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 * From i386 code copyright (C) 1995  Linus Torvalds
15 */
16
17#include <linux/signal.h>
18#include <linux/sched.h>
19#include <linux/kernel.h>
20#include <linux/errno.h>
21#include <linux/string.h>
22#include <linux/types.h>
23#include <linux/ptrace.h>
24#include <linux/mman.h>
25#include <linux/mm.h>
26#include <linux/smp.h>
27#include <linux/interrupt.h>
28#include <linux/init.h>
29#include <linux/tty.h>
30#include <linux/vt_kern.h>		/* For unblank_screen() */
31#include <linux/highmem.h>
32#include <linux/module.h>
33#include <linux/kprobes.h>
34#include <linux/hugetlb.h>
35#include <linux/syscalls.h>
36#include <linux/uaccess.h>
37#include <linux/kdebug.h>
38#include <linux/context_tracking.h>
39
40#include <asm/pgalloc.h>
41#include <asm/sections.h>
42#include <asm/traps.h>
43#include <asm/syscalls.h>
44
45#include <arch/interrupts.h>
46
47static noinline void force_sig_info_fault(const char *type, int si_signo,
48					  int si_code, unsigned long address,
49					  int fault_num,
50					  struct task_struct *tsk,
51					  struct pt_regs *regs)
52{
53	siginfo_t info;
54
55	if (unlikely(tsk->pid < 2)) {
56		panic("Signal %d (code %d) at %#lx sent to %s!",
57		      si_signo, si_code & 0xffff, address,
58		      is_idle_task(tsk) ? "the idle task" : "init");
59	}
60
61	info.si_signo = si_signo;
62	info.si_errno = 0;
63	info.si_code = si_code;
64	info.si_addr = (void __user *)address;
65	info.si_trapno = fault_num;
66	trace_unhandled_signal(type, regs, address, si_signo);
67	force_sig_info(si_signo, &info, tsk);
68}
69
70#ifndef __tilegx__
71/*
72 * Synthesize the fault a PL0 process would get by doing a word-load of
73 * an unaligned address or a high kernel address.
74 */
75SYSCALL_DEFINE1(cmpxchg_badaddr, unsigned long, address)
76{
77	struct pt_regs *regs = current_pt_regs();
78
79	if (address >= PAGE_OFFSET)
80		force_sig_info_fault("atomic segfault", SIGSEGV, SEGV_MAPERR,
81				     address, INT_DTLB_MISS, current, regs);
82	else
83		force_sig_info_fault("atomic alignment fault", SIGBUS,
84				     BUS_ADRALN, address,
85				     INT_UNALIGN_DATA, current, regs);
86
87	/*
88	 * Adjust pc to point at the actual instruction, which is unusual
89	 * for syscalls normally, but is appropriate when we are claiming
90	 * that a syscall swint1 caused a page fault or bus error.
91	 */
92	regs->pc -= 8;
93
94	/*
95	 * Mark this as a caller-save interrupt, like a normal page fault,
96	 * so that when we go through the signal handler path we will
97	 * properly restore r0, r1, and r2 for the signal handler arguments.
98	 */
99	regs->flags |= PT_FLAGS_CALLER_SAVES;
100
101	return 0;
102}
103#endif
104
105static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
106{
107	unsigned index = pgd_index(address);
108	pgd_t *pgd_k;
109	pud_t *pud, *pud_k;
110	pmd_t *pmd, *pmd_k;
111
112	pgd += index;
113	pgd_k = init_mm.pgd + index;
114
115	if (!pgd_present(*pgd_k))
116		return NULL;
117
118	pud = pud_offset(pgd, address);
119	pud_k = pud_offset(pgd_k, address);
120	if (!pud_present(*pud_k))
121		return NULL;
122
123	pmd = pmd_offset(pud, address);
124	pmd_k = pmd_offset(pud_k, address);
125	if (!pmd_present(*pmd_k))
126		return NULL;
127	if (!pmd_present(*pmd))
128		set_pmd(pmd, *pmd_k);
129	else
130		BUG_ON(pmd_ptfn(*pmd) != pmd_ptfn(*pmd_k));
131	return pmd_k;
132}
133
134/*
135 * Handle a fault on the vmalloc area.
