]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/commitdiff
arm64: ftrace: don't validate branch via PLT in ftrace_make_nop()
authorArd Biesheuvel <ard.biesheuvel@linaro.org>
Tue, 6 Jun 2017 17:00:21 +0000 (17:00 +0000)
committerWill Deacon <will.deacon@arm.com>
Wed, 7 Jun 2017 10:50:34 +0000 (11:50 +0100)
When turning branch instructions into NOPs, we attempt to validate the
action by comparing the old value at the call site with the opcode of
a direct relative branch instruction pointing at the old target.

However, these call sites are statically initialized to call _mcount(),
and may be redirected via a PLT entry if the module is loaded far away
from the kernel text, leading to false negatives and spurious errors.

So skip the validation if CONFIG_ARM64_MODULE_PLTS is configured.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
arch/arm64/kernel/ftrace.c

index 40ad08ac569af76b1f58fa775101ef264c480039..4cb576374b829b8f39dc3c05453221186125af0f 100644 (file)
@@ -84,12 +84,52 @@ int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
                    unsigned long addr)
 {
        unsigned long pc = rec->ip;
-       u32 old, new;
+       long offset = (long)pc - (long)addr;
+       bool validate = true;
+       u32 old = 0, new;
+
+       if (IS_ENABLED(CONFIG_ARM64_MODULE_PLTS) &&
+           (offset < -SZ_128M || offset >= SZ_128M)) {
+               u32 replaced;
+
+               /*
+                * 'mod' is only set at module load time, but if we end up
+                * dealing with an out-of-range condition, we can assume it
+                * is due to a module being loaded far away from the kernel.
+                */
+               if (!mod) {
+                       preempt_disable();
+                       mod = __module_text_address(pc);
+                       preempt_enable();
+
+                       if (WARN_ON(!mod))
+                               return -EINVAL;
+               }
+
+               /*
+                * The instruction we are about to patch may be a branch and
+                * link instruction that was redirected via a PLT entry. In
+                * this case, the normal validation will fail, but we can at
+                * least check that we are dealing with a branch and link
+                * instruction that points into the right module.
+                */
+               if (aarch64_insn_read((void *)pc, &replaced))
+                       return -EFAULT;
+
+               if (!aarch64_insn_is_bl(replaced) ||
+                   !within_module(pc + aarch64_get_branch_offset(replaced),
+                                  mod))
+                       return -EINVAL;
+
+               validate = false;
+       } else {
+               old = aarch64_insn_gen_branch_imm(pc, addr,
+                                                 AARCH64_INSN_BRANCH_LINK);
+       }
 
-       old = aarch64_insn_gen_branch_imm(pc, addr, AARCH64_INSN_BRANCH_LINK);
        new = aarch64_insn_gen_nop();
 
-       return ftrace_modify_code(pc, old, new, true);
+       return ftrace_modify_code(pc, old, new, validate);
 }
 
 void arch_ftrace_update_code(int command)