safe/jmp/linux-2.6
14 years agodrm/nv50: Improve PGRAPH interrupt handling.
Marcin Kościelnicki [Mon, 1 Mar 2010 00:18:39 +0000 (00:18 +0000)]
drm/nv50: Improve PGRAPH interrupt handling.

This makes nouveau recognise and report more kinds of PGRAPH errors, as
well as prevent GPU lockups resulting from some of them.

Lots of guesswork was involved and some part of this is probably
incorrect. Some potential-lockuop situations are handled by just
resetting a whole PGRAPH subunit, which doesn't sound like a "proper"
solution, but seems to work just fine... for now.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: Make ctxprog wait until interrupt handler is done.
Marcin Kościelnicki [Sun, 28 Feb 2010 23:45:38 +0000 (23:45 +0000)]
drm/nv50: Make ctxprog wait until interrupt handler is done.

This will fix races between generated ctxprogs and interrupt handler.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Fix fbcon corruption with font width not divisible by 8
Marcin Kościelnicki [Sat, 27 Feb 2010 18:13:35 +0000 (18:13 +0000)]
drm/nouveau: Fix fbcon corruption with font width not divisible by 8

NV50 is nice and has a switch that autoaligns stuff for us. Pre-NV50,
we need to align input bitmap width manually.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: Remove redundant/incorrect ctxvals initialisation.
Marcin Kościelnicki [Thu, 25 Feb 2010 18:53:37 +0000 (18:53 +0000)]
drm/nv50: Remove redundant/incorrect ctxvals initialisation.

11c/004 offset corresponds to PGRAPH reg 0x400828, and is initialised
earlier anyway by both our ctxprog generator and blob ctxvals. It's
actually incorrect with the generator, since we use different layout
on pre-NVA0.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: use ALIGN instead of open coding it
Matt Turner [Thu, 25 Feb 2010 04:27:10 +0000 (23:27 -0500)]
drm/nouveau: use ALIGN instead of open coding it

CC: Ben Skeggs <bskeggs@redhat.com>
Signed-off-by: Matt Turner <mattst88@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: report unknown connector state if lid closed
Ben Skeggs [Thu, 25 Feb 2010 02:12:09 +0000 (12:12 +1000)]
drm/nouveau: report unknown connector state if lid closed

This is in preference to disconnected.  If there's no other outputs
connected this will cause LVDS to be programmed even with the lid
closed rather than having X fail to start because of no available
outputs.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: support version 0x20 displayport tables
Ben Skeggs [Thu, 25 Feb 2010 01:53:00 +0000 (11:53 +1000)]
drm/nouveau: support version 0x20 displayport tables

Not entirely identical to 0x21, the per-encoder table header lacks the
third init table pointer.  However, our current parsing of the table
should work just fine.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Fix noaccel/nofbaccel option descriptions.
Marcin Kościelnicki [Thu, 25 Feb 2010 00:54:04 +0000 (00:54 +0000)]
drm/nouveau: Fix noaccel/nofbaccel option descriptions.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: Implement ctxprog/state generation.
Marcin Kościelnicki [Thu, 25 Feb 2010 00:54:02 +0000 (00:54 +0000)]
drm/nv50: Implement ctxprog/state generation.

This removes dependence on external firmware for NV50 generation cards.
If the generated ctxprogs don't work for you for some reason, please
report it.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: use dcb connector types throughout the driver
Ben Skeggs [Wed, 24 Feb 2010 04:01:40 +0000 (14:01 +1000)]
drm/nouveau: use dcb connector types throughout the driver

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: enable hpd on any connector we know the gpio line for
Ben Skeggs [Wed, 24 Feb 2010 03:45:57 +0000 (13:45 +1000)]
drm/nv50: enable hpd on any connector we know the gpio line for

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: use dcb connector table for creating drm connectors
Ben Skeggs [Wed, 24 Feb 2010 02:07:31 +0000 (12:07 +1000)]
drm/nouveau: use dcb connector table for creating drm connectors

This makes this code common to both the nv04 and nv50 paths.

