sh: ftrace: Make code modification NMI safe.
This cribs the x86 implementation of ftrace_nmi_enter() and friends to
make ftrace_modify_code() NMI safe, particularly on SMP configurations.
For additional notes on the problems involved, see the comment below
ftrace_call_replace().
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
diff --git a/arch/sh/Kconfig b/arch/sh/Kconfig
index b940424..5260fb55 100644
--- a/arch/sh/Kconfig
+++ b/arch/sh/Kconfig
@@ -38,6 +38,7 @@
select HAVE_DYNAMIC_FTRACE
select HAVE_FUNCTION_TRACE_MCOUNT_TEST
select HAVE_FTRACE_SYSCALLS
+ select HAVE_FTRACE_NMI_ENTER if DYNAMIC_FTRACE
select HAVE_FUNCTION_GRAPH_TRACER
select HAVE_ARCH_KGDB
select ARCH_HIBERNATION_POSSIBLE if MMU
diff --git a/arch/sh/kernel/ftrace.c b/arch/sh/kernel/ftrace.c
index 2c48e26..b6f41c1 100644
--- a/arch/sh/kernel/ftrace.c
+++ b/arch/sh/kernel/ftrace.c
@@ -62,6 +62,150 @@
return ftrace_replaced_code;
}
+/*
+ * Modifying code must take extra care. On an SMP machine, if
+ * the code being modified is also being executed on another CPU
+ * that CPU will have undefined results and possibly take a GPF.
+ * We use kstop_machine to stop other CPUS from exectuing code.
+ * But this does not stop NMIs from happening. We still need
+ * to protect against that. We separate out the modification of
+ * the code to take care of this.
+ *
+ * Two buffers are added: An IP buffer and a "code" buffer.
+ *
+ * 1) Put the instruction pointer into the IP buffer
+ * and the new code into the "code" buffer.
+ * 2) Wait for any running NMIs to finish and set a flag that says
+ * we are modifying code, it is done in an atomic operation.
+ * 3) Write the code
+ * 4) clear the flag.
+ * 5) Wait for any running NMIs to finish.
+ *
+ * If an NMI is executed, the first thing it does is to call
+ * "ftrace_nmi_enter". This will check if the flag is set to write
+ * and if it is, it will write what is in the IP and "code" buffers.
+ *
+ * The trick is, it does not matter if everyone is writing the same
+ * content to the code location. Also, if a CPU is executing code
+ * it is OK to write to that code location if the contents being written
+ * are the same as what exists.
+ */
+#define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */
+static atomic_t nmi_running = ATOMIC_INIT(0);
+static int mod_code_status; /* holds return value of text write */
+static void *mod_code_ip; /* holds the IP to write to */
+static void *mod_code_newcode; /* holds the text to write to the IP */
+
+static unsigned nmi_wait_count;
+static atomic_t nmi_update_count = ATOMIC_INIT(0);
+
+int ftrace_arch_read_dyn_info(char *buf, int size)
+{
+ int r;
+
+ r = snprintf(buf, size, "%u %u",
+ nmi_wait_count,
+ atomic_read(&nmi_update_count));
+ return r;
+}
+
+static void clear_mod_flag(void)
+{
+ int old = atomic_read(&nmi_running);
+
+ for (;;) {
+ int new = old & ~MOD_CODE_WRITE_FLAG;
+
+ if (old == new)
+ break;
+
+ old = atomic_cmpxchg(&nmi_running, old, new);
+ }
+}
+
+static void ftrace_mod_code(void)
+{
+ /*
+ * Yes, more than one CPU process can be writing to mod_code_status.
+ * (and the code itself)
+ * But if one were to fail, then they all should, and if one were
+ * to succeed, then they all should.
+ */
+ mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode,
+ MCOUNT_INSN_SIZE);
+
+ /* if we fail, then kill any new writers */
+ if (mod_code_status)
+ clear_mod_flag();
+}
+
+void ftrace_nmi_enter(void)
+{
+ if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) {
+ smp_rmb();
+ ftrace_mod_code();
+ atomic_inc(&nmi_update_count);
+ }
+ /* Must have previous changes seen before executions */
+ smp_mb();
+}
+
+void ftrace_nmi_exit(void)
+{
+ /* Finish all executions before clearing nmi_running */
+ smp_mb();
+ atomic_dec(&nmi_running);
+}
+
+static void wait_for_nmi_and_set_mod_flag(void)
+{
+ if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG))
+ return;
+
+ do {
+ cpu_relax();
+ } while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG));
+
+ nmi_wait_count++;
+}
+
+static void wait_for_nmi(void)
+{
+ if (!atomic_read(&nmi_running))
+ return;
+
+ do {
+ cpu_relax();
+ } while (atomic_read(&nmi_running));
+
+ nmi_wait_count++;
+}
+
+static int
+do_ftrace_mod_code(unsigned long ip, void *new_code)
+{
+ mod_code_ip = (void *)ip;
+ mod_code_newcode = new_code;
+
+ /* The buffers need to be visible before we let NMIs write them */
+ smp_mb();
+
+ wait_for_nmi_and_set_mod_flag();
+
+ /* Make sure all running NMIs have finished before we write the code */
+ smp_mb();
+
+ ftrace_mod_code();
+
+ /* Make sure the write happens before clearing the bit */
+ smp_mb();
+
+ clear_mod_flag();
+ wait_for_nmi();
+
+ return mod_code_status;
+}
+
static int ftrace_modify_code(unsigned long ip, unsigned char *old_code,
unsigned char *new_code)
{
@@ -86,7 +230,7 @@
return -EINVAL;
/* replace the text with the new text */
- if (probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE))
+ if (do_ftrace_mod_code(ip, new_code))
return -EPERM;
flush_icache_range(ip, ip + MCOUNT_INSN_SIZE);