The i915_vblank_swap() function schedules an automatic buffer swap
upon receipt of the vertical sync interrupt. Such an operation is
lengthy so it can't be allowed to happen in normal interrupt context,
thus the DRM implements this by scheduling the work in a kernel
softirq-scheduled tasklet. In order for the buffer swap to work
safely, the DRM's central lock must be taken, via a call to
drm_lock_take() located in drivers/char/drm/drm_irq.c within the
function drm_locked_tasklet_func(). The lock-taking logic uses a
non-interrupt-blocking spinlock to implement the manipulations needed
to take the lock. This semantic would be safe if all attempts to use
the spinlock only happen from process context. However this buffer
swap happens from softirq context which is really a form of interrupt
context. Thus we have an unsafe situation, in that
drm_locked_tasklet_func() can block on a spinlock already taken by a
thread in process context which will never get scheduled again because
of the blocked softirq tasklet. This wedges the kernel hard.
To trigger this bug, run a dual-head cloned mode configuration which
uses the i915 drm, then execute an opengl application which
synchronizes buffer swaps against the vertical sync interrupt. In my
testing, a lockup always results after running anywhere from 5 minutes
to an hour and a half. I believe dual-head is needed to really
trigger the problem because then the vertical sync interrupt handling
is no longer predictable (due to being interrupt-sourced from two
different heads running at different speeds). This raises the
probability of the tasklet trying to run while the userspace DRI is
doing things to the GPU (and manipulating the DRM lock).
The fix is to change the relevant spinlock semantics to be the
interrupt-blocking form. After this change I am no longer able to
trigger the lockup; the longest test run so far was 20 hours (test
stopped after that point).
Note: I have examined the places where this spinlock is being
employed; all are reasonably short bounded sequences and should be
suitable for interrupts being blocked without impacting overall kernel
interrupt response latency.
Signed-off-by: Mike Isely <isely@pobox.com>
The data is now in kernel space, copied in/out as appropriate according to the
This results in DRM_COPY_{TO,FROM}_USER going away, and error paths to deal
with those failures. This also means that XFree86 4.2.0 support for i810 DRM
is lost.
As a fallout, replace filp storage with file_priv storage for "unique
identifier of a client" all over the DRM. There is a 1:1 mapping, so this
should be a noop. This could be a minor performance improvement, as everything
on Linux dereferenced filp to get file_priv anyway, while only the mmap ioctls
went the other direction.
This was used to make all ioctl handlers return -errno on linux and errno on
*BSD. Instead, just return -errno in shared code, and flip sign on return from
shared code to *BSD code.
Add refcounting of user waiters to the DRM hardware lock, so that we can use the
DRM_LOCK_CONT flag more conservatively.
Also add a kernel waiter refcount that if nonzero transfers the lock for the kernel context,
when it is released. This is useful when waiting for idle and can be used
for very simple fence object driver implementations for the new memory manager.
It also resolves the AIGLX startup deadlock for the sis and the via drivers.
i810, i830 still require that the hardware lock is really taken so the deadlock remains
for those two. I'm not sure about ffb. Anyone familiar with that code?
flag DMA_QUIESCENT (typically the X server), but gets interrupted by a
signal. The locking IOCTL should then return an error, but if
DMA_QUIESCENT succeeds it returns 0, and the client falsely thinks it
has the lock. In addition The client waits for DMA_QUISCENT and
possibly DMA_READY without having the lock.
Alan Hourihane
Mike Isely
Eric Anholt
Dave Airlie
Thomas Hellstrom
Michel Dänzer
Thomas Hellstrom
Jose Fonseca
Jon Smirl
Keith Whitwell
Michel Daenzer
Gareth Hughes