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/context_tracking.h> 31#include <asm/stack.h> 32#include <asm/switch_to.h> 33#include <asm/homecache.h> 34#include <asm/syscalls.h> 35#include <asm/traps.h> 36#include <asm/setup.h> 37#include <asm/uaccess.h> 38#ifdef CONFIG_HARDWALL 39#include <asm/hardwall.h> 40#endif 41#include <arch/chip.h> 42#include <arch/abi.h> 43#include <arch/sim_def.h> 44 45/* 46 * Use the (x86) "idle=poll" option to prefer low latency when leaving the 47 * idle loop over low power while in the idle loop, e.g. if we have 48 * one thread per core and we want to get threads out of futex waits fast. 49 */ 50static int __init idle_setup(char *str) 51{ 52 if (!str) 53 return -EINVAL; 54 55 if (!strcmp(str, "poll")) { 56 pr_info("using polling idle threads\n"); 57 cpu_idle_poll_ctrl(true); 58 return 0; 59 } else if (!strcmp(str, "halt")) { 60 return 0; 61 } 62 return -1; 63} 64early_param("idle", idle_setup); 65 66void arch_cpu_idle(void) 67{ 68 __this_cpu_write(irq_stat.idle_timestamp, jiffies); 69 _cpu_idle(); 70} 71 72/* 73 * Release a thread_info structure 74 */ 75void arch_release_thread_info(struct thread_info *info) 76{ 77 struct single_step_state *step_state = info->step_state; 78 79 if (step_state) { 80 81 /* 82 * FIXME: we don't munmap step_state->buffer 83 * because the mm_struct for this process (info->task->mm) 84 * has already been zeroed in exit_mm(). Keeping a 85 * reference to it here seems like a bad move, so this 86 * means we can't munmap() the buffer, and therefore if we 87 * ptrace multiple threads in a process, we will slowly 88 * leak user memory. (Note that as soon as the last 89 * thread in a process dies, we will reclaim all user 90 * memory including single-step buffers in the usual way.) 91 * We should either assign a kernel VA to this buffer 92 * somehow, or we should associate the buffer(s) with the 93 * mm itself so we can clean them up that way. 94 */ 95 kfree(step_state); 96 } 97} 98 99static void save_arch_state(struct thread_struct *t); 100 101int copy_thread(unsigned long clone_flags, unsigned long sp, 102 unsigned long arg, struct task_struct *p) 103{ 104 struct pt_regs *childregs = task_pt_regs(p); 105 unsigned long ksp; 106 unsigned long *callee_regs; 107 108 /* 109 * Set up the stack and stack pointer appropriately for the 110 * new child to find itself woken up in __switch_to(). 111 * The callee-saved registers must be on the stack to be read; 112 * the new task will then jump to assembly support to handle 113 * calling schedule_tail(), etc., and (for userspace tasks) 114 * returning to the context set up in the pt_regs. 115 */ 116 ksp = (unsigned long) childregs; 117 ksp -= C_ABI_SAVE_AREA_SIZE; /* interrupt-entry save area */ 118 ((long *)ksp)[0] = ((long *)ksp)[1] = 0; 119 ksp -= CALLEE_SAVED_REGS_COUNT * sizeof(unsigned long); 120 callee_regs = (unsigned long *)ksp; 121 ksp -= C_ABI_SAVE_AREA_SIZE; /* __switch_to() save area */ 122 ((long *)ksp)[0] = ((long *)ksp)[1] = 0; 123 p->thread.ksp = ksp; 124 125 /* Record the pid of the task that created this one. */ 126 p->thread.creator_pid = current->pid; 127 128 if (unlikely(p->flags & PF_KTHREAD)) { 129 /* kernel thread */ 130 memset(childregs, 0, sizeof(struct pt_regs)); 131 memset(&callee_regs[2], 0, 132 (CALLEE_SAVED_REGS_COUNT - 2) * sizeof(unsigned long)); 133 callee_regs[0] = sp; /* r30 = function */ 134 callee_regs[1] = arg; /* r31 = arg */ 135 childregs->ex1 = PL_ICS_EX1(KERNEL_PL, 0); 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 /* 450 * Switch kernel SP, PC, and callee-saved registers. 451 * In the context of the new task, return the old task pointer 452 * (i.e. the task that actually called __switch_to). 