When setting up a channel fails, vmbus_add_channel_work() and vmbus_device_register() do not agree on who cleans up:
- vmbus_add_channel_kobj() drops the channel's last reference when it fails, and its callers then use the channel and drop it again (patch 1; this is what our static analysis reported, for the sub-channel path). Paths that fail before it put a kobject that was never initialized. - When vmbus_device_register() fails, the device's release calls hv_process_channel_removal() on a channel that was not rescinded, which hits BUG_ON(!channel->rescind); vmbus_add_channel_work() would then clean the channel up a second time (patch 2). Patch 2 leaves the channel to vmbus_add_channel_work() when device_register() fails, and keeps the device when only the channel's sysfs entries fail, since a driver may be bound by then. Changes in v3 (patch 2 only, after the Sashiko review of v2): - vmbus_match() reads the channel under rcu_read_lock(), as it is freed with kfree_rcu(). - The channel no longer points to the device after device_register() fails and the device is freed. - A channel of a device without a channels directory is left out of sysfs instead of being refused. v2 refused a sub-channel offered while the primary channel's device was still being registered, which could leave a driver waiting for it. v2: https://lore.kernel.org/linux-hyperv/[email protected]/ Changes in v2 (patch 2 only, after the Sashiko review of v1): - vmbus_match() skips a device whose channel pointer was cleared, which a concurrent driver_attach() can still see. - Failures after device_register() (the channels directory, the primary channel's kobject and its sysfs group) no longer unregister the device: a bound driver such as hv_sock, or one that opened sub-channels, can keep the channel past its remove callback unless the channel was rescinded. v1: https://lore.kernel.org/linux-hyperv/[email protected]/ Testing: QEMU (x86_64, KVM, kvm=off with the hv-* enlightenments, vmbus-bridge and hv-balloon) with KASAN, hv_vmbus and hv_balloon as modules loaded from init after failslab is armed, on mainline d135ab453a74. failslab is limited to allocations under vmbus_device_register(), and fails the N-th one when the hv-balloon offer is processed. Without the series, every failure hits the BUG_ON. With it, for N = 1 to 95: - 53 failures in device_register() leave the offer unbound with "unable to add child device object"; - 31 failures in the channels directory, the channel's kobject or its sysfs group leave the device registered and bound to hv_balloon, with the channel missing from sysfs; unbinding the driver and binding it again works; - the other runs fail an allocation whose failure is handled, or none. No run reports anything (no BUG, KASAN, WARNING or refcount_t message). v2 and v3 give the same results. The sub-channel paths are not covered: QEMU's VMBus offers no sub-channels, cannot rescind a channel, and hv_balloon cannot be unloaded. The failures need an allocation failure while the guest processes a channel offer from the host. Yuqiang Sun (2): Drivers: hv: vmbus: Keep the channel when vmbus_add_channel_kobj() fails Drivers: hv: vmbus: Do not remove the channel when device registration fails drivers/hv/channel_mgmt.c | 12 +++-- drivers/hv/hyperv_vmbus.h | 1 + drivers/hv/vmbus_drv.c | 97 +++++++++++++++++++++++++-------------- 3 files changed, 71 insertions(+), 39 deletions(-) base-commit: d135ab453a7400a34930bdda1ef4bdc8acfac80e -- 2.53.0