For the moment, we keep the previous behaviour with HDMI/eDP connectors
and report them as DVI-D/DP instead.  This will be fixed once the rest
of the code has been fixed to deal with those types.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: construct a connector table for cards that lack a real one
Ben Skeggs [Wed, 24 Feb 2010 01:53:59 +0000 (11:53 +1000)]
drm/nouveau: construct a connector table for cards that lack a real one

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: check for known dcb connector types
Ben Skeggs [Wed, 24 Feb 2010 01:09:20 +0000 (11:09 +1000)]
drm/nouveau: check for known dcb connector types

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: parse dcb gpio/connector tables after encoders
Ben Skeggs [Wed, 24 Feb 2010 00:48:16 +0000 (10:48 +1000)]
drm/nouveau: parse dcb gpio/connector tables after encoders

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: reorganise bios header, add dcb connector type enums
Ben Skeggs [Wed, 24 Feb 2010 00:31:39 +0000 (10:31 +1000)]
drm/nouveau: reorganise bios header, add dcb connector type enums

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: merge nvbios and nouveau_bios_info
Ben Skeggs [Wed, 24 Feb 2010 00:03:05 +0000 (10:03 +1000)]
drm/nouveau: merge nvbios and nouveau_bios_info

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: merge parsed_dcb and bios_parsed_dcb into dcb_table
Ben Skeggs [Tue, 23 Feb 2010 23:56:18 +0000 (09:56 +1000)]
drm/nouveau: merge parsed_dcb and bios_parsed_dcb into dcb_table

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: rename parsed_dcb_gpio to dcb_gpio_table
Ben Skeggs [Tue, 23 Feb 2010 23:46:27 +0000 (09:46 +1000)]
drm/nouveau: rename parsed_dcb_gpio to dcb_gpio_table

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: allow retrieval of vbios image from debugfs
Ben Skeggs [Tue, 23 Feb 2010 01:39:02 +0000 (11:39 +1000)]
drm/nouveau: allow retrieval of vbios image from debugfs

It's very useful to be able to access this without additional tools for
debugging purposes.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix missing spin_unlock in failure path
Luca Barbieri [Sat, 20 Feb 2010 10:08:56 +0000 (11:08 +0100)]
drm/nouveau: fix missing spin_unlock in failure path

Found by sparse.

Signed-off-by: Luca Barbieri <luca@luca-barbieri.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix i2ctable bounds checking
Marcin Slusarz [Mon, 15 Feb 2010 22:24:49 +0000 (23:24 +0100)]
drm/nouveau: fix i2ctable bounds checking

i2c_entries seems to be the number of i2c entries,
so with index equal to this number, we could read
invalid data from i2ctable. Fix it.

Signed-off-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix nouveau_i2c_find bounds checking
Marcin Slusarz [Wed, 17 Feb 2010 18:04:01 +0000 (19:04 +0100)]
drm/nouveau: fix nouveau_i2c_find bounds checking

Reported-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix pramdac_table range checking
Marcin Slusarz [Wed, 17 Feb 2010 18:04:00 +0000 (19:04 +0100)]
drm/nouveau: fix pramdac_table range checking

get_tmds_index_reg reads some value from stack when mlv happens
to be equal to size of pramdac_table array. Fix it.

Reported-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: new gem pushbuf interface, bump to 0.0.16
Ben Skeggs [Fri, 12 Feb 2010 00:27:35 +0000 (10:27 +1000)]
drm/nouveau: new gem pushbuf interface, bump to 0.0.16

This commit breaks the userspace interface, and requires a new libdrm for
nouveau to operate again.

The multiple GEM_PUSHBUF ioctls that were present in 0.0.15 for
compatibility purposes are now gone, and replaced with the new ioctl which
allows for multiple push buffers to be submitted (necessary for hw index
buffers in the nv50 3d driver) and relocations to be applied on any buffer.

