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?
Memory types are either fixed (on-card or pre-bound AGP) or not fixed
(dynamically bound) to an aperture. They also carry information about:
1) Whether they can be mapped cached.
2) Whether they are at all mappable.
3) Whether they need an ioremap to be accessible from kernel space.
In this way VRAM memory and, for example, pre-bound AGP appear
identical to the memory manager.
This also makes support for unmappable VRAM simple to implement.
* Pulled in some registers from nv10reg.h. Needed for context switching.
* Filled in nv30 graphics context (based on nv40_graph.c).
* Figure out nv30 context table, set up on context creation. Allows the cards automatic switching to work.
This comes from the Re: [patch] paravirt: isolate module ops on lkml
It needs some testing, please report any regressions caused.
Signed-off-by: Dave Airlie <airlied@linux.ie>
The vmalloc_32 function together with the memset to clear
the new pages are replaced with a vmalloc_user.
A pre-2.6.18 compat vmalloc_user is added.
Please replace any breakage on machines with > 1GB of memory.
ioremap must be balanced by an iounmap and failing to do so can result
in a memory leak.
Tested (compilation only) to make sure the files are compiling without
any warning/error due to new changes
Signed-off-by: Amol Lad <amol@verismonetworks.com>
Signed-off-by: Dave Airlie <airlied@linux.ie>
I need the following patch to fix compilation of
latest drm/linux-core on my ppc64 machine.
/home/mb/develop/git/drm/linux-core/savage_bci.c: In function ‘savage_driver_firstopen’:
/home/mb/develop/git/drm/linux-core/savage_bci.c:587: error: ‘DRM_MTRR_WC’ undeclared (first use in this function)
/home/mb/develop/git/drm/linux-core/savage_bci.c:587: error: (Each undeclared identifier is reported only once
/home/mb/develop/git/drm/linux-core/savage_bci.c:587: error: for each function it appears in.)
/home/mb/develop/git/drm/linux-core/savage_bci.c: In function ‘savage_driver_lastclose’:
/home/mb/develop/git/drm/linux-core/savage_bci.c:664: error: ‘DRM_MTRR_WC’ undeclared (first use in this function)
I looked at in-kernel drmP.h and it actually
has the same fix in it.
Signed-off-by: Michael Buesch <mb@bu3sch.de>
The current version didn't build on BSD, where the new functionality isn't used
yet anyway. Whoever changes that will hopefully be able to make the OSes share
this file as well.
This will come in very handy for tiled buffers on intel hardware.
Also add some padding to interface structures to allow future binary backwards
compatible changes.
This makes an allocated block actually align itself and returns any
wasted space to the manager.
Also add some functions to grow and shrink the managed area.
This will be used in the future to manage the buffer object swap cache.
otherwise data will be missing, which becomes apparent when the kernel evicts
batch buffers which are likely to be written into in the evicted state,
and then rebound to the AGP aperture.
This means we cannot rely on the AGP module to flush the
cache for us.
driver we're on.
Avoid global cache flushes before inserting pages.
In general, they are never mapped, and not accessed through the kernel map, so
a cache flush should not be necessary. The exception is pages that are bound
cached. We might need a cache flush for those.
The mm_lock function is used when leaving vt. It evicts _all_ buffers.
Buffers with the DRM_BO_NO_MOVE attribute set will be guaranteed to
get the same offset when / if they are rebound.
Fix buffer bound caching policy changing, Allow
on-the-fly changing of caching policy on bound buffers if the hardware
supports it.
Allow drivers to use driver-specific AGP memory types for TTM AGP pages.
Will make AGP drivers much easier to migrate.
Adapt for new functions in the 2.6.19 kernel.
Remove the ability to have multiple regions in one TTM.
This simplifies a lot of code.
Remove the ability to access TTMs from user space.
We don't need it anymore without ttm regions.
Don't change caching policy for evicted buffers. Instead change it only
when the buffer is accessed by the CPU (on the first page fault).