136 */
137static inline int vmalloc_fault(pgd_t *pgd, unsigned long address)
138{
139	pmd_t *pmd_k;
140	pte_t *pte_k;
141
142	/* Make sure we are in vmalloc area */
143	if (!(address >= VMALLOC_START && address < VMALLOC_END))
144		return -1;
145
146	/*
147	 * Synchronize this task's top level page-table
148	 * with the 'reference' page table.
149	 */
150	pmd_k = vmalloc_sync_one(pgd, address);
151	if (!pmd_k)
152		return -1;
153	pte_k = pte_offset_kernel(pmd_k, address);
154	if (!pte_present(*pte_k))
155		return -1;
156	return 0;
157}
158
159/* Wait until this PTE has completed migration. */
160static void wait_for_migration(pte_t *pte)
161{
162	if (pte_migrating(*pte)) {
163		/*
164		 * Wait until the migrater fixes up this pte.
165		 * We scale the loop count by the clock rate so we'll wait for
166		 * a few seconds here.
167		 */
168		int retries = 0;
169		int bound = get_clock_rate();
170		while (pte_migrating(*pte)) {
171			barrier();
172			if (++retries > bound)
173				panic("Hit migrating PTE (%#llx) and page PFN %#lx still migrating",
174				      pte->val, pte_pfn(*pte));
175		}
176	}
177}
178
179/*
180 * It's not generally safe to use "current" to get the page table pointer,
181 * since we might be running an oprofile interrupt in the middle of a
182 * task switch.
183 */
184static pgd_t *get_current_pgd(void)
185{
186	HV_Context ctx = hv_inquire_context();
187	unsigned long pgd_pfn = ctx.page_table >> PAGE_SHIFT;
188	struct page *pgd_page = pfn_to_page(pgd_pfn);
189	BUG_ON(PageHighMem(pgd_page));
190	return (pgd_t *) __va(ctx.page_table);
191}
192
193/*
194 * We can receive a page fault from a migrating PTE at any time.
195 * Handle it by just waiting until the fault resolves.
196 *
197 * It's also possible to get a migrating kernel PTE that resolves
198 * itself during the downcall from hypervisor to Linux.  We just check
199 * here to see if the PTE seems valid, and if so we retry it.
200 *
201 * NOTE! We MUST NOT take any locks for this case.  We may be in an
202 * interrupt or a critical region, and must do as little as possible.
203 * Similarly, we can't use atomic ops here, since we may be handling a
204 * fault caused by an atomic op access.
205 *
206 * If we find a migrating PTE while we're in an NMI context, and we're
207 * at a PC that has a registered exception handler, we don't wait,
208 * since this thread may (e.g.) have been interrupted while migrating
209 * its own stack, which would then cause us to self-deadlock.
210 */
211static int handle_migrating_pte(pgd_t *pgd, int fault_num,
212				unsigned long address, unsigned long pc,
213				int is_kernel_mode, int write)
214{
215	pud_t *pud;
216	pmd_t *pmd;
217	pte_t *pte;
218	pte_t pteval;
219
220	if (pgd_addr_invalid(address))
221		return 0;
222
223	pgd += pgd_index(address);
224	pud = pud_offset(pgd, address);
225	if (!pud || !pud_present(*pud))
226		return 0;
227	pmd = pmd_offset(pud, address);
228	if (!pmd || !pmd_present(*pmd))
229		return 0;
230	pte = pmd_huge_page(*pmd) ? ((pte_t *)pmd) :
231		pte_offset_kernel(pmd, address);
232	pteval = *pte;
233	if (pte_migrating(pteval)) {
234		if (in_nmi() && search_exception_tables(pc))
235			return 0;
236		wait_for_migration(pte);
237		return 1;
238	}
239
240	if (!is_kernel_mode || !pte_present(pteval))
241		return 0;
242	if (fault_num == INT_ITLB_MISS) {
243		if (pte_exec(pteval))
244			return 1;
245	} else if (write) {
246		if (pte_write(pteval))
247			return 1;
248	} else {
249		if (pte_read(pteval))
250			return 1;
251	}
252
253	return 0;
254}
255
256/*
257 * This routine is responsible for faulting in user pages.