A number of other ioctls (CARD_INIT, GEM_PIN, GEM_UNPIN) that were needed
for userspace modesetting have also been removed.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
14 years agodrm/nv50: make pushbuf dma object cover entire vm
Ben Skeggs [Fri, 12 Feb 2010 01:11:54 +0000 (11:11 +1000)]
drm/nv50: make pushbuf dma object cover entire vm

This allows us to submit push buffers from any memtype to the hardware.
We'll need this ability for VRAM index buffers at some point.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: remove PUSHBUF_CAL macro
Ben Skeggs [Thu, 11 Feb 2010 23:05:42 +0000 (09:05 +1000)]
drm/nouveau: remove PUSHBUF_CAL macro

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: switch to indirect push buffer controls
Ben Skeggs [Thu, 11 Feb 2010 06:37:26 +0000 (16:37 +1000)]
drm/nv50: switch to indirect push buffer controls

PFIFO on G80 and up has a new mode where the main ring buffer is simply a
ring of pointers to indirect buffers containing the actual command/data
packets.  In order to be able to implement index buffers in the 3D driver
we need to be able to submit data-only push buffers right after the cmd
packet header, which is only possible using the new command submission
method.

This commit doesn't make it possible to implement index buffers yet, some
userspace interface changes will be required, but it does allow for
testing/debugging of the hardware-side support in the meantime.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: protect channel create/destroy and irq handler with a spinlock
Maarten Maathuis [Mon, 1 Feb 2010 19:58:27 +0000 (20:58 +0100)]
drm/nouveau: protect channel create/destroy and irq handler with a spinlock

The nv50 pgraph handler (for example) could reenable pgraph fifo access
and that would be bad when pgraph context is being unloaded (we need the
guarantee a ctxprog isn't running).

Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: fix vram ptes on IGPs to point at stolen system memory
Ben Skeggs [Fri, 19 Feb 2010 22:10:11 +0000 (08:10 +1000)]
drm/nv50: fix vram ptes on IGPs to point at stolen system memory

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: fix instmem binding on IGPs to point at stolen system memory
Ben Skeggs [Fri, 19 Feb 2010 22:06:36 +0000 (08:06 +1000)]
drm/nv50: fix instmem binding on IGPs to point at stolen system memory

This also modifies the unused PRAMIN PT entries to be all zeroes, can't
really recall why I used 9/0 initially, just that it didn't work for
some reason.  It was likely masking a bug elsewhere that's since been
fixed.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: improve vram page table construction
Ben Skeggs [Thu, 11 Feb 2010 01:31:44 +0000 (11:31 +1000)]
drm/nv50: improve vram page table construction

This commit changes nouveau to construct PTEs which look very much like
the ones the binary driver creates.

I presume that filling multiple PTEs identically with length flags and
the physical address of the start of a block of VRAM is a hint to the
memory controller that it need not perform additional page table lookups
for that range of addresses.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: more efficient clearing of gpu page table entries
Ben Skeggs [Thu, 11 Feb 2010 00:25:53 +0000 (10:25 +1000)]
drm/nv50: more efficient clearing of gpu page table entries

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: make nv50_mem_vm_{bind,unbind} operate only on vram
Ben Skeggs [Thu, 11 Feb 2010 00:23:30 +0000 (10:23 +1000)]
drm/nv50: make nv50_mem_vm_{bind,unbind} operate only on vram

GART is handled elsewhere, no reason to have the code for it here too.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Fix up pre-nv17 analog load detection.
Francisco Jerez [Wed, 10 Feb 2010 18:25:58 +0000 (19:25 +0100)]
drm/nouveau: Fix up pre-nv17 analog load detection.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Force TV encoder DPMS reinit after resume.
Francisco Jerez [Thu, 11 Feb 2010 11:47:40 +0000 (12:47 +0100)]
drm/nouveau: Force TV encoder DPMS reinit after resume.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: use mutex for vbios lock
Ben Skeggs [Tue, 16 Feb 2010 01:14:14 +0000 (11:14 +1000)]
drm/nouveau: use mutex for vbios lock

