ftrace: avoid modifying kprobe'd records
Abhishek Sagar [Sat, 21 Jun 2008 18:20:29 +0000 (23:20 +0530)]
Avoid modifying the mcount call-site if there is a kprobe installed on it.
These records are not marked as failed however. This allowed the filter
rules on them to remain up-to-date. Whenever the kprobe on the corresponding
record is removed, the record gets updated as normal.

Signed-off-by: Abhishek Sagar <sagar.abhishek@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

kernel/trace/ftrace.c

index d123816..85e8413 100644 (file)
@@ -21,6 +21,7 @@
 #include <linux/hardirq.h>
 #include <linux/kthread.h>
 #include <linux/uaccess.h>
+#include <linux/kprobes.h>
 #include <linux/ftrace.h>
 #include <linux/sysctl.h>
 #include <linux/ctype.h>
@@ -500,6 +501,10 @@ static void ftrace_replace_code(int enable)
                        if (rec->flags & FTRACE_FL_FAILED)
                                continue;
 
+                       /* ignore updates to this record's mcount site */
+                       if (get_kprobe((void *)rec->ip))
+                               continue;
+
                        failed = __ftrace_replace_code(rec, old, new, enable);
                        if (failed && (rec->flags & FTRACE_FL_CONVERTED)) {
                                rec->flags |= FTRACE_FL_FAILED;
@@ -692,11 +697,11 @@ unsigned long             ftrace_update_tot_cnt;
 
 static int __ftrace_update_code(void *ignore)
 {
+       int i, save_ftrace_enabled;
+       cycle_t start, stop;
        struct dyn_ftrace *p;
        struct hlist_node *t, *n;
-       int save_ftrace_enabled;
-       cycle_t start, stop;
-       int i;
+       struct hlist_head *head, temp_list;
 
        /* Don't be recording funcs now */
        ftrace_record_suspend++;
@@ -708,8 +713,11 @@ static int __ftrace_update_code(void *ignore)
 
        /* No locks needed, the machine is stopped! */
        for (i = 0; i < FTRACE_HASHSIZE; i++) {
+               INIT_HLIST_HEAD(&temp_list);
+               head = &ftrace_hash[i];
+
                /* all CPUS are stopped, we are safe to modify code */
-               hlist_for_each_entry_safe(p, t, n, &ftrace_hash[i], node) {
+               hlist_for_each_entry_safe(p, t, n, head, node) {
                        /* Skip over failed records which have not been
                         * freed. */
                        if (p->flags & FTRACE_FL_FAILED)
@@ -723,6 +731,19 @@ static int __ftrace_update_code(void *ignore)
                        if (p->flags & (FTRACE_FL_CONVERTED))
                                break;
 
+                       /* Ignore updates to this record's mcount site.
+                        * Reintroduce this record at the head of this
+                        * bucket to attempt to "convert" it again if
+                        * the kprobe on it is unregistered before the
+                        * next run. */
+                       if (get_kprobe((void *)p->ip)) {
+                               ftrace_del_hash(p);
+                               INIT_HLIST_NODE(&p->node);
+                               hlist_add_head(&p->node, &temp_list);
+                               continue;
+                       }
+
+                       /* convert record (i.e, patch mcount-call with NOP) */
                        if (ftrace_code_disable(p)) {
                                p->flags |= FTRACE_FL_CONVERTED;
                                ftrace_update_cnt++;
@@ -734,6 +755,12 @@ static int __ftrace_update_code(void *ignore)
                                }
                        }
                }
+
+               hlist_for_each_entry_safe(p, t, n, &temp_list, node) {
+                       hlist_del(&p->node);
+                       INIT_HLIST_NODE(&p->node);
+                       hlist_add_head(&p->node, head);
+               }
        }
 
        stop = ftrace_now(raw_smp_processor_id());