258 * It passes the work off to one of the appropriate routines.
259 * It returns true if the fault was successfully handled.
260 */
261static int handle_page_fault(struct pt_regs *regs,
262			     int fault_num,
263			     int is_page_fault,
264			     unsigned long address,
265			     int write)
266{
267	struct task_struct *tsk;
268	struct mm_struct *mm;
269	struct vm_area_struct *vma;
270	unsigned long stack_offset;
271	int fault;
272	int si_code;
273	int is_kernel_mode;
274	pgd_t *pgd;
275	unsigned int flags;
276
277	/* on TILE, protection faults are always writes */
278	if (!is_page_fault)
279		write = 1;
280
281	flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
282
283	is_kernel_mode = !user_mode(regs);
284
285	tsk = validate_current();
286
287	/*
288	 * Check to see if we might be overwriting the stack, and bail
289	 * out if so.  The page fault code is a relatively likely
290	 * place to get trapped in an infinite regress, and once we
291	 * overwrite the whole stack, it becomes very hard to recover.
292	 */
293	stack_offset = stack_pointer & (THREAD_SIZE-1);
294	if (stack_offset < THREAD_SIZE / 8) {
295		pr_alert("Potential stack overrun: sp %#lx\n", stack_pointer);
296		show_regs(regs);
297		pr_alert("Killing current process %d/%s\n",
298			 tsk->pid, tsk->comm);
299		do_group_exit(SIGKILL);
300	}
301
302	/*
303	 * Early on, we need to check for migrating PTE entries;
304	 * see homecache.c.  If we find a migrating PTE, we wait until
305	 * the backing page claims to be done migrating, then we proceed.
306	 * For kernel PTEs, we rewrite the PTE and return and retry.
307	 * Otherwise, we treat the fault like a normal "no PTE" fault,
308	 * rather than trying to patch up the existing PTE.
309	 */
310	pgd = get_current_pgd();
311	if (handle_migrating_pte(pgd, fault_num, address, regs->pc,
312				 is_kernel_mode, write))
313		return 1;
314
315	si_code = SEGV_MAPERR;
316
317	/*
318	 * We fault-in kernel-space virtual memory on-demand. The
319	 * 'reference' page table is init_mm.pgd.
320	 *
321	 * NOTE! We MUST NOT take any locks for this case. We may
322	 * be in an interrupt or a critical region, and should
323	 * only copy the information from the master page table,
324	 * nothing more.
325	 *
326	 * This verifies that the fault happens in kernel space
327	 * and that the fault was not a protection fault.
328	 */
329	if (unlikely(address >= TASK_SIZE &&
330		     !is_arch_mappable_range(address, 0))) {
331		if (is_kernel_mode && is_page_fault &&
332		    vmalloc_fault(pgd, address) >= 0)
333			return 1;
334		/*
335		 * Don't take the mm semaphore here. If we fixup a prefetch
336		 * fault we could otherwise deadlock.
337		 */
338		mm = NULL;  /* happy compiler */
339		vma = NULL;
340		goto bad_area_nosemaphore;
341	}
342
343	/*
344	 * If we're trying to touch user-space addresses, we must
345	 * be either at PL0, or else with interrupts enabled in the
346	 * kernel, so either way we can re-enable interrupts here
347	 * unless we are doing atomic access to user space with
348	 * interrupts disabled.
349	 */
350	if (!(regs->flags & PT_FLAGS_DISABLE_IRQ))
351		local_irq_enable();
352
353	mm = tsk->mm;
354
355	/*
356	 * If we're in an interrupt, have no user context or are running in an
357	 * region with pagefaults disabled then we must not take the fault.
358	 */
359	if (pagefault_disabled() || !mm) {
360		vma = NULL;  /* happy compiler */
361		goto bad_area_nosemaphore;
362	}
363
364	if (!is_kernel_mode)
365		flags |= FAULT_FLAG_USER;
366
367	/*
368	 * When running in the kernel we expect faults to occur only to
369	 * addresses in user space.  All other faults represent errors in the
370	 * kernel and should generate an OOPS.  Unfortunately, in the case of an
371	 * erroneous fault occurring in a code path which already holds mmap_sem
372	 * we will deadlock attempting to validate the fault against the
373	 * address space.  Luckily the kernel only validly references user
374	 * space from well defined areas of code, which are listed in the
375	 * exceptions table.