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agonouveau: fix state detection with switchable graphics
Matthew Garrett [Wed, 10 Feb 2010 10:50:34 +0000 (20:50 +1000)]
nouveau: fix state detection with switchable graphics

Signed-off-by: Matthew Garrett <mjg@redhat.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: move dereferences after null checks
Marcin Slusarz [Sat, 30 Jan 2010 14:41:00 +0000 (15:41 +0100)]
drm/nouveau: move dereferences after null checks

Reported-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: make the pgraph irq handler loop like the pre-nv50 version
Maarten Maathuis [Wed, 20 Jan 2010 18:54:34 +0000 (19:54 +0100)]
drm/nv50: make the pgraph irq handler loop like the pre-nv50 version

Unset the bit that indicates that a ctxprog can continue at the end.

Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: delete ramfc object after disabling fifo, not before
Maarten Maathuis [Mon, 1 Feb 2010 17:47:52 +0000 (18:47 +0100)]
drm/nv50: delete ramfc object after disabling fifo, not before

ramfc is zero'ed upon destruction, so it's safer to do things in the right
order.

Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: avoid unloading pgraph context when ctxprog is running
Maarten Maathuis [Mon, 1 Feb 2010 17:32:09 +0000 (18:32 +0100)]
drm/nv50: avoid unloading pgraph context when ctxprog is running

- We need to disable pgraph fifo access before checking the current channel,
  otherwise we could still hit a running ctxprog.
- The writes to 0x400500 are already handled by pgraph->fifo_access and are
  therefore redundant, moreover pgraph fifo access should not be reenabled
  before current context is set as invalid. So remove them altogether.

Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: align size of buffer object to the right boundaries.
Maarten Maathuis [Sun, 27 Dec 2009 11:22:07 +0000 (12:22 +0100)]
drm/nv50: align size of buffer object to the right boundaries.

- In the current situation the padding that is added is dangerous to write
  to, userspace could potentially overwrite parts of another bo.
- Depth and stencil buffers are supposed to be large enough in general so
  the waste of memory should be acceptable.
- Alternatives are hiding the padding from users or splitting vram into 2
  zones.

Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: disregard dac outputs in nv50_sor_dpms()
Ben Skeggs [Tue, 9 Feb 2010 02:30:35 +0000 (12:30 +1000)]
drm/nv50: disregard dac outputs in nv50_sor_dpms()

Fixes DVI+VGA on my 9400, and likely a lot of other configurations that
got broken by the previos DVI-over-DP fix.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: prevent multiple init tables being parsed at the same time
Ben Skeggs [Tue, 9 Feb 2010 00:22:29 +0000 (10:22 +1000)]
drm/nv50: prevent multiple init tables being parsed at the same time

With DVI and DP plugged, the DVI clock change interrupts being run can
cause DP link training to fail.  This adds a spinlock around init table
parsing to prevent this.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: make dp auxch xfer len check for reads only
Ben Skeggs [Tue, 9 Feb 2010 00:08:34 +0000 (10:08 +1000)]
drm/nouveau: make dp auxch xfer len check for reads only

Writes don't return a count, and adding the check broke native DP.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv40: make INIT_COMPUTE_MEM a NOP, just like nv50
Ben Skeggs [Fri, 22 Jan 2010 03:17:28 +0000 (13:17 +1000)]
drm/nv40: make INIT_COMPUTE_MEM a NOP, just like nv50

It appears we aren't required to do memory sizing ourselves on nv40
either.  NV40 init tables read a strap from PEXTDEV_BOOT_0 into a
CRTC register, and then later use that value to select a memory
configuration (written to PFB_CFG0, just like INIT_COMPUTE_MEM on
earlier cards) with INIT_IO_RESTRICT_PROG.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Add proper vgaarb support.
Marcin Kościelnicki [Mon, 8 Feb 2010 00:20:17 +0000 (00:20 +0000)]
drm/nouveau: Add proper vgaarb support.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Fix fbcon on mixed pre-NV50 + NV50 multicard.
Marcin Kościelnicki [Wed, 27 Jan 2010 14:03:18 +0000 (14:03 +0000)]
drm/nouveau: Fix fbcon on mixed pre-NV50 + NV50 multicard.

