In order to correctly prevent the invalid reuse of a purged buffer, we
need to track such events and warn the user before something bad
happens.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Whilst cleaning up the patches for submission, I mis-classified non-dirty
objects as purgeable. This was causing the backing pages for those
objects to be evicted under memory-pressure, discarding valid and
unreplaceable texture data.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As evict_something() is called by routines that do not repeatedly search
again, try harder in the initial search to find an object that matches
the request.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
First the routine attempted to unlock a mutex it did not own along the
error path.
Secondly the routine should never be called on any list but the inactive
one, since we attempt to unbind those objects, so fix the calling semantics.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
By adding tracepoint equivalents for WATCH_BUF/EXEC we are able to monitor
the lifetimes of objects, requests and significant events. These events can
then be probed using the tracing frameworks, such as systemtap and, in
particular, perf.
For example to record the stack trace for every GPU stall during a run, use
$ perf record -e i915:i915_gem_request_wait_begin -c 1 -g
And
$ perf report
to view the results.
[Updated to fix compilation issues caused.]
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Ben Gamari <bgamari@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
If the presumed_offset as feed to userspace and returned to the kernel
from a previous execbuffer is still valid, then we do not need to rewrite
the relocation entry and may skip the offset sanity checks.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Eric noted a potential concern with the low bits not being strictly used
as part of the absolute offset (instead part of the command stream to the
GPU), but in practice that should not be an issue.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Andy Whitcroft <apw@canonical.com>
Cc: Eric Anholt <eric@anholt.net>
CC: stable@kernel.org
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Due to the necessity of having to take the struct_mutex, the i915
shrinker can not free the inactive lists if we fail to allocate memory
whilst processing a batch buffer, triggering an OOM and an ENOMEM that
is reported back to userspace. In order to fare better under such
circumstances we need to manually retry a failed allocation after
evicting inactive buffers.
To do so involves 3 steps:
1. Marking the backing shm pages as NORETRY.
2. Updating the get_pages() callers to evict something on failure and then
retry.
3. Revamping the evict something logic to be smarter about the required
buffer size and prefer to use volatile or clean inactive pages.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Similar to the madvise() concept, the application may wish to mark some
data as volatile. That is in the event of memory pressure the kernel is
free to discard such buffers safe in the knowledge that the application
can recreate them on demand, and is simply using these as a cache.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
This should help GEM handle memory pressure sitatuions more gracefully.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
There is no need to store the gtt_alignment as it is either explicitly
set according to the hardware requirements (e.g. scanout) or the
minimum alignment is computed on demand.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
If we failed to set the domain, the buffer was no longer being tracked
on any list.
Cc: stable@kernel.org
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
There is a very real possibility that multiple CPUs will notice that the
GPU is wedged. This introduces all sorts of potential race conditions.
Make the wedged flag atomic to mitigate this risk.
Signed-off-by: Ben Gamari <bgamari.foss@gmail.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
We set a periodic timer to check on the GPU, resetting it every time a
batch is completed. If the timer elapses, we check acthd. If acthd
hasn't changed in two timer periods, we assume the chip is wedged.
This is implemented in such a way that it leaves the option open to
employ adaptive timer intervals in the future. One could wait until
several timer periods have elapsed before declaring the chip dead. If
the chip comes back after several periods but before the "dead"
threshold, the timer interval or dead threshold could be raised.
It is important to note that while checking for active requests, we need
to account for the fact that requests are removed from the list (i.e.
retired) in a deferred work queue handler. This means that merely
checking for an empty request_list is insufficient; the list could be
non-empty yet the GPU still idle, causing the hangcheck timer to
incorrectly mark the GPU as wedged (it took me a while to figure that
out---sigh...)
Signed-off-by: Ben Gamari <bgamari.foss@gmail.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
We'll need it in i915_irq.c for checking whether there are outstanding
requests. Also, the function really ought to return a bool, not an int.
Signed-off-by: Ben Gamari <bgamari.foss@gmail.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
i915_wait_request() only checks mm.wedged after it interacts with the
hardware, generally causing the driver to lock up waiting for a wedged
chip. Make sure we check mm.wedged as the first thing we do.