376	 *
377	 * As the vast majority of faults will be valid we will only perform
378	 * the source reference check when there is a possibility of a deadlock.
379	 * Attempt to lock the address space, if we cannot we then validate the
380	 * source.  If this is invalid we can skip the address space check,
381	 * thus avoiding the deadlock.
382	 */
383	if (!down_read_trylock(&mm->mmap_sem)) {
384		if (is_kernel_mode &&
385		    !search_exception_tables(regs->pc)) {
386			vma = NULL;  /* happy compiler */
387			goto bad_area_nosemaphore;
388		}
389
390retry:
391		down_read(&mm->mmap_sem);
392	}
393
394	vma = find_vma(mm, address);
395	if (!vma)
396		goto bad_area;
397	if (vma->vm_start <= address)
398		goto good_area;
399	if (!(vma->vm_flags & VM_GROWSDOWN))
400		goto bad_area;
401	if (regs->sp < PAGE_OFFSET) {
402		/*
403		 * accessing the stack below sp is always a bug.
404		 */
405		if (address < regs->sp)
406			goto bad_area;
407	}
408	if (expand_stack(vma, address))
409		goto bad_area;
410
411/*
412 * Ok, we have a good vm_area for this memory access, so
413 * we can handle it..
414 */
415good_area:
416	si_code = SEGV_ACCERR;
417	if (fault_num == INT_ITLB_MISS) {
418		if (!(vma->vm_flags & VM_EXEC))
419			goto bad_area;
420	} else if (write) {
421#ifdef TEST_VERIFY_AREA
422		if (!is_page_fault && regs->cs == KERNEL_CS)
423			pr_err("WP fault at " REGFMT "\n", regs->eip);
424#endif
425		if (!(vma->vm_flags & VM_WRITE))
426			goto bad_area;
427		flags |= FAULT_FLAG_WRITE;
428	} else {
429		if (!is_page_fault || !(vma->vm_flags & VM_READ))
430			goto bad_area;
431	}
432
433	/*
434	 * If for any reason at all we couldn't handle the fault,
435	 * make sure we exit gracefully rather than endlessly redo
436	 * the fault.
437	 */
438	fault = handle_mm_fault(mm, vma, address, flags);
439
440	if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current))
441		return 0;
442
443	if (unlikely(fault & VM_FAULT_ERROR)) {
444		if (fault & VM_FAULT_OOM)
445			goto out_of_memory;
446		else if (fault & VM_FAULT_SIGSEGV)
447			goto bad_area;
448		else if (fault & VM_FAULT_SIGBUS)
449			goto do_sigbus;
450		BUG();
451	}
452	if (flags & FAULT_FLAG_ALLOW_RETRY) {
453		if (fault & VM_FAULT_MAJOR)
454			tsk->maj_flt++;
455		else
456			tsk->min_flt++;
457		if (fault & VM_FAULT_RETRY) {
458			flags &= ~FAULT_FLAG_ALLOW_RETRY;
459			flags |= FAULT_FLAG_TRIED;
460
461			 /*
462			  * No need to up_read(&mm->mmap_sem) as we would
463			  * have already released it in __lock_page_or_retry
464			  * in mm/filemap.c.
465			  */
466			goto retry;
467		}
468	}
469
470#if CHIP_HAS_TILE_DMA()
471	/* If this was a DMA TLB fault, restart the DMA engine. */
472	switch (fault_num) {
473	case INT_DMATLB_MISS:
474	case INT_DMATLB_MISS_DWNCL:
475	case INT_DMATLB_ACCESS:
476	case INT_DMATLB_ACCESS_DWNCL:
477		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
478		break;
479	}
480#endif
481
482	up_read(&mm->mmap_sem);
483	return 1;
484
485/*
486 * Something tried to access memory that isn't in our memory map..
