arch/loongarch/mm/fault.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
   4  *
   5  * Derived from MIPS:
   6  * Copyright (C) 1995 - 2000 by Ralf Baechle
   7  */
   8 #include <linux/context_tracking.h>
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/interrupt.h>
  12 #include <linux/kernel.h>
  13 #include <linux/entry-common.h>
  14 #include <linux/errno.h>
  15 #include <linux/string.h>
  16 #include <linux/types.h>
  17 #include <linux/ptrace.h>
  18 #include <linux/ratelimit.h>
  19 #include <linux/mman.h>
  20 #include <linux/mm.h>
  21 #include <linux/smp.h>
  22 #include <linux/kdebug.h>
  23 #include <linux/perf_event.h>
  24 #include <linux/uaccess.h>
  25 #include <linux/kfence.h>
  26
  27 #include <asm/branch.h>
  28 #include <asm/exception.h>
  29 #include <asm/mmu_context.h>
  30 #include <asm/ptrace.h>
  31
  32 int show_unhandled_signals = 1;
  33
  34 static int __kprobes spurious_fault(unsigned long write, unsigned long address)
  35 {
  36         pgd_t *pgd;
  37         p4d_t *p4d;
  38         pud_t *pud;
  39         pmd_t *pmd;
  40         pte_t *pte;
  41
  42         if (!(address & __UA_LIMIT))
  43                 return 0;
  44
  45         pgd = pgd_offset_k(address);
  46         if (!pgd_present(pgdp_get(pgd)))
  47                 return 0;
  48
  49         p4d = p4d_offset(pgd, address);
  50         if (!p4d_present(p4dp_get(p4d)))
  51                 return 0;
  52
  53         pud = pud_offset(p4d, address);
  54         if (!pud_present(pudp_get(pud)))
  55                 return 0;
  56
  57         pmd = pmd_offset(pud, address);
  58         if (!pmd_present(pmdp_get(pmd)))
  59                 return 0;
  60
  61         if (pmd_leaf(*pmd)) {
  62                 return write ? pmd_write(pmdp_get(pmd)) : 1;
  63         } else {
  64                 pte = pte_offset_kernel(pmd, address);
  65                 if (!pte_present(ptep_get(pte)))
  66                         return 0;
  67
  68                 return write ? pte_write(ptep_get(pte)) : 1;
  69         }
  70 }
  71
  72 static void __kprobes no_context(struct pt_regs *regs,
  73                         unsigned long write, unsigned long address)
  74 {
  75         const int field = sizeof(unsigned long) * 2;
  76
  77         if (spurious_fault(write, address))
  78                 return;
  79
  80         /* Are we prepared to handle this kernel fault?  */
  81         if (fixup_exception(regs))
  82                 return;
  83
  84         if (kfence_handle_page_fault(address, write, regs))
  85                 return;
  86
  87         /*
  88          * Oops. The kernel tried to access some bad page. We'll have to
  89          * terminate things with extreme prejudice.
  90          */
  91         bust_spinlocks(1);
  92
  93         pr_alert("CPU %d Unable to handle kernel paging request at "
  94                "virtual address %0*lx, era == %0*lx, ra == %0*lx\n",
  95                raw_smp_processor_id(), field, address, field, regs->csr_era,
  96                field,  regs->regs[1]);
  97         die("Oops", regs);
  98 }
  99
 100 static void __kprobes do_out_of_memory(struct pt_regs *regs,
 101                         unsigned long write, unsigned long address)
 102 {
 103         /*
 104          * We ran out of memory, call the OOM killer, and return the userspace
 105          * (which will retry the fault, or kill us if we got oom-killed).
 106          */
 107         if (!user_mode(regs)) {
 108                 no_context(regs, write, address);
 109                 return;
 110         }
 111         pagefault_out_of_memory();
 112 }
 113
 114 static void __kprobes do_sigbus(struct pt_regs *regs,
 115                 unsigned long write, unsigned long address, int si_code)
 116 {
 117         /* Kernel mode? Handle exceptions or die */
 118         if (!user_mode(regs)) {
 119                 no_context(regs, write, address);
 120                 return;
 121         }
 122
 123         /*
 124          * Send a sigbus, regardless of whether we were in kernel
 125          * or user mode.
 126          */
 127         current->thread.csr_badvaddr = address;
 128         current->thread.trap_nr = read_csr_excode();
 129         force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
 130 }
 131
 132 static void __kprobes do_sigsegv(struct pt_regs *regs,
 133                 unsigned long write, unsigned long address, int si_code)
 134 {
 135         const int field = sizeof(unsigned long) * 2;
 136         static DEFINE_RATELIMIT_STATE(ratelimit_state, 5 * HZ, 10);
 137
 138         /* Kernel mode? Handle exceptions or die */
 139         if (!user_mode(regs)) {
 140                 no_context(regs, write, address);
 141                 return;
 142         }
 143
 144         /* User mode accesses just cause a SIGSEGV */
 145         current->thread.csr_badvaddr = address;
 146         if (!write)
 147                 current->thread.error_code = 1;
 148         else
 149                 current->thread.error_code = 2;
 150         current->thread.trap_nr = read_csr_excode();
 151
 152         if (show_unhandled_signals &&
 153             unhandled_signal(current, SIGSEGV) && __ratelimit(&ratelimit_state)) {
 154                 pr_info("do_page_fault(): sending SIGSEGV to %s for invalid %s %0*lx\n",
 155                         current->comm,
 156                         write ? "write access to" : "read access from",
 157                         field, address);
 158                 pr_info("era = %0*lx in", field,
 159                         (unsigned long) regs->csr_era);
 160                 print_vma_addr(KERN_CONT " ", regs->csr_era);
 161                 pr_cont("\n");
 162                 pr_info("ra  = %0*lx in", field,
 163                         (unsigned long) regs->regs[1]);
 164                 print_vma_addr(KERN_CONT " ", regs->regs[1]);
 165                 pr_cont("\n");
 166         }
 167         force_sig_fault(SIGSEGV, si_code, (void __user *)address);
 168 }
 169
 170 /*
 171  * This routine handles page faults.  It determines the address,
 172  * and the problem, and then passes it off to one of the appropriate
 173  * routines.
