Virtual address need to be fenced to know when we can safely remove it.
This patch also properly clear the pagetable. Previously it was
serouisly broken.
Kernel 3.5/3.4 need a similar patch but adapted for difference in mutex locking.
v2: For to update pagetable when unbinding bo (don't bailout if
bo_va->valid is true).
v3: Add kernel 3.5/3.4 comment.
v4: Fix compilation warnings.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Const IBs are executed on the CE not the CP, so we can't
fence them in the normal way.
So submit them directly before the IB instead, just as
the documentation says.
v2: keep the extra documentation
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
GPU reset need to be exclusive, one happening at a time. For this
add a rw semaphore so that any path that trigger GPU activities
have to take the semaphore as a reader thus allowing concurency.
The GPU reset path take the semaphore as a writer ensuring that
no concurrent reset take place.
v2: init rw semaphore
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Don't return success if scheduling the IB fails, otherwise
we end up with an oops in ttm_eu_fence_buffer_objects.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Try to remove or replace the cs_mutex with a
vm_mutex where it is still needed.
v2: fix locking order
v3: rebased on drm-next
Signed-off-by: Christian König <deathsimple@vodafone.de>
Move inter ring syncing with semaphores into the
existing ring allocations, with that we need to
lock the ring mutex only once.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
radeon_cs_parser_init is called by both the legacy UMS
CS ioctl and the KMS CS ioctl. Protect KMS specific
pieces of the code by checking that rdev is not NULL.
Reported-by: Michael Burian <michael.burian@sbg.at>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
Just move its only caller into the same file as it and make it static.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
No need to malloc it any more.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
It never really belonged there in the first place.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Instead of offset + size keep start and end offset directly.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
We need to sync with the GFX ring as ttm might have schedule bo move
on it and new command scheduled for other ring need to wait for bo
data to be in place.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed by: Christian König <christian.koenig@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Instead of all this humpy pumpy with recursive
mutex (which also fixes only halve of the problem)
move the actual gpu reset out of the fence code,
return -EDEADLK and then reset the gpu in the
calling ioctl function.
v2: Split removal of radeon_mutex into separate patch.
Return -EAGAIN if reset is successful.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Rings need to lock in order, otherwise
the ring subsystem can deadlock.
v2: fix error handling and number of locked doublewords.
v3: stop creating unneeded semaphores.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
On coherent systems (not-AGP) the IB should be in cached memory so should
be just as fast, so we can avoid copying to temporary pages and just use it
directly.
provides minor speedups on rv530: gears ~1820->1860, ipers: 29.9->30.6,
but always good to use less CPU if we can.
v3: cleanup unneeded bits.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Very basic implementation for picking the ring priority.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This adds a new chunk id to the CS ioctl to support the
INDIRECT_BUFFER_CONST packet.
On SI, the CP adds a new engine called the CE (Constant Engine)
which runs simulatenously with the DE (Drawing Engine, formerly
called the ME). This allows the CP to process two related IBs
simultaneously. The CE is tasked with loading the constant data
(constant buffers, resource descriptors, samplers, etc.) while
the DE loads context register state and issues drawing commands.
It's up to the userspace application to sychronize the CE and the
DE using special synchronization packets.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The function radeon_cs_parser_init is only called from two places, in
drivers/gpu/drm/radeon/radeon_cs.c and drivers/gpu/drm/radeon/r600_cs.c.
In each case, if the call fails another function is called that frees all
of the kdata and dpage information in the chunks array. So this
information should not be freed in radeon_cs_parser_init as well.
Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Not all rings use PM4, so the cs_parser also needs to be per ring.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The function radeon_bo_list_validate can cause a
bo to move, resulting in a different sync_obj
and a dependency to wait for this move to finish.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
If accel is not working many subsystem such as the ib pool might not be
initialized properly that can lead to segfault inside kernel when cs
ioctl is call with non working acceleration. To avoid this make sure
the accel working flag is false when an error in GPU startup happen and
return EBUSY from cs ioctl if accel is not working.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Add a VM manager enabled field and use it to check if
vm is enabled.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: jglisse@redhat.com
Signed-off-by: Dave Airlie <airlied@redhat.com>
Use semaphores to sync buffers across rings in the CS
ioctl. Add a reloc flag to allow userspace to skip
sync for buffers.
agd5f: port to latest CS ioctl changes.
v2: add ring lock/unlock to make sure changes hit the ring.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.
First 16M of virtual address space is reserved by the kernel.
Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).
Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.
v2: agd5f fixes:
- Add vram base offset for vram pages. The GPU physical address of a
vram page is FB_OFFSET + page offset. FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1
v3: agd5f:
- integrate with the semaphore/multi-ring stuff
v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
chunk
v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path
v6:
- fix update of page table to only happen on valid mem placement
v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
to then do ring specific scheduling if vm ib scheduling function
v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support
v9:
- rebuild page table only when bind and incrementaly depending
on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup
v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
the callback and use the existing ib_execute callback for VM IBs.
v11:
- rebase on top of lastest Linus
v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()
v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS
v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm
v15:
- rebase
v16:
- cleanup left over of vm init/fini
v17: agd5f:
- cs checker
v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM. Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode. Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.
v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes
v20: agd5f:
- further CS ioctl cleanups from Christian's comments
v21: agd5f:
- integrate CS checker improvements
v22: agd5f:
- final cleanups for release, only allow VM CS on cayman
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This avoid to waste ib pool size and avoid a bunch of wait for
previous ib to finish.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Give all asic and radeon_ring_* functions a
radeon_cp parameter, so they know the ring to work with.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Only check the previously checked relocs for
duplicates. Also leaving the handle uninitialized
isn't such a good idea.
Signed-off-by: Christian König <deathsimple@vodafone.de>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This adds a new optional chunk to the CS ioctl that specifies optional flags
to the CS parser. Why this is useful is explained below. Note that some regs
no longer need the NOP relocation packet if this feature is enabled.
Tested on r300g and r600g with this flag disabled and enabled.
Assume there are two contexts sharing the same mipmapped tiled texture.
One context wants to render into the first mipmap and the other one
wants to render into the last mipmap. As you probably know, the hardware
has a MACRO_SWITCH feature, which turns off macro tiling for small mipmaps,
but that only applies to samplers.
(at least on r300-r500, though later hardware likely behaves the same)
So we want to just re-set the tiling flags before rendering (writing
packets), right? ... No. The contexts run in parallel, so they may
set the tiling flags simultaneously and then fire their command streams
also simultaneously. The last one setting the flags wins, the other one
loses.
Another problem is when one context wants to render into the first and
the last mipmap in one CS. Impossible. It must flush before changing
tiling flags and do the rendering into the smaller mipmaps in another CS.
Yet another problem is that writing copy_blit in userspace would be a mess
involving re-setting tiling flags to please the kernel, and causing races
with other contexts at the same time.
The only way out of this is to send tiling flags with each CS, ideally
with each relocation. But we already do that through the registers.
So let's just use what we have in the registers.
Signed-off-by: Marek Olšák <maraeo@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This was only the case if the GPU reset was triggered from the CS ioctl,
otherwise other processes could happily enter the CS ioctl and wreak havoc
during the GPU reset.
This is a little complicated because the GPU reset can be triggered from the
CS ioctl, in which case we're already holding the mutex, or from other call
paths, in which case we need to lock the mutex. AFAICT the mutex API doesn't
allow recursive locking or finding out the mutex owner, so we need to handle
this with helper functions which allow recursive locking from the same
process.
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Wierdly the kms parser never initialised the family, it wasn't really used
much, but the fmt checker patch started using it and it fell over.
Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>
... and switch it to container_of upcasting.
v2: converted new pageflip code-paths.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Rather than re-implementing in the Radeon driver,
Use the execbuf / cs / pushbuf utilities that comes with TTM.
This comes with an even greater benefit now that many spinlocks have been
optimized away...
Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This is consistent with trying to access a filename that not exist
within a directory which is a good analogy here. The main reason for the
change is that it is easy to confuse the error code of EBADF as an
performing an ioctl on an invalid file descriptor (rather than an
unknown object).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This simplify and improve GPU reset for R1XX-R6XX hw, it's
not 100% reliable here are result:
- R1XX/R2XX works bunch of time in a row, sometimes it
seems it can work indifinitly
- R3XX/R3XX the most unreliable one, sometimes you will be
able to reset few times, sometimes not even once
- R5XX more reliable than previous hw, seems to work most
of the times but once in a while it fails for no obvious
reasons (same status than previous reset just no same
happy ending)
- R6XX/R7XX are lot more reliable with this patch, still
it seems that it can fail after a bunch (reset every
2sec for 3hour bring down the GPU & computer)
This have been tested on various hw, for some odd reasons
i wasn't able to lockup RS480/RS690 (while they use to
love locking up).
Note that on R1XX-R5XX the cursor will disapear after
lockup haven't checked why, switch to console and back
to X will restore cursor.
Next step is to record the bogus command that leaded to
the lockup.
V2 Fix r6xx resume path to avoid reinitializing blit
module, use the gpu_lockup boolean to avoid entering
inifinite waiting loop on fence while reiniting the GPU
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
When there is allocation failure in radeon_cs_parser_relocs parser->nrelocs
is not cleaned. This causes NULL pointer defeference in radeon_cs_parser_fini
when clean up code is trying to loop over the relocation array and free the
objects.
