Fix locking in the page fault handler. A (somewhat malicious) userland

program could force a copyin/copyout from/to memory mapped through the GTT,
forcing a "locking against myself" panic.  The intel-gpu-tools "package" has
a test for this.  The problem can be circumvented by making the fault handler
fail if we already hold the (DRM) lock.  This will make the copyin/copyout
return with EFAULT making the caller fall back on a "slow path".

This makes it obvious that using a shared (read) lock here doesn't make any
sense.  So use an exclusive (write) lock like everywhere else in the inteldrm
code.

0 files changed, 0 insertions, 0 deletions