e1000e: locking bug introduced by commit 67fd4fcb
Commit 67fd4fcb (e1000e: convert to stats64) added the ability to update statistics more accurately and on-demand through the net_device_ops .ndo_get_stats64 hook, but introduced a locking bug on 82577/8/9 when linked at half-duplex (seen on kernels with CONFIG_DEBUG_ATOMIC_SLEEP=y and CONFIG_PROVE_LOCKING=y). The commit introduced code paths that caused a mutex to be locked in atomic contexts, e.g. an rcu_read_lock is held when irqbalance reads the stats from /sys/class/net/ethX/statistics causing the mutex to be locked to read the Phy half-duplex statistics registers. The mutex was originally introduced to prevent concurrent accesses of resources (the NVM and Phy) shared by the driver, firmware and hardware a few years back when there was an issue with the NVM getting corrupted. It was later split into two mutexes - one for the NVM and one for the Phy when it was determined the NVM, unlike the Phy, should not be protected by the software/firmware/hardware semaphore (arbitration of which is done in part with the SWFLAG bit in the EXTCNF_CTRL register). This latter semaphore should be sufficient to prevent resource contention of the Phy in the driver (i.e. the mutex for Phy accesses is not needed), but to be sure the mutex is replaced with an atomic bit flag which will warn if any contention is possible. Also add additional debug output to help determine when the sw/fw/hw semaphore is owned by the firmware or hardware. Signed-off-by:Bruce Allan <bruce.w.allan@intel.com> Reported-by:
Francois Romieu <romieu@fr.zoreil.com> Tested-by:
Jeff Pieper <jeffrey.e.pieper@intel.com>
parent
96cd8951
Please register or sign in to comment