We used single shared fbops struct and patched it at fb init time with
pointers to the right variant. On mixed multicard, this meant that
it was either sending NV50-style commands to all cards, or NV04-style
commands to all cards.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrivers/gpu/drm/nouveau/nouveau_grctx.c: correct NULL test
Julia Lawall [Tue, 2 Feb 2010 22:40:30 +0000 (14:40 -0800)]
drivers/gpu/drm/nouveau/nouveau_grctx.c: correct NULL test

Test the just-allocated value for NULL rather than some other value.

The semantic patch that makes this change is as follows:
(http://coccinelle.lip6.fr/)

// <smpl>
@@
expression x,y;
statement S;
@@

x = \(kmalloc\|kcalloc\|kzalloc\)(...);
(
if ((x) == NULL) S
|
if (
-   y
+   x
       == NULL)
 S
)
// </smpl>

Signed-off-by: Julia Lawall <julia@diku.dk>
Cc: David Airlie <airlied@linux.ie>
Cc: Ben Skeggs <bskeggs@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: call ttm_bo_wait with the bo lock held to prevent hang
Luca Barbieri [Sat, 30 Jan 2010 22:21:38 +0000 (23:21 +0100)]
drm/nouveau: call ttm_bo_wait with the bo lock held to prevent hang

nouveau_gem_ioctl_cpu_prep calls ttm_bo_wait without the bo lock held.
ttm_bo_wait unlocks that lock, and so must be called with it held.

Currently this bug causes libdrm nouveau_bo_busy() to hang the machine.

Signed-off-by: Luca Barbieri <luca at luca-barbieri.com>
Acked-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Fixup semaphores on pre-nv50 cards.
Francisco Jerez [Sat, 30 Jan 2010 17:28:00 +0000 (18:28 +0100)]
drm/nouveau: Fixup semaphores on pre-nv50 cards.

Apparently, they generate a PFIFO interrupt each time one of the
semaphore methods is executed if its ctxdma wasn't manually marked as
valid. This patch makes it flip the valid bit in response to the
DMA_SEMAPHORE method (which triggers the IRQ even for a valid ctxdma).

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Add getparam to get available PGRAPH units.
Marcin Kościelnicki [Tue, 26 Jan 2010 18:39:20 +0000 (18:39 +0000)]
drm/nouveau: Add getparam to get available PGRAPH units.

On nv50, this will be needed by applications using CUDA to know
how much stack/local memory to allocate.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Add module options to disable acceleration.
Marcin Kościelnicki [Tue, 26 Jan 2010 14:00:42 +0000 (14:00 +0000)]
drm/nouveau: Add module options to disable acceleration.

noaccel=1 disables all acceleration and doesn't even attempt
initialising PGRAPH+PFIFO, nofbaccel=1 only makes fbcon unaccelerated.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix non-vram notifier blocks
Ben Skeggs [Wed, 27 Jan 2010 04:29:05 +0000 (14:29 +1000)]
drm/nouveau: fix non-vram notifier blocks

Due to a thinko, these were previously forced to VRAM even if we allocated
them in GART.

This commit fixes that bug, but keeps the previous behaviour of using VRAM
by default until it's been tested properly across more chipsets.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: prevent switching off SOR when in use for DVI-over-DP
Ben Skeggs [Fri, 22 Jan 2010 00:57:01 +0000 (10:57 +1000)]
drm/nv50: prevent switching off SOR when in use for DVI-over-DP

Another hack because of us exposing each encoder block's function as
an encoder rather than exposing a single encoder that deals with them
all.