This tremendously speeds up eviction rates.
Current code is safe for kernels <= 2.6.14.
Should also be OK with 2.6.19 and above.
Initialize the spinlock unconditionally when struct drm_device is filled in,
and return early in drm_locked_tasklet() if the driver doesn't support IRQs.
It looks like this would have caused signals to always get sent on the next
vertical blank, regardless of the sequence number.
(cherry picked from cf6b2c5299 commit)
When this flag is set and the target sequence is missed, wait for the next
vertical blank instead of returning immediately.
(cherry picked from 89e323e490 commit)
This makes it easier for userspace to know when it needs to allocate an ID.
Also free drawable information memory when it's no longer needed.
(cherry picked from df7551ef73 commit)
Actually make the existing ioctls for adding and removing drawables do
something useful, and add another ioctl for the X server to update drawable
information. The only kind of drawable information tracked so far is cliprects.
(cherry picked from 29598e5253 commit)
When the vertical blank interrupt is enabled for both pipes, pipe A is
considered primary and pipe B secondary. When it's only enabled for one pipe,
it's always considered primary for backwards compatibility.
(cherry picked from 0c7d7f4361 commit)
Actually make the existing ioctls for adding and removing drawables do
something useful, and add another ioctl for the X server to update drawable
information. The only kind of drawable information tracked so far is cliprects.
When the vertical blank interrupt is enabled for both pipes, pipe A is
considered primary and pipe B secondary. When it's only enabled for one pipe,
it's always considered primary for backwards compatibility.
Change the fence object interface somewhat to allow some more flexibility.
Make list IOCTLS really restartable.
Try to avoid busy-waits in the kernel using immediate return to user-space with an -EAGAIN.
Add a function to free buffers on hold for destruction if their
fence object has expired.
Add a timer to periodically call that function when there are
buffers pending deletion.
First warning result from open-coded PTR_ERR,
the rest is caused by code like this:
*(u32 *) ((u32) buf_priv->kernel_virtual + used)
I've also fixed a missing PTR_ERR in i830_dma.c
From: Denis Vlasenko <vda.linux@googlemail.com>
Signed-off-by: Dave Airlie <airlied@linux.ie>
0x00 EXE fence. Signals when command stream interpreter has reached the point
where the fence was emitted.
0x01 FLUSH fence. Signals when command stream interpreter has reached the point
where the fence was emitted, and all previous drawing operations have been
completed and flushed.
Implements busy wait (for fastest response time / high CPU) and
lazy wait (User interrupt or timer driven).
First warning result from open-coded PTR_ERR,
the rest is caused by code like this:
*(u32 *) ((u32) buf_priv->kernel_virtual + used)
I've also fixed a missing PTR_ERR in i830_dma.c
From: Denis Vlasenko <vda.linux@googlemail.com>
Signed-off-by: Dave Airlie <airlied@linux.ie>
32-bit physical device addresses are mapped directly to user-tokens. No
duplicate maps are allowed, and the addresses are assumed to be outside
of the range 0x10000000 through 0x30000000. The user-token is identical
to the 32-bit physical start-address of the map.
64-bit physical device addressed are mapped to user-tokens in the range
0x10000000 to 0x30000000 with page-size increments. The user_token should
not be interpreted as an address.
Other map types, like upcoming TTM maps are mapped to user-tokens in the
range
0x10000000 to 0x30000000 with page-size increments. The user_token should
not be interpreted as an address.
This keeps compatibility with buggy drivers, while still implementing a
hashed map lookup. The SiS and via device driver major bumps are
reverted.
0x10000000 to 0x90000000 in PAGE_SIZE increments.
Implement hashed map lookups.
This potentially breaks both 2D and 3D drivers. If so, the corresponding
2D and 3D driver should be fixed, and it's corresponding drm device driver
should have its major bumped as soon as possible.
Bump sis and via drm device driver majors.
