compat_sys_execve(const char __user *path,
583 const compat_uptr_t __user *argv,
584 const compat_uptr_t __user *envp,
585 struct pt_regs *regs)
590 filename = getname(path);
591 error = PTR_ERR(filename);
592 if (IS_ERR(filename))
594 error = compat_do_execve(filename, argv, envp, regs);
601 unsigned long get_wchan(struct task_struct *p)
603 struct KBacktraceIterator kbt;
605 if (!p || p == current || p->state == TASK_RUNNING)
608 for (KBacktraceIterator_init(&kbt, p, NULL);
609 !KBacktraceIterator_end(&kbt);
610 KBacktraceIterator_next(&kbt)) {
611 if (!in_sched_functions(kbt.it.pc))
619 * We pass in lr as zero (cleared in kernel_thread) and the caller
620 * part of the backtrace ABI on the stack also zeroed (in copy_thread)
621 * so that backtraces will stop with this function.
622 * Note that we don't use r0, since copy_thread() clears it.
624 static void start_kernel_thread(int dummy, int (*fn)(int), int arg)
630 * Create a kernel thread
632 int kernel_thread(int (*fn)(void *), void * arg, unsigned long flags)
636 memset(®s, 0, sizeof(regs));
637 regs.ex1 = PL_ICS_EX1(KERNEL_PL, 0); /* run at kernel PL, no ICS */
638 regs.pc = (long) start_kernel_thread;
639 regs.flags = PT_FLAGS_CALLER_SAVES; /* need to restore r1 and r2 */
640 regs.regs[1] = (long) fn; /* function pointer */
641 regs.regs[2] = (long) arg; /* parameter register */
643 /* Ok, create the new process.. */
644 return do_fork(flags | CLONE_VM | CLONE_UNTRACED, 0, ®s,
647 EXPORT_SYMBOL(kernel_thread);
649 /* Flush thread state. */
650 void flush_thread(void)
656 * Free current thread data structures etc..
658 void exit_thread(void)
663 void show_regs(struct pt_regs *regs)
665 struct task_struct *tsk = validate_current();
669 pr_err(" Pid: %d, comm: %20s, CPU: %d\n",
670 tsk->pid, tsk->comm, smp_processor_id());
672 for (i = 0; i < 51; i += 3)
673 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
674 i, regs->regs[i], i+1, regs->regs[i+1],
675 i+2, regs->regs[i+2]);
676 pr_err(" r51: "REGFMT" r52: "REGFMT" tp : "REGFMT"\n",
677 regs->regs[51], regs->regs[52], regs->tp);
678 pr_err(" sp : "REGFMT" lr : "REGFMT"\n", regs->sp, regs->lr);
680 for (i = 0; i < 52; i += 4)
681 pr_err(" r%-2d: "REGFMT" r%-2d: "REGFMT
682 " r%-2d: "REGFMT" r%-2d: "REGFMT"\n",
683 i, regs->regs[i], i+1, regs->regs[i+1],
684 i+2, regs->regs[i+2], i+3, regs->regs[i+3]);
685 pr_err(" r52: "REGFMT" tp : "REGFMT" sp : "REGFMT" lr : "REGFMT"\n",
686 regs->regs[52], regs->tp, regs->sp, regs->lr);
688 pr_err(" pc : "REGFMT" ex1: %ld faultnum: %ld\n",
689 regs->pc, regs->ex1, regs->faultnum);
691 dump_stack_regs(regs);