453 * Pass the value to use for SYSTEM_SAVE_K_0 when we reset our sp. 454 */ 455 return __switch_to(prev, next, next_current_ksp0(next)); 456} 457 458/* 459 * This routine is called on return from interrupt if any of the 460 * TIF_WORK_MASK flags are set in thread_info->flags. It is 461 * entered with interrupts disabled so we don't miss an event 462 * that modified the thread_info flags. If any flag is set, we 463 * handle it and return, and the calling assembly code will 464 * re-disable interrupts, reload the thread flags, and call back 465 * if more flags need to be handled. 466 * 467 * We return whether we need to check the thread_info flags again 468 * or not. Note that we don't clear TIF_SINGLESTEP here, so it's 469 * important that it be tested last, and then claim that we don't 470 * need to recheck the flags. 471 */ 472int do_work_pending(struct pt_regs *regs, u32 thread_info_flags) 473{ 474 /* If we enter in kernel mode, do nothing and exit the caller loop. */ 475 if (!user_mode(regs)) 476 return 0; 477 478 user_exit(); 479 480 /* Enable interrupts; they are disabled again on return to caller. */ 481 local_irq_enable(); 482 483 if (thread_info_flags & _TIF_NEED_RESCHED) { 484 schedule(); 485 return 1; 486 } 487#if CHIP_HAS_TILE_DMA() 488 if (thread_info_flags & _TIF_ASYNC_TLB) { 489 do_async_page_fault(regs); 490 return 1; 491 } 492#endif 493 if (thread_info_flags & _TIF_SIGPENDING) { 494 do_signal(regs); 495 return 1; 496 } 497 if (thread_info_flags & _TIF_NOTIFY_RESUME) { 498 clear_thread_flag(TIF_NOTIFY_RESUME); 499 tracehook_notify_resume(regs); 500 return 1; 501 } 502 if (thread_info_flags & _TIF_SINGLESTEP) 503 single_step_once(regs); 504 505 user_enter(); 506 507 return 0; 508} 509 510unsigned long get_wchan(struct task_struct *p) 511{ 512 struct KBacktraceIterator kbt; 513 514 if (!p || p == current || p->state == TASK_RUNNING) 515 return 0; 516 517 for (KBacktraceIterator_init(&kbt, p, NULL); 518 !KBacktraceIterator_end(&kbt); 519 KBacktraceIterator_next(&kbt)) { 520 if (!in_sched_functions(kbt.it.pc)) 521 return kbt.it.pc; 522 } 523 524 return 0; 525} 526 527/* Flush thread state. */ 528void flush_thread(void) 529{ 530 /* Nothing */ 531} 532 533/* 534 * Free current thread data structures etc.. 535 */ 536void exit_thread(void) 537{ 538#ifdef CONFIG_HARDWALL 539 /* 540 * Remove the task from the list of tasks that are associated 541 * with any live hardwalls. (If the task that is exiting held 542 * the last reference to a hardwall fd, it would already have 543 * been released and deactivated at this point.) 544 */ 545 hardwall_deactivate_all(current); 546#endif 547} 548 549void show_regs(struct pt_regs *regs) 550{ 551 struct task_struct *tsk = validate_current(); 552 int i; 553 554 if (tsk != &corrupt_current) 555 show_regs_print_info(KERN_ERR); 556#ifdef __tilegx__ 557 for (i = 0; i < 17; i++) 558 pr_err(" r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT "\n", 559 i, regs->regs[i], i+18, regs->regs[i+18], 560 i+36, regs->regs[i+36]); 561 pr_err(" r17: " REGFMT " r35: " REGFMT " tp : " REGFMT "\n", 562 regs->regs[17], regs->regs[35], regs->tp); 563 pr_err(" sp : " REGFMT " lr : " REGFMT "\n", regs->sp, regs->lr); 564#else 565 for (i = 0; i < 13; i++) 566 pr_err(" r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT " r%-2d: " REGFMT "\n", 567 i, regs->regs[i], i+14, regs->regs[i+14], 568 i+27, regs->regs[i+27], i+40, regs->regs[i+40]); 569 pr_err(" r13: " REGFMT " tp : " REGFMT " sp : " REGFMT " lr : " REGFMT "\n", 570 regs->regs[13], regs->tp, regs->sp, regs->lr); 571#endif 572 pr_err(" pc : " REGFMT " ex1: %ld faultnum: %ld\n", 573 regs->pc, regs->ex1, regs->faultnum); 574 575 dump_stack_regs(regs); 576} 577