487 * Fix it, but check if it's kernel or user first..
488 */
489bad_area:
490	up_read(&mm->mmap_sem);
491
492bad_area_nosemaphore:
493	/* User mode accesses just cause a SIGSEGV */
494	if (!is_kernel_mode) {
495		/*
496		 * It's possible to have interrupts off here.
497		 */
498		local_irq_enable();
499
500		force_sig_info_fault("segfault", SIGSEGV, si_code, address,
501				     fault_num, tsk, regs);
502		return 0;
503	}
504
505no_context:
506	/* Are we prepared to handle this kernel fault?  */
507	if (fixup_exception(regs))
508		return 0;
509
510/*
511 * Oops. The kernel tried to access some bad page. We'll have to
512 * terminate things with extreme prejudice.
513 */
514
515	bust_spinlocks(1);
516
517	/* FIXME: no lookup_address() yet */
518#ifdef SUPPORT_LOOKUP_ADDRESS
519	if (fault_num == INT_ITLB_MISS) {
520		pte_t *pte = lookup_address(address);
521
522		if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
523			pr_crit("kernel tried to execute non-executable page - exploit attempt? (uid: %d)\n",
524				current->uid);
525	}
526#endif
527	if (address < PAGE_SIZE)
528		pr_alert("Unable to handle kernel NULL pointer dereference\n");
529	else
530		pr_alert("Unable to handle kernel paging request\n");
531	pr_alert(" at virtual address " REGFMT ", pc " REGFMT "\n",
532		 address, regs->pc);
533
534	show_regs(regs);
535
536	if (unlikely(tsk->pid < 2)) {
537		panic("Kernel page fault running %s!",
538		      is_idle_task(tsk) ? "the idle task" : "init");
539	}
540
541	/*
542	 * More FIXME: we should probably copy the i386 here and
543	 * implement a generic die() routine.  Not today.
544	 */
545#ifdef SUPPORT_DIE
546	die("Oops", regs);
547#endif
548	bust_spinlocks(1);
549
550	do_group_exit(SIGKILL);
551
552/*
553 * We ran out of memory, or some other thing happened to us that made
554 * us unable to handle the page fault gracefully.
555 */
556out_of_memory:
557	up_read(&mm->mmap_sem);
558	if (is_kernel_mode)
559		goto no_context;
560	pagefault_out_of_memory();
561	return 0;
562
563do_sigbus:
564	up_read(&mm->mmap_sem);
565
566	/* Kernel mode? Handle exceptions or die */
567	if (is_kernel_mode)
568		goto no_context;
569
570	force_sig_info_fault("bus error", SIGBUS, BUS_ADRERR, address,
571			     fault_num, tsk, regs);
572	return 0;
573}
574
575#ifndef __tilegx__
576
577/* We must release ICS before panicking or we won't get anywhere. */
578#define ics_panic(fmt, ...)					\
579do {								\
580	__insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);	\
581	panic(fmt, ##__VA_ARGS__);				\
582} while (0)
583
584/*
585 * When we take an ITLB or DTLB fault or access violation in the
586 * supervisor while the critical section bit is set, the hypervisor is
587 * reluctant to write new values into the EX_CONTEXT_K_x registers,
588 * since that might indicate we have not yet squirreled the SPR
589 * contents away and can thus safely take a recursive interrupt.
590 * Accordingly, the hypervisor passes us the PC via SYSTEM_SAVE_K_2.
591 *
592 * Note that this routine is called before homecache_tlb_defer_enter(),
593 * which means that we can properly unlock any atomics that might
594 * be used there (good), but also means we must be very sensitive
595 * to not touch any data structures that might be located in memory
596 * that could migrate, as we could be entering the kernel on a dataplane
597 * cpu that has been deferring kernel TLB updates.  This means, for
598 * example, that we can't migrate init_mm or its pgd.
