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>