fault.c

  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 * Copyright (C) 2009 Sunplus Core Technology Co., Ltd.
  4 *  Lennox Wu <lennox.wu@sunplusct.com>
  5 *  Chen Liqin <liqin.chen@sunplusct.com>
  6 * Copyright (C) 2012 Regents of the University of California
  7 */
  8
  9
 10#include <linux/mm.h>
 11#include <linux/kernel.h>
 12#include <linux/interrupt.h>
 13#include <linux/perf_event.h>
 14#include <linux/signal.h>
 15#include <linux/uaccess.h>
 16#include <linux/kprobes.h>
 17#include <linux/kfence.h>
 18#include <linux/entry-common.h>
 19
 20#include <asm/ptrace.h>
 21#include <asm/tlbflush.h>
 22
 23#define CREATE_TRACE_POINTS
 24#include <trace/events/exceptions.h>
 25
 26#include "../kernel/head.h"
 27
 28static void show_pte(unsigned long addr)
 29{
 30	pgd_t *pgdp, pgd;
 31	p4d_t *p4dp, p4d;
 32	pud_t *pudp, pud;
 33	pmd_t *pmdp, pmd;
 34	pte_t *ptep, pte;
 35	struct mm_struct *mm = current->mm;
 36
 37	if (!mm)
 38		mm = &init_mm;
 39
 40	pr_alert("Current %s pgtable: %luK pagesize, %d-bit VAs, pgdp=0x%016llx\n",
 41		 current->comm, PAGE_SIZE / SZ_1K, VA_BITS,
 42		 mm == &init_mm ? (u64)__pa_symbol(mm->pgd) : virt_to_phys(mm->pgd));
 43
 44	pgdp = pgd_offset(mm, addr);
 45	pgd = pgdp_get(pgdp);
 46	pr_alert("[%016lx] pgd=%016lx", addr, pgd_val(pgd));
 47	if (pgd_none(pgd) || pgd_bad(pgd) || pgd_leaf(pgd))
 48		goto out;
 49
 50	p4dp = p4d_offset(pgdp, addr);
 51	p4d = p4dp_get(p4dp);
 52	pr_cont(", p4d=%016lx", p4d_val(p4d));
 53	if (p4d_none(p4d) || p4d_bad(p4d) || p4d_leaf(p4d))
 54		goto out;
 55
 56	pudp = pud_offset(p4dp, addr);
 57	pud = pudp_get(pudp);
 58	pr_cont(", pud=%016lx", pud_val(pud));
 59	if (pud_none(pud) || pud_bad(pud) || pud_leaf(pud))
 60		goto out;
 61
 62	pmdp = pmd_offset(pudp, addr);
 63	pmd = pmdp_get(pmdp);
 64	pr_cont(", pmd=%016lx", pmd_val(pmd));
 65	if (pmd_none(pmd) || pmd_bad(pmd) || pmd_leaf(pmd))
 66		goto out;
 67
 68	ptep = pte_offset_map(pmdp, addr);
 69	if (!ptep)
 70		goto out;
 71
 72	pte = ptep_get(ptep);
 73	pr_cont(", pte=%016lx", pte_val(pte));
 74	pte_unmap(ptep);
 75out:
 76	pr_cont("\n");
 77}
 78
 79static void die_kernel_fault(const char *msg, unsigned long addr,
 80		struct pt_regs *regs)
 81{
 82	bust_spinlocks(1);
 83
 84	pr_alert("Unable to handle kernel %s at virtual address " REG_FMT "\n", msg,
 85		addr);
 86
 87	bust_spinlocks(0);
 88	show_pte(addr);
 89	die(regs, "Oops");
 90	make_task_dead(SIGKILL);
 91}
 92
 93static inline void no_context(struct pt_regs *regs, unsigned long addr)
 94{
 95	const char *msg;
 96
 97	/* Are we prepared to handle this kernel fault? */
 98	if (fixup_exception(regs))
 99		return;
100
101	/*
102	 * Oops. The kernel tried to access some bad page. We'll have to
103	 * terminate things with extreme prejudice.
104	 */
105	if (addr < PAGE_SIZE)
106		msg = "NULL pointer dereference";
107	else {
108		if (kfence_handle_page_fault(addr, regs->cause == EXC_STORE_PAGE_FAULT, regs))
109			return;
110
111		msg = "paging request";
112	}
113
114	die_kernel_fault(msg, addr, regs);
115}
116
117static inline void mm_fault_error(struct pt_regs *regs, unsigned long addr, vm_fault_t fault)
118{
119	if (!user_mode(regs)) {
120		no_context(regs, addr);
121		return;
122	}
123
124	if (fault & VM_FAULT_OOM) {
125		/*
126		 * We ran out of memory, call the OOM killer, and return the userspace
127		 * (which will retry the fault, or kill us if we got oom-killed).