599 */
600struct intvec_state do_page_fault_ics(struct pt_regs *regs, int fault_num,
601				      unsigned long address,
602				      unsigned long info)
603{
604	unsigned long pc = info & ~1;
605	int write = info & 1;
606	pgd_t *pgd = get_current_pgd();
607
608	/* Retval is 1 at first since we will handle the fault fully. */
609	struct intvec_state state = {
610		do_page_fault, fault_num, address, write, 1
611	};
612
613	/* Validate that we are plausibly in the right routine. */
614	if ((pc & 0x7) != 0 || pc < PAGE_OFFSET ||
615	    (fault_num != INT_DTLB_MISS &&
616	     fault_num != INT_DTLB_ACCESS)) {
617		unsigned long old_pc = regs->pc;
618		regs->pc = pc;
619		ics_panic("Bad ICS page fault args: old PC %#lx, fault %d/%d at %#lx",
620			  old_pc, fault_num, write, address);
621	}
622
623	/* We might be faulting on a vmalloc page, so check that first. */
624	if (fault_num != INT_DTLB_ACCESS && vmalloc_fault(pgd, address) >= 0)
625		return state;
626
627	/*
628	 * If we faulted with ICS set in sys_cmpxchg, we are providing
629	 * a user syscall service that should generate a signal on
630	 * fault.  We didn't set up a kernel stack on initial entry to
631	 * sys_cmpxchg, but instead had one set up by the fault, which
632	 * (because sys_cmpxchg never releases ICS) came to us via the
633	 * SYSTEM_SAVE_K_2 mechanism, and thus EX_CONTEXT_K_[01] are
634	 * still referencing the original user code.  We release the
635	 * atomic lock and rewrite pt_regs so that it appears that we
636	 * came from user-space directly, and after we finish the
637	 * fault we'll go back to user space and re-issue the swint.
638	 * This way the backtrace information is correct if we need to
639	 * emit a stack dump at any point while handling this.
640	 *
641	 * Must match register use in sys_cmpxchg().
642	 */
643	if (pc >= (unsigned long) sys_cmpxchg &&
644	    pc < (unsigned long) __sys_cmpxchg_end) {
645#ifdef CONFIG_SMP
646		/* Don't unlock before we could have locked. */
647		if (pc >= (unsigned long)__sys_cmpxchg_grab_lock) {
648			int *lock_ptr = (int *)(regs->regs[ATOMIC_LOCK_REG]);
649			__atomic_fault_unlock(lock_ptr);
650		}
651#endif
652		regs->sp = regs->regs[27];
653	}
654
655	/*
656	 * We can also fault in the atomic assembly, in which
657	 * case we use the exception table to do the first-level fixup.
658	 * We may re-fixup again in the real fault handler if it
659	 * turns out the faulting address is just bad, and not,
660	 * for example, migrating.
661	 */
662	else if (pc >= (unsigned long) __start_atomic_asm_code &&
663		   pc < (unsigned long) __end_atomic_asm_code) {
664		const struct exception_table_entry *fixup;
665#ifdef CONFIG_SMP
666		/* Unlock the atomic lock. */
667		int *lock_ptr = (int *)(regs->regs[ATOMIC_LOCK_REG]);
668		__atomic_fault_unlock(lock_ptr);
669#endif
670		fixup = search_exception_tables(pc);
671		if (!fixup)
672			ics_panic("ICS atomic fault not in table: PC %#lx, fault %d",
673				  pc, fault_num);
674		regs->pc = fixup->fixup;
675		regs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
676	}
677
678	/*
679	 * Now that we have released the atomic lock (if necessary),
680	 * it's safe to spin if the PTE that caused the fault was migrating.
681	 */
682	if (fault_num == INT_DTLB_ACCESS)
683		write = 1;
684	if (handle_migrating_pte(pgd, fault_num, address, pc, 1, write))
685		return state;
686
687	/* Return zero so that we continue on with normal fault handling. */
688	state.retval = 0;
689	return state;
690}
691
692#endif /* !__tilegx__ */
693
694/*
695 * This routine handles page faults.  It determines the address, and the
696 * problem, and then passes it handle_page_fault() for normal DTLB and
697 * ITLB issues, and for DMA or SN processor faults when we are in user
698 * space.  For the latter, if we're in kernel mode, we just save the
699 * interrupt away appropriately and return immediately.  We can't do
700 * page faults for user code while in kernel mode.