Fix adds a check for a possible NULL pointer in clean up code.
Signed-off-by: Pauli Nieminen <suokkos@gmail.com>
Cc: stable@kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
We can get this if the user moves the mouse when we are waiting to move
some stuff around in the validate. Don't fail.
Cc: stable@kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
Previous code did associate fence to bo before the fence was emited
and it also didn't lock protected access to ttm sync_obj member.
Both of this flaw leads to possible race between different code
path. This patch fix this by associating fence only once the fence
is emitted and properly lock protect access to sync_obj member of
ttm.
Fix:
https://bugs.freedesktop.org/show_bug.cgi?id=26438
and likely similar others bugs
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Mostly obvious simplifications.
The i915 pread/pwrite ioctls, intel_overlay_put_image and
nouveau_gem_new were incorrectly using the locked versions
without locking: this is also fixed in this patch.
Signed-off-by: Luca Barbieri <luca@luca-barbieri.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
If radeon_cs_parser_init() fails, radeon_cs_ioctl() calls
radeon_cs_parser_fini() with the non-zero error value. The latter dereferenced
parser->ib which hasn't been initialized yet -> boom. Add a test for parser->ib
being non-NULL before dereferencing it.
Signed-off-by: Michel Dänzer <daenzer@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This patch workaround a possible security issue which can allow
user to abuse drm on r6xx/r7xx hw to access any system ram memory.
This patch doesn't break userspace, it detect "valid" old use of
CB_COLOR[0-7]_FRAG & CB_COLOR[0-7]_TILE registers and overwritte
the address these registers are pointing to with the one of the
last color buffer. This workaround will work for old mesa &
xf86-video-ati and any old user which did use similar register
programming pattern as those (we expect that there is no others
user of those ioctl except possibly a malicious one). This patch
add a warning if it detects such usage, warning encourage people
to update their mesa & xf86-video-ati. New userspace will submit
proper relocation.
Fix for xf86-video-ati / mesa (this kernel patch is enough to
prevent abuse, fix for userspace are to set proper cs stream and
avoid kernel warning) :
http://cgit.freedesktop.org/xorg/driver/xf86-video-ati/commit/?id=95d63e408cc88b6934bec84a0b1ef94dfe8bee7bhttp://cgit.freedesktop.org/mesa/mesa/commit/?id=46dc6fd3ed5ef96cda53641a97bc68c3bc104a9f
Abusing this register to perform system ram memory is not easy,
here is outline on how it could be achieve. First attacker must
have access to the drm device and be able to submit command stream
throught cs ioctl. Then attacker must build a proper command stream
for r6xx/r7xx hw which will abuse the FRAG or TILE buffer to
overwrite the GPU GART which is in VRAM. To achieve so attacker
as to setup CB_COLOR[0-7]_FRAG or CB_COLOR[0-7]_TILE to point
to the GPU GART, then it has to find a way to write predictable
value into those buffer (with little cleverness i believe this
can be done but this is an hard task). Once attacker have such
program it can overwritte GPU GART to program GPU gart to point
anywhere in system memory. It then can reusse same method as he
used to reprogram GART to overwritte the system ram through the
GART mapping. In the process the attacker has to be carefull to
not overwritte any sensitive area of the GART table, like ring
or IB gart entry as it will more then likely lead to GPU lockup.
Bottom line is that i think it's very hard to use this flaw
to get system ram access but in theory one can achieve so.
Side note: I am not aware of anyone ever using the GPU as an
attack vector, nevertheless we take great care in the opensource
driver to try to detect and forbid malicious use of GPU. I don't
think the closed source driver are as cautious as we are.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@linux.ie>
The locking & protection of radeon object was somewhat messy.
This patch completely rework it to now use ttm reserve as a
protection for the radeon object structure member. It also
shrink down the various radeon object structure by removing
field which were redondant with the ttm information. Last it
converts few simple functions to inline which should with
performances.
airlied: rebase on top of r600 and other changes.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Thanks to Michel for pointing this out to me, this is
why I need to get more sleep, over complicate this a bit.
Signed-off-by: Dave Airlie <airlied@redhat.com>
This avoids needing to do a kmalloc > PAGE_SIZE for the main
indirect buffer chunk, it adds an accessor for all reads from
the chunk and caches a single page at a time for subsequent
reads.
changes since v1:
Use a two page pool which should be the most common case
a single packet spanning > PAGE_SIZE will be hit, but I'm
having trouble seeing anywhere we currently generate anything like that.
hopefully proper short page copying at end
added parser_error flag to set deep errors instead of having to test
every ib value fetch.
fixed bug in patch that went to list.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Jerome Glisse