On Mon, Nov 26, 2018 at 12:26:17AM +0000, kys@linuxonhyperv.com wrote:
From: Dexuan Cui decui@microsoft.com
We can concurrently try to open the same sub-channel from 2 paths:
path #1: vmbus_onoffer() -> vmbus_process_offer() -> handle_sc_creation(). path #2: storvsc_probe() -> storvsc_connect_to_vsp() -> -> storvsc_channel_init() -> handle_multichannel_storage() -> -> vmbus_are_subchannels_present() -> handle_sc_creation().
They conflict with each other, but it was not an issue before the recent commit ae6935ed7d42 ("vmbus: split ring buffer allocation from open"), because at the beginning of vmbus_open() we checked newchannel->state so only one path could succeed, and the other would return with -EINVAL.
After ae6935ed7d42, the failing path frees the channel's ringbuffer by vmbus_free_ring(), and this causes a panic later.
Commit ae6935ed7d42 itself is good, and it just reveals the longstanding race. We can resolve the issue by removing path #2, i.e. removing the second vmbus_are_subchannels_present() in handle_multichannel_storage().
BTW, the comment "Check to see if sub-channels have already been created" in handle_multichannel_storage() is incorrect: when we unload the driver, we first close the sub-channel(s) and then close the primary channel, next the host sends rescind-offer message(s) so primary->sc_list will become empty. This means the first vmbus_are_subchannels_present() in handle_multichannel_storage() is never useful.
Fixes: ae6935ed7d42 ("vmbus: split ring buffer allocation from open") Cc: stable@vger.kernel.org Cc: Long Li longli@microsoft.com Cc: Stephen Hemminger sthemmin@microsoft.com Cc: K. Y. Srinivasan kys@microsoft.com Cc: Haiyang Zhang haiyangz@microsoft.com Signed-off-by: Dexuan Cui decui@microsoft.com Signed-off-by: K. Y. Srinivasan kys@microsoft.com
Just a heads-up: ae6935ed7d42 ("vmbus: split ring buffer allocation from open") was merged in the 4.20 merge window, so this fix won't actually apply to any of the current stable trees.
However, it's good to have tags (fixes + cc: stable) here since this fix might end up (for whatever reason) getting merged only for 4.21, which will then make these tags relevant.
-- Thanks, Sasha