701 */
702static inline void __do_page_fault(struct pt_regs *regs, int fault_num,
703				   unsigned long address, unsigned long write)
704{
705	int is_page_fault;
706
707#ifdef CONFIG_KPROBES
708	/*
709	 * This is to notify the fault handler of the kprobes.  The
710	 * exception code is redundant as it is also carried in REGS,
711	 * but we pass it anyhow.
712	 */
713	if (notify_die(DIE_PAGE_FAULT, "page fault", regs, -1,
714		       regs->faultnum, SIGSEGV) == NOTIFY_STOP)
715		return;
716#endif
717
718#ifdef __tilegx__
719	/*
720	 * We don't need early do_page_fault_ics() support, since unlike
721	 * Pro we don't need to worry about unlocking the atomic locks.
722	 * There is only one current case in GX where we touch any memory
723	 * under ICS other than our own kernel stack, and we handle that
724	 * here.  (If we crash due to trying to touch our own stack,
725	 * we're in too much trouble for C code to help out anyway.)
726	 */
727	if (write & ~1) {
728		unsigned long pc = write & ~1;
729		if (pc >= (unsigned long) __start_unalign_asm_code &&
730		    pc < (unsigned long) __end_unalign_asm_code) {
731			struct thread_info *ti = current_thread_info();
732			/*
733			 * Our EX_CONTEXT is still what it was from the
734			 * initial unalign exception, but now we've faulted
735			 * on the JIT page.  We would like to complete the
736			 * page fault however is appropriate, and then retry
737			 * the instruction that caused the unalign exception.
738			 * Our state has been "corrupted" by setting the low
739			 * bit in "sp", and stashing r0..r3 in the
740			 * thread_info area, so we revert all of that, then
741			 * continue as if this were a normal page fault.
742			 */
743			regs->sp &= ~1UL;
744			regs->regs[0] = ti->unalign_jit_tmp[0];
745			regs->regs[1] = ti->unalign_jit_tmp[1];
746			regs->regs[2] = ti->unalign_jit_tmp[2];
747			regs->regs[3] = ti->unalign_jit_tmp[3];
748			write &= 1;
749		} else {
750			pr_alert("%s/%d: ICS set at page fault at %#lx: %#lx\n",
751				 current->comm, current->pid, pc, address);
752			show_regs(regs);
753			do_group_exit(SIGKILL);
754		}
755	}
756#else
757	/* This case should have been handled by do_page_fault_ics(). */
758	BUG_ON(write & ~1);
759#endif
760
761#if CHIP_HAS_TILE_DMA()
762	/*
763	 * If it's a DMA fault, suspend the transfer while we're
764	 * handling the miss; we'll restart after it's handled.  If we
765	 * don't suspend, it's possible that this process could swap
766	 * out and back in, and restart the engine since the DMA is
767	 * still 'running'.
768	 */
769	if (fault_num == INT_DMATLB_MISS ||
770	    fault_num == INT_DMATLB_ACCESS ||
771	    fault_num == INT_DMATLB_MISS_DWNCL ||
772	    fault_num == INT_DMATLB_ACCESS_DWNCL) {
773		__insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
774		while (__insn_mfspr(SPR_DMA_USER_STATUS) &
775		       SPR_DMA_STATUS__BUSY_MASK)
776			;
777	}
778#endif
779
780	/* Validate fault num and decide if this is a first-time page fault. */
781	switch (fault_num) {
782	case INT_ITLB_MISS:
783	case INT_DTLB_MISS:
784#if CHIP_HAS_TILE_DMA()
785	case INT_DMATLB_MISS:
786	case INT_DMATLB_MISS_DWNCL:
787#endif
788		is_page_fault = 1;
789		break;
790
791	case INT_DTLB_ACCESS:
792#if CHIP_HAS_TILE_DMA()
793	case INT_DMATLB_ACCESS:
794	case INT_DMATLB_ACCESS_DWNCL:
795#endif
796		is_page_fault = 0;
797		break;
798
799	default:
800		panic("Bad fault number %d in do_page_fault", fault_num);
801	}
802
803#if CHIP_HAS_TILE_DMA()
804	if (!user_mode(regs)) {
805		struct async_tlb *async;
806		switch (fault_num) {
807#if CHIP_HAS_TILE_DMA()
808		case INT_DMATLB_MISS:
809		case INT_DMATLB_ACCESS:
810		case INT_DMATLB_MISS_DWNCL:
811		case INT_DMATLB_ACCESS_DWNCL:
812			async = &current->thread.dma_async_tlb;
813			break;
814#endif
815		default:
816			async = NULL;
817		}
818		if (async) {
819
820			/*
821			 * No vmalloc check required, so we can allow
822			 * interrupts immediately at this point.