128		 */
129		pagefault_out_of_memory();
130		return;
131	} else if (fault & (VM_FAULT_SIGBUS | VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE)) {
132		/* Kernel mode? Handle exceptions or die */
133		do_trap(regs, SIGBUS, BUS_ADRERR, addr);
134		return;
135	} else if (fault & VM_FAULT_SIGSEGV) {
136		do_trap(regs, SIGSEGV, SEGV_MAPERR, addr);
137		return;
138	}
139
140	BUG();
141}
142
143static inline void
144bad_area_nosemaphore(struct pt_regs *regs, int code, unsigned long addr)
145{
146	/*
147	 * Something tried to access memory that isn't in our memory map.
148	 * Fix it, but check if it's kernel or user first.
149	 */
150	/* User mode accesses just cause a SIGSEGV */
151	if (user_mode(regs)) {
152		do_trap(regs, SIGSEGV, code, addr);
153		return;
154	}
155
156	no_context(regs, addr);
157}
158
159static inline void
160bad_area(struct pt_regs *regs, struct mm_struct *mm, int code,
161	 unsigned long addr)
162{
163	mmap_read_unlock(mm);
164
165	bad_area_nosemaphore(regs, code, addr);
166}
167
168static inline void vmalloc_fault(struct pt_regs *regs, int code, unsigned long addr)
169{
170	pgd_t *pgd, *pgd_k;
171	pud_t *pud_k;
172	p4d_t *p4d_k;
173	pmd_t *pmd_k;
174	pte_t *pte_k;
175	int index;
176	unsigned long pfn;
177
178	/* User mode accesses just cause a SIGSEGV */
179	if (user_mode(regs))
180		return do_trap(regs, SIGSEGV, code, addr);
181
182	/*
183	 * Synchronize this task's top level page-table
184	 * with the 'reference' page table.
185	 *
186	 * Do _not_ use "tsk->active_mm->pgd" here.
187	 * We might be inside an interrupt in the middle
188	 * of a task switch.
189	 */
190	index = pgd_index(addr);
191	pfn = csr_read(CSR_SATP) & SATP_PPN;
192	pgd = (pgd_t *)pfn_to_virt(pfn) + index;
193	pgd_k = init_mm.pgd + index;
194
195	if (!pgd_present(pgdp_get(pgd_k))) {
196		no_context(regs, addr);
197		return;
198	}
199	set_pgd(pgd, pgdp_get(pgd_k));
200
201	p4d_k = p4d_offset(pgd_k, addr);
202	if (!p4d_present(p4dp_get(p4d_k))) {
203		no_context(regs, addr);
204		return;
205	}
206
207	pud_k = pud_offset(p4d_k, addr);
208	if (!pud_present(pudp_get(pud_k))) {
209		no_context(regs, addr);
210		return;
211	}
212	if (pud_leaf(pudp_get(pud_k)))
213		goto flush_tlb;
214
215	/*
216	 * Since the vmalloc area is global, it is unnecessary
217	 * to copy individual PTEs
218	 */
219	pmd_k = pmd_offset(pud_k, addr);
220	if (!pmd_present(pmdp_get(pmd_k))) {
221		no_context(regs, addr);
222		return;
223	}
224	if (pmd_leaf(pmdp_get(pmd_k)))
225		goto flush_tlb;
226
227	/*
228	 * Make sure the actual PTE exists as well to
229	 * catch kernel vmalloc-area accesses to non-mapped
230	 * addresses. If we don't do this, this will just
231	 * silently loop forever.
232	 */
233	pte_k = pte_offset_kernel(pmd_k, addr);
234	if (!pte_present(ptep_get(pte_k))) {
235		no_context(regs, addr);
236		return;
237	}
238
239	/*
240	 * The kernel assumes that TLBs don't cache invalid
241	 * entries, but in RISC-V, SFENCE.VMA specifies an
242	 * ordering constraint, not a cache flush; it is
243	 * necessary even after writing invalid entries.
244	 */
245flush_tlb:
246	local_flush_tlb_page(addr);
247}
248
249static inline bool access_error(unsigned long cause, struct vm_area_struct *vma)
250{
251	switch (cause) {
252	case EXC_INST_PAGE_FAULT:
253		if (!(vma->vm_flags & VM_EXEC)) {
254			return true;
255		}
256		break;
257	case EXC_LOAD_PAGE_FAULT:
258		/* Write implies read */
259		if (!(vma->vm_flags & (VM_READ | VM_WRITE))) {
260			return true;
261		}
262		break;
263	case EXC_STORE_PAGE_FAULT:
264		if (!(vma->vm_flags & VM_WRITE)) {
265			return true;
266		}
267		break;
268	default:
269		panic("%s: unhandled cause %lu", __func__, cause);
270	}
271	return false;
272}
273
274/*
275 * This routine handles page faults.  It determines the address and the
276 * problem, and then passes it off to one of the appropriate routines.
277 */
278void handle_page_fault(struct pt_regs *regs)
279{
280	struct task_struct *tsk;
281	struct vm_area_struct *vma;
282	struct mm_struct *mm;
283	unsigned long addr, cause;
284	unsigned int flags = FAULT_FLAG_DEFAULT;
285	int code = SEGV_MAPERR;
286	vm_fault_t fault;
287
288	cause = regs->cause;
289	addr = regs->badaddr;
290
291	tsk = current;
292	mm = tsk->mm;
293
294	if (kprobe_page_fault(regs, cause))
295		return;
296
297	if (user_mode(regs))
298		trace_page_fault_user(addr, regs, cause);
299	else
300		trace_page_fault_kernel(addr, regs, cause);
301
302	/*
303	 * Fault-in kernel-space virtual memory on-demand.