Reported-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ben Gamari <bgamari.foss@gmail.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
drm_ht_remove_item() does not handle removing an absent item and the hlist
in particular is incorrectly initialised. The easy remedy is simply skip
calling i915_gem_free_mmap_offset() unless we have actually created the
offset and associated ht entry.
This also fixes the mishandling of a partially constructed offset which
leaves pointers initialized after freeing them along the
i915_gem_create_mmap_offset() error paths.
In particular this should fix the oops found here:
https://bugs.launchpad.net/ubuntu/+source/xserver-xorg-video-intel/+bug/415357/comments/8
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
Cc: stable@kernel.org
Ever since we enabled GEM, the pre-9xx chipsets (particularly 865) have had
serious stability issues. Back in May a wbinvd was added to the DRM to
work around much of the problem. Some failure remained -- easily visible
by dragging a window around on an X -retro desktop, or by looking at bugzilla.
The chipset flush was on the right track -- hitting the right amount of
memory, and it appears to be the only way to flush on these chipsets, but the
flush page was mapped uncached. As a result, the writes trying to clear the
writeback cache ended up bypassing the cache, and not flushing anything! The
wbinvd would flush out other writeback data and often cause the data we wanted
to get flushed, but not always. By removing the setting of the page to UC
and instead just clflushing the data we write to try to flush it, we get the
desired behavior with no wbinvd.
This exports clflush_cache_range(), which was laying around and happened to
basically match the code I was otherwise going to copy from the DRM.
Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Brice Goglin <Brice.Goglin@ens-lyon.org>
Cc: stable@kernel.org
Otherwise, some other userland writing into its buffer may race to land
writes either after the CPU thinks it's got a coherent view, or after its
GTT entries have been redirected to point at the scratch page. Either
result is unpleasant.
Signed-off-by: Eric Anholt <eric@anholt.net>
According to the docs, the ringbuffer is not allowed to wrap in the middle
of an instruction.
G45 PRM, Vol 1b, p101:
While the “free space” wrap may allow commands to be wrapped around the
end of the Ring Buffer, the wrap should only occur between commands.
Padding (with NOP) may be required to follow this restriction.
Do as commanded.
[Having seen bug reports where there is evidence of split commands, but
apparently the GPU has continued on merrily before a bizarre and untimely
death, this may or may not fix a few random hangs.]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
CC: Eric Anholt <eric@anholt.net>
Signed-off-by: Eric Anholt <eric@anholt.net>
There are several sources of unnecessary power consumption on Intel
graphics systems. The first is the LVDS clock. TFTs don't suffer from
persistence issues like CRTs, and so we can reduce the LVDS refresh rate
when the screen is idle. It will be automatically upclocked when
userspace triggers graphical activity. Beyond that, we can enable memory
self refresh. This allows the memory to go into a lower power state when
the graphics are idle. Finally, we can drop some clocks on the gpu
itself. All of these things can be reenabled between frames when GPU
activity is triggered, and so there should be no user visible graphical
changes.
Signed-off-by: Jesse Barnes <jesse.barnes@intel.com>
Signed-off-by: Matthew Garrett <mjg@redhat.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
Remember to release the local reference if we fail to wait on
the rendering.
(Also whilst in the vicinity add some whitespace so that the phasing of
the operations is clearer.)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
Several functions in the GEM kernel API used int as handle type, but
user API has it __u32 which is also the intended type.
Replace int with u32.
Signed-off-by: Pekka Paalanen <pq@iki.fi>
Signed-off-by: Dave Airlie <airlied@redhat.com>
As of 52dc7d32b8, we could leave an old
linear GTT mapping in place, so that apps trying to GTT-mapped write in
tiled data wouldn't get the fence added, and garbage would get displayed.
Signed-off-by: Eric Anholt <eric@anholt.net>
As we call unmap_mapping_range() twice in identical fashion, refactor
and attempt to explain why we need to call unmap_mapping_range().
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
Unitialized fence register could leads to corrupted display. Problem
encountered on MacBooks (revision 1 and 2), directly booting from EFI
or through BIOS emulation.