 174  */
 175 static void __kprobes __do_page_fault(struct pt_regs *regs,
 176                         unsigned long write, unsigned long address)
 177 {
 178         int si_code = SEGV_MAPERR;
 179         unsigned int flags = FAULT_FLAG_DEFAULT;
 180         struct task_struct *tsk = current;
 181         struct mm_struct *mm = tsk->mm;
 182         struct vm_area_struct *vma = NULL;
 183         vm_fault_t fault;
 184
 185         if (kprobe_page_fault(regs, current->thread.trap_nr))
 186                 return;
 187
 188         /*
 189          * We fault-in kernel-space virtual memory on-demand. The
 190          * 'reference' page table is init_mm.pgd.
 191          *
 192          * NOTE! We MUST NOT take any locks for this case. We may
 193          * be in an interrupt or a critical region, and should
 194          * only copy the information from the master page table,
 195          * nothing more.
 196          */
 197         if (address & __UA_LIMIT) {
 198                 if (!user_mode(regs))
 199                         no_context(regs, write, address);
 200                 else
 201                         do_sigsegv(regs, write, address, si_code);
 202                 return;
 203         }
 204
 205         /*
 206          * If we're in an interrupt or have no user
 207          * context, we must not take the fault..
 208          */
 209         if (faulthandler_disabled() || !mm) {
 210                 do_sigsegv(regs, write, address, si_code);
 211                 return;
 212         }
 213
 214         if (user_mode(regs))
 215                 flags |= FAULT_FLAG_USER;
 216
 217         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 218 retry:
 219         vma = lock_mm_and_find_vma(mm, address, regs);
 220         if (unlikely(!vma))
 221                 goto bad_area_nosemaphore;
 222         goto good_area;
 223
 224 /*
 225  * Something tried to access memory that isn't in our memory map..
 226  * Fix it, but check if it's kernel or user first..
 227  */
 228 bad_area:
 229         mmap_read_unlock(mm);
 230 bad_area_nosemaphore:
 231         do_sigsegv(regs, write, address, si_code);
 232         return;
 233
 234 /*
 235  * Ok, we have a good vm_area for this memory access, so
 236  * we can handle it..
 237  */
 238 good_area:
 239         si_code = SEGV_ACCERR;
 240
 241         if (write) {
 242                 flags |= FAULT_FLAG_WRITE;
 243                 if (!(vma->vm_flags & VM_WRITE))
 244                         goto bad_area;
 245         } else {
 246                 if (!(vma->vm_flags & VM_EXEC) && address == exception_era(regs))
 247                         goto bad_area;
 248                 if (!(vma->vm_flags & (VM_READ | VM_WRITE)) && address != exception_era(regs))
 249                         goto bad_area;
 250         }
 251
 252         /*
 253          * If for any reason at all we couldn't handle the fault,
 254          * make sure we exit gracefully rather than endlessly redo
 255          * the fault.
 256          */
 257         fault = handle_mm_fault(vma, address, flags, regs);
 258
 259         if (fault_signal_pending(fault, regs)) {
 260                 if (!user_mode(regs))
 261                         no_context(regs, write, address);
 262                 return;
 263         }
 264
 265         /* The fault is fully completed (including releasing mmap lock) */
 266         if (fault & VM_FAULT_COMPLETED)
 267                 return;
 268
 269         if (unlikely(fault & VM_FAULT_RETRY)) {
 270                 flags |= FAULT_FLAG_TRIED;
 271
 272                 /*
 273                  * No need to mmap_read_unlock(mm) as we would
 274                  * have already released it in __lock_page_or_retry
 275                  * in mm/filemap.c.
 276                  */
 277                 goto retry;
 278         }
 279         if (unlikely(fault & VM_FAULT_ERROR)) {
 280                 mmap_read_unlock(mm);
 281                 if (fault & VM_FAULT_OOM) {
 282                         do_out_of_memory(regs, write, address);
 283                         return;
 284                 } else if (fault & VM_FAULT_SIGSEGV) {
 285                         do_sigsegv(regs, write, address, si_code);
 286                         return;
 287                 } else if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
 288                         do_sigbus(regs, write, address, si_code);
 289                         return;
 290                 }
 291                 BUG();
 292         }
 293
 294         mmap_read_unlock(mm);
 295 }
 296
 297 asmlinkage void __kprobes do_page_fault(struct pt_regs *regs,
 298                         unsigned long write, unsigned long address)
 299 {
 300         irqentry_state_t state = irqentry_enter(regs);
 301
 302         /* Enable interrupt if enabled in parent context */
 303         if (likely(regs->csr_prmd & CSR_PRMD_PIE))
 304                 local_irq_enable();
 305
 306         __do_page_fault(regs, write, address);
 307
 308         local_irq_disable();
 309
 310         irqentry_exit(regs, state);
 311 }