304	 * The 'reference' page table is init_mm.pgd.
305	 *
306	 * NOTE! We MUST NOT take any locks for this case. We may
307	 * be in an interrupt or a critical region, and should
308	 * only copy the information from the master page table,
309	 * nothing more.
310	 */
311	if ((!IS_ENABLED(CONFIG_MMU) || !IS_ENABLED(CONFIG_64BIT)) &&
312	    unlikely(addr >= VMALLOC_START && addr < VMALLOC_END)) {
313		vmalloc_fault(regs, code, addr);
314		return;
315	}
316
317	/* Enable interrupts if they were enabled in the parent context. */
318	if (!regs_irqs_disabled(regs))
319		local_irq_enable();
320
321	/*
322	 * If we're in an interrupt, have no user context, or are running
323	 * in an atomic region, then we must not take the fault.
324	 */
325	if (unlikely(faulthandler_disabled() || !mm)) {
326		tsk->thread.bad_cause = cause;
327		no_context(regs, addr);
328		return;
329	}
330
331	if (user_mode(regs))
332		flags |= FAULT_FLAG_USER;
333
334	if (!user_mode(regs) && addr < TASK_SIZE && unlikely(!(regs->status & SR_SUM))) {
335		if (fixup_exception(regs))
336			return;
337
338		die_kernel_fault("access to user memory without uaccess routines", addr, regs);
339	}
340
341	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, addr);
342
343	if (cause == EXC_STORE_PAGE_FAULT)
344		flags |= FAULT_FLAG_WRITE;
345	else if (cause == EXC_INST_PAGE_FAULT)
346		flags |= FAULT_FLAG_INSTRUCTION;
347	if (!(flags & FAULT_FLAG_USER))
348		goto lock_mmap;
349
350	vma = lock_vma_under_rcu(mm, addr);
351	if (!vma)
352		goto lock_mmap;
353
354	if (unlikely(access_error(cause, vma))) {
355		vma_end_read(vma);
356		count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
357		tsk->thread.bad_cause = cause;
358		bad_area_nosemaphore(regs, SEGV_ACCERR, addr);
359		return;
360	}
361
362	fault = handle_mm_fault(vma, addr, flags | FAULT_FLAG_VMA_LOCK, regs);
363	if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
364		vma_end_read(vma);
365
366	if (!(fault & VM_FAULT_RETRY)) {
367		count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
368		goto done;
369	}
370	count_vm_vma_lock_event(VMA_LOCK_RETRY);
371	if (fault & VM_FAULT_MAJOR)
372		flags |= FAULT_FLAG_TRIED;
373
374	if (fault_signal_pending(fault, regs)) {
375		if (!user_mode(regs))
376			no_context(regs, addr);
377		return;
378	}
379lock_mmap:
380
381retry:
382	vma = lock_mm_and_find_vma(mm, addr, regs);
383	if (unlikely(!vma)) {
384		tsk->thread.bad_cause = cause;
385		bad_area_nosemaphore(regs, code, addr);
386		return;
387	}
388
389	/*
390	 * Ok, we have a good vm_area for this memory access, so
391	 * we can handle it.
392	 */
393	code = SEGV_ACCERR;
394
395	if (unlikely(access_error(cause, vma))) {
396		tsk->thread.bad_cause = cause;
397		bad_area(regs, mm, code, addr);
398		return;
399	}
400
401	/*
402	 * If for any reason at all we could not handle the fault,
403	 * make sure we exit gracefully rather than endlessly redo
404	 * the fault.
405	 */
406	fault = handle_mm_fault(vma, addr, flags, regs);
407
408	/*
409	 * If we need to retry but a fatal signal is pending, handle the
410	 * signal first. We do not need to release the mmap_lock because it
411	 * would already be released in __lock_page_or_retry in mm/filemap.c.
412	 */
413	if (fault_signal_pending(fault, regs)) {
414		if (!user_mode(regs))
415			no_context(regs, addr);
416		return;
417	}
418
419	/* The fault is fully completed (including releasing mmap lock) */
420	if (fault & VM_FAULT_COMPLETED)
421		return;
422
423	if (unlikely(fault & VM_FAULT_RETRY)) {
424		flags |= FAULT_FLAG_TRIED;
425
426		/*
427		 * No need to mmap_read_unlock(mm) as we would
428		 * have already released it in __lock_page_or_retry
429		 * in mm/filemap.c.
430		 */
431		goto retry;
432	}
433
434	mmap_read_unlock(mm);
435
436done:
437	if (unlikely(fault & VM_FAULT_ERROR)) {
438		tsk->thread.bad_cause = cause;
439		mm_fault_error(regs, addr, fault);
440		return;
441	}
442	return;
443}