https://bugzilla.kernel.org/show_bug.cgi?id=209253 --- Comment #12 from Alex Williamson (alex.williamson@xxxxxxxxxx) --- Created attachment 293281 --> https://bugzilla.kernel.org/attachment.cgi?id=293281&action=edit Test fix for ioeventfd_write traces For those experiencing the issue described by Ian in comment 1 and the first issue from Martin in comment 2, please try this patch if you're able. vfio_pci_ioeventfd_handler() is called in a spinlock context with interrupts disabled and tries to acquire a read lock on the memory semaphore to verify the device memory is enabled. The down_read() call can sleep and therefore should instead be called from a thread context if there is contention. TBH, I've never seen it contended, so in practice the thread is never really used, but this seems to solve the specific case Ian has identified and it would have been introduced by the commit noted in comment 5. This change should be specific to configurations with NVIDIA GPUs assigned. I'm curious about the traces with vfio_msihandler as well, if anyone can provide a reproducer of that it would be appreciated. -- You are receiving this mail because: You are watching the assignee of the bug.