The SiS and Unichrome 3D drivers are fixed in Mesa CVS HEAD and
mesa_6_4_branch.
drm_mtrr_{add,del} for handling the MTRR setup. Still has a LOR issue
with DRM_VERIFYAREA_READ/DRM_COPY_FROM_USER_UNCHECKED in savage_bci.c
-- this won't work with the fine-grained locking in use, and just doing
a single copyin to a temporary will probably work fine. Also note that
the module leaks approximately 4 kb on unload.
radeon_cp.c to use a drm_local_map_t-type mapping (drm_core_ioremap
rather than drm_ioremap), which contains private device mapping
information on BSD. I also changed the ati_pcigart interface to use
"void *" for pointers to kva rather than "unsigned long". While PCIGART
support appears to be broken on FreeBSD currently, I think this is not
new, and BusType PCI remains working on my r100 in Linux.
drm_agp_foo_ioctl pair. Modifies the MGA DRM to use the drm_agp_foo
functions instead of the drm_foo_agp functions. The drm_foo_agp
functions are no longer exported by drm.ko.
Ensures that dma->seg_count and dma->page_count are properly set in
drm_addbufs_{agp,sg,fb}. drm_addbufs_pci was already correct.
Ensures that mga_do_agp_dma_bootstrap correctly sets agp_buffer_token.
At this point PCI DMA is still broken.
Xorg bug: #4797 Reviewed by: Dave Airlie, Eric Anholt Signed-off-by: Ian
Romanick <idr@us.ibm.com>
me to match other drivers and avoid ifdeffing. The linux via_drv.c will
be moved from shared-core to linux-core soon by repocopy.
Submitted by: Jake Burkholder <jake@FreeBSD.org> Tested by: unichrome
caller on fb / agp memory alloc and free. Otherwise malicious clients
can register allocations on other clients or free memory used by other
clients which will lead to severe memory manager inconsistensies.
understandable: preinit -> load postinit -> (removed) presetup ->
firstopen postsetup -> (removed) open_helper -> open prerelease ->
preclose free_filp_priv -> postclose pretakedown -> lastclose
postcleanup -> unload release -> reclaim_buffers_locked version ->
(removed)
postinit and version were replaced with generic code in the Linux DRM
(drivers now set their version numbers and description in the driver
structure, like on BSD). postsetup wasn't used at all. Fixes the savage
hooks for initializing and tearing down mappings at the right times.
Testing involved at least starting X, running glxgears, killing
glxgears, exiting X, and repeating.
Tested on: FreeBSD (g200, g400, r200, r128) Linux (r200, savage4)
buffers that have been allocated from AGP. This includes some new
capable(CAP_SYS_ADMIN) checks, these functions are also protected by
the root requirement on the IOCTL macros.
code duplication, and it also hands you the map pointer so you don't
need to re-find it.
- Remove the permanent maps flag. Instead, for register and framebuffer
maps, we always check whether there's already a map of that type and
offset around. Move the Radeon map initialization into presetup (first
open) so it happens again after every takedown.
- Remove the split cleanup of maps between driver takedown (last close) and
cleanup (module unload). Instead, always tear down maps on takedown,
and drivers can recreate them on first open.
- Make MGA always use addmap, instead of allocating consistent memory in
the PCI case and then faking up a map for it, which accomplished nearly
the same thing, in a different order. Note that the maps are exposed to
the user again: we may want to expose a flag to avoid this, but it's
not a security concern, and saves us a lot of code.
- Remove rmmaps in the MGA driver. Since the function is only called during
takedown anyway, we can let them die a natural death.
- Make removal of maps happen in one function, which is called by both
drm_takedown and drm_rmmap_ioctl.
Reviewed by: idr (previous revision) Tested on: mga (old/new/pci dma),
radeon, savage
This patch adds serveral new ioctls and a new query to get_param query to
support PCI MGA cards.
