The resent conversion of the cpu hotplug support in the uncore driver
introduced a regression due to the way the callbacks are invoked at
initialization time.
The old code called the prepare/starting/online function on each online cpu
as a block. The new code registers the hotplug callbacks in the core for
each state. The core invokes the callbacks at each registration on all
online cpus.
The code implicitely relied on the prepare/starting/online callbacks being
called as combo on a particular cpu, which was not obvious and completely
undocumented.
The resulting subtle wreckage happens due to the way how the uncore code
manages shared data structures for cpus which share an uncore resource in
hardware. The sharing is determined in the cpu starting callback, but the
prepare callback allocates per cpu data for the upcoming cpu because
potential sharing is unknown at this point. If the starting callback finds
a online cpu which shares the hardware resource it takes a refcount on the
percpu data of that cpu and puts the own data structure into a
'free_at_online' pointer of that shared data structure. The online callback
frees that.
With the old model this worked because in a starting callback only one non
unused structure (the one of the starting cpu) was available. The new code
allocates the data structures for all cpus when the prepare callback is
registered.
Now the starting function iterates through all online cpus and looks for a
data structure (skipping its own) which has a matching hardware id. The id
member of the data structure is initialized to 0, but the hardware id can
be 0 as well. The resulting wreckage is:
CPU0 finds a matching id on CPU1, takes a refcount on CPU1 data and puts
its own data structure into CPU1s data structure to be freed.
CPU1 skips CPU0 because the data structure is its allegedly unsued own.
It finds a matching id on CPU2, takes a refcount on CPU1 data and puts
its own data structure into CPU2s data structure to be freed.
....
Now the online callbacks are invoked.
CPU0 has a pointer to CPU1s data and frees the original CPU0 data. So
far so good.
CPU1 has a pointer to CPU2s data and frees the original CPU1 data, which
is still referenced by CPU0 ---> Booom
So there are two issues to be solved here:
1) The id field must be initialized at allocation time to a value which
cannot be a valid hardware id, i.e. -1
This prevents the above scenario, but now CPU1 and CPU2 both stick their
own data structure into the free_at_online pointer of CPU0. So we leak
CPU1s data structure.
2) Fix the memory leak described in #1
Instead of having a single pointer, use a hlist to enqueue the
superflous data structures which are then freed by the first cpu
invoking the online callback.
Ideally we should know the sharing _before_ invoking the prepare callback,
but that's way beyond the scope of this bug fix.
[ tglx: Rewrote changelog ]
Fixes: 96b2bd3866 ("perf/x86/amd/uncore: Convert to hotplug state machine")
Reported-and-tested-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Borislav Petkov <bp@suse.de>
Link: http://lkml.kernel.org/r/20160909160822.lowgmkdwms2dheyv@linutronix.de
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
7d762e49c2