(bug #21710 at freedestop.org)
Signed-off-by: Grégoire Henry <henry@pps.jussieu.fr>
Signed-off-by: Eric Anholt <eric@anholt.net>
It hasn't been used in ages, and having the user tell your how much
memory is being freed at free time is a recipe for disaster even if it
was ever used.
Signed-off-by: Eric Anholt <eric@anholt.net>
The fence register value also depends upon the stride of the object, so we
need to clear the fence if that is changed as well.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[anholt: Added 8xx and 965 paths, and renamed the confusing
i915_gem_object_tiling_ok function to i915_gem_object_fence_offset_ok]
Signed-off-by: Eric Anholt <eric@anholt.net>
With the work by Jesse Barnes to eliminate allocation of fences during
execbuffer, it becomes possible to write to the scan-out buffer with it
never acquiring a fence (simply by only ever writing to the object using
tiled GPU commands and never writing to it via the GTT). So for pre-i965
chipsets which require fenced access for tiled scan-out buffers, we need
to obtain a fence register.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
After performing an operation over the page list for a buffer retrieved by
i915_gem_object_get_pages() the pages need to be returned with
i915_gem_object_put_pages(). This was not being observed for the phys
objects which were thus leaking references to their backing pages.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
CC: Dave Airlie <airlied@gmail.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
To differentiate between encountering an out-of-memory error with running
out of space in the aperture, use ENOSPC for the later.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
The batch buffer may be shared with another read buffer, so we should not
ignore any previously set domains, but just or in the command domain (and
check that the buffer is not writable).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
By sending a broken execbuffer (its length was not suitably aligned) I
triggered an operation upon a freed object. The invalid alignment was
discovered after updating the write_domain on the object but before the
object was placed on the active queue. So during the unwind process
following the error, the now freed object attempts to flush its
non-existent, but outstanding, GPU writes causing this use-after-free.
[drm:i915_dispatch_gem_execbuffer] *ERROR* alignment
[drm:i915_gem_execbuffer] *ERROR* dispatch failed -22
WARNING: at lib/kref.c:43 warn_slowpath_null+0x10/0x15()
Modules linked in:
Pid: 4552, comm: lt-csi-drm Not tainted 2.6.30-rc6 #423
Call Trace:
[<c0119ef3>] warn_slowpath_fmt+0x57/0x6d
[<c014de24>] ? get_pageblock_migratetype+0x18/0x1e
[<c014e8fd>] ? free_hot_page+0xa/0xc
[<c014e915>] ? __free_pages+0x16/0x1f
[<c0153ebf>] ? shmem_truncate_range+0x63e/0x656
[<c015fb2f>] ? slob_page_alloc+0x146/0x1c8
[<c0119f19>] warn_slowpath_null+0x10/0x15
[<c01f55f2>] kref_get+0x1b/0x21
[<c02605db>] i915_gem_object_move_to_active+0x1f/0x56
[<c0261302>] i915_add_request+0x156/0x19a
[<c026136e>] i915_gem_object_flush_gpu_write_domain+0x28/0x3f
[<c0261eca>] i915_gem_object_unbind+0x4a/0x124
[<c0261fd7>] i915_gem_free_object+0x33/0x9b
[<c0250d6b>] drm_gem_object_free+0x28/0x4a
[<c0250d43>] ? drm_gem_object_free+0x0/0x4a
[<c01f55ce>] kref_put+0x38/0x41
[<c0250cbf>] drm_gem_object_unreference+0x11/0x13
[<c0250d06>] drm_gem_object_handle_unreference+0x1e/0x21
[<c0250d13>] drm_gem_object_release_handle+0xa/0xe
[<c01f3e6b>] idr_for_each+0x5f/0x98
[<c0250d09>] ? drm_gem_object_release_handle+0x0/0xe
[<c0250daf>] drm_gem_release+0x22/0x34
[<c025046f>] drm_release+0x1e8/0x3c4
[<c0162d25>] __fput+0xaf/0x146
[<c0162dce>] fput+0x12/0x14
[<c01605ef>] filp_close+0x48/0x52
[<c011b182>] put_files_struct+0x57/0x9b
[<c011b1e4>] exit_files+0x1e/0x20
[<c011c6b6>] do_exit+0x16d/0x511
[<c03704ab>] ? __schedule+0x3d4/0x3e5
[<c0103f0d>] ? handle_irq+0xd/0x69
[<c011caa7>] do_group_exit+0x4d/0x73
[<c011cae0>] sys_exit_group+0x13/0x17
[<c010268c>] sysenter_do_call+0x12/0x2b
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
Update interrupt handling methods for IGDNG with new registers
for display and graphics interrupt functions. As we won't use
irq-based vblank sync in dri2, so display interrupt on new chip
will be used for hotplug only in future.
