arm64: introduce interfaces to hotpatch kernel and module code
authorJiang Liu <liuj97@gmail.com>
Tue, 7 Jan 2014 14:17:09 +0000 (22:17 +0800)
committerCatalin Marinas <catalin.marinas@arm.com>
Wed, 8 Jan 2014 15:21:29 +0000 (15:21 +0000)
Introduce three interfaces to patch kernel and module code:
aarch64_insn_patch_text_nosync():
patch code without synchronization, it's caller's responsibility
to synchronize all CPUs if needed.
aarch64_insn_patch_text_sync():
patch code and always synchronize with stop_machine()
aarch64_insn_patch_text():
patch code and synchronize with stop_machine() if needed

Reviewed-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Jiang Liu <liuj97@gmail.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
arch/arm64/include/asm/insn.h
arch/arm64/kernel/insn.c

index 1bdc44c274562d2a01678aa6fda415151fc57ec0..bf8085fdc140f1be04369f326459e4d07e8d8aae 100644 (file)
@@ -18,6 +18,9 @@
 #define        __ASM_INSN_H
 #include <linux/types.h>
 
+/* A64 instructions are always 32 bits. */
+#define        AARCH64_INSN_SIZE               4
+
 /*
  * ARM Architecture Reference Manual for ARMv8 Profile-A, Issue A.a
  * Section C3.1 "A64 instruction index by encoding":
@@ -70,8 +73,13 @@ __AARCH64_INSN_FUNCS(hint,   0xFFFFF01F, 0xD503201F)
 
 bool aarch64_insn_is_nop(u32 insn);
 
+int aarch64_insn_read(void *addr, u32 *insnp);
+int aarch64_insn_write(void *addr, u32 insn);
 enum aarch64_insn_encoding_class aarch64_get_insn_class(u32 insn);
-
 bool aarch64_insn_hotpatch_safe(u32 old_insn, u32 new_insn);
 
+int aarch64_insn_patch_text_nosync(void *addr, u32 insn);
+int aarch64_insn_patch_text_sync(void *addrs[], u32 insns[], int cnt);
+int aarch64_insn_patch_text(void *addrs[], u32 insns[], int cnt);
+
 #endif /* __ASM_INSN_H */
index 56a2498ab3e3215b1d556d0720d5f69e4a0ade66..b9dac57e580a3fcad2a8e2a0aa5f32da86fcc898 100644 (file)
  */
 #include <linux/compiler.h>
 #include <linux/kernel.h>
+#include <linux/smp.h>
+#include <linux/stop_machine.h>
+#include <linux/uaccess.h>
+#include <asm/cacheflush.h>
 #include <asm/insn.h>
 
 static int aarch64_insn_encoding_class[] = {
@@ -60,6 +64,28 @@ bool __kprobes aarch64_insn_is_nop(u32 insn)
        }
 }
 
+/*
+ * In ARMv8-A, A64 instructions have a fixed length of 32 bits and are always
+ * little-endian.
+ */
+int __kprobes aarch64_insn_read(void *addr, u32 *insnp)
+{
+       int ret;
+       u32 val;
+
+       ret = probe_kernel_read(&val, addr, AARCH64_INSN_SIZE);
+       if (!ret)
+               *insnp = le32_to_cpu(val);
+
+       return ret;
+}
+
+int __kprobes aarch64_insn_write(void *addr, u32 insn)
+{
+       insn = cpu_to_le32(insn);
+       return probe_kernel_write(addr, &insn, AARCH64_INSN_SIZE);
+}
+
 static bool __kprobes __aarch64_insn_hotpatch_safe(u32 insn)
 {
        if (aarch64_get_insn_class(insn) != AARCH64_INSN_CLS_BR_SYS)
@@ -89,3 +115,96 @@ bool __kprobes aarch64_insn_hotpatch_safe(u32 old_insn, u32 new_insn)
        return __aarch64_insn_hotpatch_safe(old_insn) &&
               __aarch64_insn_hotpatch_safe(new_insn);
 }
+
+int __kprobes aarch64_insn_patch_text_nosync(void *addr, u32 insn)
+{
+       u32 *tp = addr;
+       int ret;
+
+       /* A64 instructions must be word aligned */
+       if ((uintptr_t)tp & 0x3)
+               return -EINVAL;
+
+       ret = aarch64_insn_write(tp, insn);
+       if (ret == 0)
+               flush_icache_range((uintptr_t)tp,
+                                  (uintptr_t)tp + AARCH64_INSN_SIZE);
+
+       return ret;
+}
+
+struct aarch64_insn_patch {
+       void            **text_addrs;
+       u32             *new_insns;
+       int             insn_cnt;
+       atomic_t        cpu_count;
+};
+
+static int __kprobes aarch64_insn_patch_text_cb(void *arg)
+{
+       int i, ret = 0;
+       struct aarch64_insn_patch *pp = arg;
+
+       /* The first CPU becomes master */
+       if (atomic_inc_return(&pp->cpu_count) == 1) {
+               for (i = 0; ret == 0 && i < pp->insn_cnt; i++)
+                       ret = aarch64_insn_patch_text_nosync(pp->text_addrs[i],
+                                                            pp->new_insns[i]);
+               /*
+                * aarch64_insn_patch_text_nosync() calls flush_icache_range(),
+                * which ends with "dsb; isb" pair guaranteeing global
+                * visibility.
+                */
+               atomic_set(&pp->cpu_count, -1);
+       } else {
+               while (atomic_read(&pp->cpu_count) != -1)
+                       cpu_relax();
+               isb();
+       }
+
+       return ret;
+}
+
+int __kprobes aarch64_insn_patch_text_sync(void *addrs[], u32 insns[], int cnt)
+{
+       struct aarch64_insn_patch patch = {
+               .text_addrs = addrs,
+               .new_insns = insns,
+               .insn_cnt = cnt,
+               .cpu_count = ATOMIC_INIT(0),
+       };
+
+       if (cnt <= 0)
+               return -EINVAL;
+
+       return stop_machine(aarch64_insn_patch_text_cb, &patch,
+                           cpu_online_mask);
+}
+
+int __kprobes aarch64_insn_patch_text(void *addrs[], u32 insns[], int cnt)
+{
+       int ret;
+       u32 insn;
+
+       /* Unsafe to patch multiple instructions without synchronizaiton */
+       if (cnt == 1) {
+               ret = aarch64_insn_read(addrs[0], &insn);
+               if (ret)
+                       return ret;
+
+               if (aarch64_insn_hotpatch_safe(insn, insns[0])) {
+                       /*
+                        * ARMv8 architecture doesn't guarantee all CPUs see
+                        * the new instruction after returning from function
+                        * aarch64_insn_patch_text_nosync(). So send IPIs to
+                        * all other CPUs to achieve instruction
+                        * synchronization.
+                        */
+                       ret = aarch64_insn_patch_text_nosync(addrs[0], insns[0]);
+                       kick_all_cpus_sync();
+                       return ret;
+               }
+       }
+
+       return aarch64_insn_patch_text_sync(addrs, insns, cnt);
+}