Home Home > GIT Browse
diff options
authorSteven Rostedt <rostedt@goodmis.org>2008-10-15 18:21:44 -0400
committerGreg Kroah-Hartman <gregkh@suse.de>2008-10-15 16:02:33 -0700
commitd23d43386311fde5f11e06c16d4185e94a8d6d06 (patch)
parent3fa8749e584b55f1180411ab1b51117190bac1e5 (diff)
disable CONFIG_DYNAMIC_FTRACE due to possible memory corruption on module unload
While debugging the e1000e corruption bug with Intel, we discovered today that the dynamic ftrace code in mainline is the likely source of this bug. For the stable kernel we are providing the only viable fix patch: labeling CONFIG_DYNAMIC_FTRACE as broken. (see the patch below) We will follow up with a backport patch that contains the fixes. But since the fixes are not a one liner, the safest approach for now is to disable the code in question. The cause of the bug is due to the way the current code in mainline handles dynamic ftrace. When dynamic ftrace is turned on, it also turns on CONFIG_FTRACE which enables the -pg config in gcc that places a call to mcount at every function call. With just CONFIG_FTRACE this causes a noticeable overhead. CONFIG_DYNAMIC_FTRACE works to ease this overhead by dynamically updating the mcount call sites into nops. The problem arises when we trace functions and modules are unloaded. The first time a function is called, it will call mcount and the mcount call will call ftrace_record_ip. This records the calling site and stores it in a preallocated hash table. Later on a daemon will wake up and call kstop_machine and convert any mcount callers into nops. The evolution of this code first tried to do this without the kstop_machine and used cmpxchg to update the callers as they were called. But I was informed that this is dangerous to do on SMP machines if another CPU is running that same code. The solution was to do this with kstop_machine. We still used cmpxchg to test if the code that we are modifying is indeed code that we expect to be before updating it - as a final line of defense. But on 32bit machines, ioremapped memory and modules share the same address space. When a module would load its code into memory and execute some code, that would register the function. On module unload, ftrace incorrectly did not zap these functions from its hash (this was the bug). The cmpxchg could have saved us in most cases (via luck) - but with ioremap-ed memory that was exactly the wrong thing to do - the results of cmpxchg on device memory are undefined. (and will likely result in a write) The pending .28 ftrace tree does not have this bug anymore, as a general push towards more robustness of code patching, this is done differently: we do not use cmpxchg and we do a WARN_ON and turn the tracer off if anything deviates from its expected state. Furthermore, patch sites are statically identified during build time so there's no runtime discovery of dynamic code areas anymore, and no room for code unmaps to cause the hash to become out of date. We believe the fragility of dynamic patching has been sufficiently addressed in the development code via the static patching method, but further suggestions to make it more robust are welcome. Signed-off-by: Steven Rostedt <srostedt@goodmis.org> Acked-by: Ingo Molnar <mingo@elte.hu> Acked-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
1 files changed, 2 insertions, 1 deletions
diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig
index 263e9e6bbd60..aa53fdd040e1 100644
--- a/kernel/trace/Kconfig
+++ b/kernel/trace/Kconfig
@@ -103,7 +103,8 @@ config CONTEXT_SWITCH_TRACER
all switching of tasks.
- bool "enable/disable ftrace tracepoints dynamically"
+ bool "enable/disable ftrace tracepoints dynamically (BROKEN)"
+ depends on BROKEN
depends on FTRACE
default y