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
keithp didn't like the original 20ms plan because a cooperative client could
be starved by an uncooperative client. There may even have been problems
with cooperative clients versus cooperative clients. So keithp changed
throttle to just wait for the second to last seqno emitted by that client.
It worked well, until we started getting more round-trips to the server
due to DRI2 -- the server throttles in BlockHandler, and so if you did more
than one round trip after finishing your frame, you'd end up unintentionally
syncing to the swap.
Fix this by keeping track of the client's requests, so the client can wait
when it has an outstanding request over 20ms old. This should have
non-starving behavior, good behavior in the presence of restarts, and less
waiting. Improves high-settings openarena performance on my GM45 by 50%.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
This could be triggered by a gtt mapping fault on 965 that decides to
remove the fence from another object that happens to be active currently.
Since the other object doesn't rely on the fence reg for its execution, we
don't wait for it to finish. We'll soon be not waiting on 915 most of the
time as well, so just drop the BUG_ON.
Signed-off-by: Eric Anholt <eric@anholt.net>
When a GEM object is evicted from the GTT we set it to the CPU domain,
as it might get swapped in and out or ever mmapped regularly. If the
object is mmapped through the GTT it can still get evicted in this way
by other objects requiring GTT space. When the GTT mapping is touched
again we fault it back into the GTT, but fail to set it back to the
GTT domain. This means we fail to flush any cached CPU writes to the
pages backing the object which will then happen "eventually", typically
after we write to the page through the uncached GTT mapping.
[anholt: Note that userland does do a set_domain(GTT, GTT) when starting
to access the GTT mapping. That covers getting the existing mapping of the
object synchronized if it's bound to the GTT. But set_domain(GTT, GTT)
doesn't do anything if the object is currently unbound. This fix covers the
transition to being bound for GTT mapping.]
Fixes glyph and other pixmap corruption during swapping. fd.o bug #21790
Signed-off-by: Kristian Høgsberg <krh@redhat.com>
Signed-off-by: Eric Anholt <eric@anholt.net>
On the 865, but not the 855, the clflush we do appears to not actually make
it out to the hardware all the time. An easy way to safely reproduce was
X -retro, which would show that some of the blits involved in drawing the
lovely root weave didn't make it out to the hardware. Those blits are 32
bytes each, and 1-2 would be missing at various points around the screen.
Other experimentation (doing more clflush, doing more AGP chipset flush,
poking at some more device registers to maybe trigger more flushing) didn't
help. krh came up with the wbinvd as a way to successfully get all those
blits to appear.
Signed-off-by: Eric Anholt <eric@anholt.net>
The pitch field is an exponent on pre-965, so we were rejecting buffers
on 8xx that we shouldn't have. 915 got lucky in that the largest legal
value happened to match (8KB / 512 = 0x10), but 8xx has a smaller tile width.
Additionally, we programmed that bad value into the register on 8xx, so the
only pitch that would work correctly was 4096 (512-1023 pixels), while others
would probably give bad rendering or hangs.
Signed-off-by: Eric Anholt <eric@anholt.net>
fd.o bug #20473.
For some reason we never added 8xx desktop cursor support to the
kernel. This patch fixes that.
[krh: Also set the size on pre-i915 hw.]
Tested-by: Kristian Høgsberg <krh@redhat.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Eric Anholt <eric@anholt.net>