Two ioctls were added to implement interrupt based waiting. With this
change, the client-side driver no longer needs to map the primary DMA
region or the MMIO region. Previously, end-of-frame waiting was done by
busy waiting in the client-side driver until one of the MMIO registers
(the current DMA pointer) matched a pointer to the end of primary DMA
space. By using interrupts, the busy waiting and the extra mappings are
removed.
A third ioctl was added to bootstrap DMA. This ioctl, which is used by the
X-server, moves a *LOT* of code from the X-server into the kernel. This
allows the kernel to do whatever needs to be done to setup DMA buffers.
The entire process and the locations of the buffers are hidden from
user-mode.
Additionally, a get_param query was added to differentiate between G4x0
cards and G550 cards. A gap was left in the numbering sequence so that,
if needed, G450 cards could be distinguished from G400 cards. According
to Ville Syrjälä, the G4x0 cards and the G550 cards handle
anisotropic filtering differently. This seems the most compatible way
to let the client-side driver know which card it's own. Doing this very
small change now eliminates the need to bump the DRM minor version
twice.
http://marc.theaimsgroup.com/?l=dri-devel&m=106625815319773&w=2
A number of ioctl handlers in linux-core were also modified so that they
could be called in-kernel. In these cases, the in-kernel callable
version kept the existing name (e.g., drm_agp_acquire) and the ioctl
handler added _ioctl to the name (e.g., drm_agp_acquire_ioctl).
This patch also replaces the drm_agp_do_release function with
drm_agp_release. drm_agp_release (drm_core_agp_release in the previous
patch) is very similar to drm_agp_do_release, and I saw no reason to
have both.
This commit *breaks the build* on BSD. Eric said that he would make the
required updates to the BSD side soon.
Xorg bug: 3259 Reviewed by: Eric Anholt
There were two problems. First, the 'warp' and 'primary' pointers weren't
cleared, so mga_do_cleanup_dma, which gets called multiple times, would
try to ioremapfree them multiple times. This resulted in the new error
messages to syslog. The second problem was the, since the dev_private
structure isn't reallocated and cleaned out in mga_do_init_dma, when
the server is reloaded idle-waits would wait for impossible values.
I have given this patch some more riggorous testing. This includes:
- Load module, start server, run GL app, stop server, unload module.
- Load module, start server, run GL app, stop server, unload module, reload
module, restart server, run GL app.
- Load module, start server, run GL app, stop server, restart server, run
GL app, stop server, unload module.
In all three cases, everything worked as expected. Please let me know if
there are any further regressions with this patch.
Xorg bug: 3408 Reported by: Chris Rankin
linux-core to free pci memory without freeing the structure. Linux-core
internals often create pci dma handle structures on the stack due to
the lack of a drm_local_map_t to store them in properly. Fix the
original drm_pci_free to actually free the dma handle structure instead
of leaking it.
Signed-off-by: Sergey Vlasov <vsu@altlinux.ru>
There's two choices when fb is or isn't loaded as we treat ourselves as a
PCI driver in the latter case.
If we are a PCI driver, then register the suspend/resume functions
directly. If not, then we register as a sysdev and pick up the
suspend/resume actions and pump them down into a generic *power
function.
It'll be nice when this little mess is sorted out with regard to being a
real PCI driver ;-/
that a device absolutely is, absolutely is not, or may or may not be
AGP. Modify the i915 DRM to use this to force all i9x5 devices to be
"AGP" (even the PCI-e devices).
Reported by: Lukas Hejtmanek
is now allocated (and partially filled in) by the new
mga_driver_preinit function.
This allows the driver to detect the type of card (i.e., G200 class vs.
G400 class) on its own. The chipset value passed to mga_dma_init is now
ignored. This same technique is used by the radeon DRM.
As a result of this, mga_driver_pretakedown was converted to
mga_driver_postcleanup. This routine gets called in some other places
than might be expected, and it sets the dev_private pointer to NULL.