823			 */
824			local_irq_enable();
825
826			set_thread_flag(TIF_ASYNC_TLB);
827			if (async->fault_num != 0) {
828				panic("Second async fault %d; old fault was %d (%#lx/%ld)",
829				      fault_num, async->fault_num,
830				      address, write);
831			}
832			BUG_ON(fault_num == 0);
833			async->fault_num = fault_num;
834			async->is_fault = is_page_fault;
835			async->is_write = write;
836			async->address = address;
837			return;
838		}
839	}
840#endif
841
842	handle_page_fault(regs, fault_num, is_page_fault, address, write);
843}
844
845void do_page_fault(struct pt_regs *regs, int fault_num,
846		   unsigned long address, unsigned long write)
847{
848	enum ctx_state prev_state = exception_enter();
849	__do_page_fault(regs, fault_num, address, write);
850	exception_exit(prev_state);
851}
852
853#if CHIP_HAS_TILE_DMA()
854/*
855 * This routine effectively re-issues asynchronous page faults
856 * when we are returning to user space.
857 */
858void do_async_page_fault(struct pt_regs *regs)
859{
860	struct async_tlb *async = &current->thread.dma_async_tlb;
861
862	/*
863	 * Clear thread flag early.  If we re-interrupt while processing
864	 * code here, we will reset it and recall this routine before
865	 * returning to user space.
866	 */
867	clear_thread_flag(TIF_ASYNC_TLB);
868
869	if (async->fault_num) {
870		/*
871		 * Clear async->fault_num before calling the page-fault
872		 * handler so that if we re-interrupt before returning
873		 * from the function we have somewhere to put the
874		 * information from the new interrupt.
875		 */
876		int fault_num = async->fault_num;
877		async->fault_num = 0;
878		handle_page_fault(regs, fault_num, async->is_fault,
879				  async->address, async->is_write);
880	}
881}
882#endif /* CHIP_HAS_TILE_DMA() */
883
884
885void vmalloc_sync_all(void)
886{
887#ifdef __tilegx__
888	/* Currently all L1 kernel pmd's are static and shared. */
889	BUILD_BUG_ON(pgd_index(VMALLOC_END - PAGE_SIZE) !=
890		     pgd_index(VMALLOC_START));
891#else
892	/*
893	 * Note that races in the updates of insync and start aren't
894	 * problematic: insync can only get set bits added, and updates to
895	 * start are only improving performance (without affecting correctness
896	 * if undone).
897	 */
898	static DECLARE_BITMAP(insync, PTRS_PER_PGD);
899	static unsigned long start = PAGE_OFFSET;
900	unsigned long address;
901
902	BUILD_BUG_ON(PAGE_OFFSET & ~PGDIR_MASK);
903	for (address = start; address >= PAGE_OFFSET; address += PGDIR_SIZE) {
904		if (!test_bit(pgd_index(address), insync)) {
905			unsigned long flags;
906			struct list_head *pos;
907
908			spin_lock_irqsave(&pgd_lock, flags);
909			list_for_each(pos, &pgd_list)
910				if (!vmalloc_sync_one(list_to_pgd(pos),
911								address)) {
912					/* Must be at first entry in list. */
913					BUG_ON(pos != pgd_list.next);
914					break;
915				}
916			spin_unlock_irqrestore(&pgd_lock, flags);
917			if (pos != pgd_list.next)
918				set_bit(pgd_index(address), insync);
919		}
920		if (address == start && test_bit(pgd_index(address), insync))
921			start = address + PGDIR_SIZE;
922	}
923#endif
924}
925