That little gem took over an hour to track down. :(
platform-specific drm_device_is_agp function. Added implementation of
this function the the Linux-specific portion of the MGA driver to
detect PCI G450 cards. Added code to the Linux-specific portion of the
generic DRM layer to not initialize AGP infrastructure if the card is
not AGP (this matches what already existed in BSD).
Bumped the driver date and the driver patch-level for MGA.
This mostly fixes bugzilla #3248. The BSD side still needs an
implementation of mga_driver_device_is_agp.
with IOMMUs and such. There is one usage of the forbidden vtophys()
left in drm_scatter.c which will be fixed up soon. This required a KPI
change for drm_pci_alloc/free() to return/use a drm_dma_handle_t that
keeps track of os-specific bits, rather than just passing around the
vaddr/busaddr/size.
Submitted by: Tonnerre Lombard (partially) Tested on: FreeBSD: Rage128
AGP/PCI Linux: Savage4 AGP/PCI
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.
New PCI command parser. Moved from via_dma.c to via_verifier.c so functions
with similar functionality are close to eachother.
Moved video related functions to via_video.c, which might be extended in
the future, as new video functionality is added.
New device-specific generic IRQ IOCTL, similar to the general VBLANK IOCTL,
but with support for multiple device IRQ sources and functionality.
Support for Unichrome Pro PM800/CN400 video DMA commands in verifier and
PCI parser.
Support for Unichrome Pro PM800/CN400 HQV IRQs in the new generic IRQ
IOCTL.
Bumped minor. New version 2.6.0.
what wait_event_interruptible_timeout() does, use the function and just
change the return values appropriately.
Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
so we need to consult the EFI memory map before we try to set the write
combine attribute of a page. This patch will try to map a page write
combined if it's not an AGP page and the EFI memory map says it's ok,
otherwise it falls back to a regular, uncached mapping. Can someone
please apply this to the drm tree?
From: Jesse Barnes
privileges on Radeon hardware. Essentially, a malicious program could
submit a packet containing an offset (possibly in main memory) to be
rendered from/to, while a separate thread switched that offset in
userspace rapidly between a valid value and an invalid one.
radeon_check_and_fixup_offset() would pull the offset in from user
space, check it, and spit it back out to user space to be copied in
later by the emit code. It would sometimes catch the bad value, but
sometimes the malicious program could modify it after the check and get
an invalid offset rendered from/to.
Fix this by allocating a temporary buffer and copying the data in at once.
While here, make the cliprects stuff not do the VERIFYAREA_READ and
COPY_FROM_USER_UNCHECKED gymnastics, avoiding a lock order reversal on
FreeBSD. Performance impact is negligible -- no difference on r200 to
~1% improvement on rv200 in quake3 tests (P4 1Ghz, demofour at
1024x768, n=4 or 5).
FreeBSD. Add drm_get_resource_{start|len} so linux-specific stuff
doesn't need to be in shared code.
- Fix mach64 build by using __DECONST to work around passing a const
pointer to useracc, which is unfortunately not marked const.
- Get rid of a lot of maplist code by not having dev->maplist be a pointer,
and by sticking the link entries directly in drm_local_map_t rather
than having a separate structure for the linked list.
- Factor out map uninit and removal into its own routine, rather than
duplicating in both drm_takedown() and drm_rmmap().
- Hook up more driver functions, and correct FreeBSD-specific bits of
radeon_cp.c, making radeon work.
- Baby steps towards using bus_space as we should.
Dave Airlie
Maarten Maathuis
Jay Estabrook
Thomas Hellstrom
Michel Dänzer
Adam Jackson
Thomas Hellstrom
Stephane Marchesin
Matthieu Castet
Jeremy Kolb
Christoph Hellwig
Dave Airlie
Ben Skeggs
Dave Airlie
Michael Buesch
Dave Airlie
Tilman Sauerbeck
George Sapountzis
Felix Kuhling
Dave Airlie
Eric Anholt
Alan Hourihane
Thomas Hellstrom
Eric Anholt
Ian Romanick
Adam Jackson
Jon Smirl