A proper fix will come, it's just rather invasive so this hack will
do until then.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: fail auxch transaction if reply count not what we expect
Ben Skeggs [Thu, 21 Jan 2010 23:10:05 +0000 (09:10 +1000)]
drm/nv50: fail auxch transaction if reply count not what we expect

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix failure path if userspace specifies no valid memtypes
Ben Skeggs [Thu, 21 Jan 2010 05:03:23 +0000 (15:03 +1000)]
drm/nouveau: fix failure path if userspace specifies no valid memtypes

We need to add the buffer to the list even if we fail, otherwise the
validate_fini() call won't unreserve + unreference the GEM object,
making TTM very unhappy.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: report LVDS as disconnected if lid closed
Ben Skeggs [Mon, 18 Jan 2010 01:42:37 +0000 (11:42 +1000)]
drm/nouveau: report LVDS as disconnected if lid closed

Also adds a module option to ignore the status reported via ACPI, in case
we hit systems with broken ACPI.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: prevent accidently turning off encoders we're actually using
Ben Skeggs [Sun, 17 Jan 2010 22:52:35 +0000 (08:52 +1000)]
drm/nv50: prevent accidently turning off encoders we're actually using

On most cards the DisplayPort connector is created with 2 encoders sharing
a single SOR (for native DP, and for DVI-over-DP).  The previous logic
for turning off unused encoders didn't take into account that we could
have multiple drm_encoders on a single hw encoder and ended up turning off
encoders that were actually being used still.

This patch fixes that issue.  We probably want to look at something a bit
better later on, and only expose one drm_encoder per hw encoder block.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: fix alignment of per-channel fifo cache
Ben Skeggs [Sun, 17 Jan 2010 22:33:04 +0000 (08:33 +1000)]
drm/nv50: fix alignment of per-channel fifo cache

GPU pointer to the structure is shifted right by 10 bits, so we need to
align to 1024 bytes, not 256.

Reported-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Evict buffers in VRAM before freeing sgdma
Luca Barbieri [Sat, 16 Jan 2010 14:30:15 +0000 (15:30 +0100)]
drm/nouveau: Evict buffers in VRAM before freeing sgdma

Currently, we take down the sgdma engine without evicting all buffers
from VRAM.

The TTM device release will try to evict anything in VRAM to GART
memory, but this will fail since sgdma has already been taken down.

This causes an infinite loop in kernel mode on module unload.
It usually doesn't happen because there aren't any buffer on close.
However, if the GPU is locked up, this condition is easily triggered.

This patch fixes it in the simplest way possible by cleaning VRAM
right before cleaning SGDMA memory.

Signed-off-by: Luca Barbieri <luca@luca-barbieri.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Acknowledge DMA_VTX_PROTECTION PGRAPH interrupts
Luca Barbieri [Sat, 16 Jan 2010 14:27:51 +0000 (15:27 +0100)]
drm/nouveau: Acknowledge DMA_VTX_PROTECTION PGRAPH interrupts

Currently Nouveau is unable to dismiss DMA_VTX_PROTECTION errors,
which results in an infinite loop in the interrupt handler.

These errors are caused both by bugs in the Gallium driver and by
user-specified index buffers with out of bounds indices.

By mmio-tracing the nVidia drivers, I found out how this is done.
On DMA_VTX_PROTECTION, The nVidia driver reads the register 0x402000,
always getting the value 4, and then writes 4 back to 0x402000.

This patch adds that logic by reading 0x402000 and writing the same
value back.
It's unclear what should happen if the value read is not 4, and
the current approach might not be the correct one.

To test this, modify mesa/progs/trivial/vbo-drawrange.c, defining
ELTOBJ to 1 and replacing indices with huge out of bounds integers.

Without this patch, the GPU and/or kernel should lock up.
With this patch, it should misrender as expected but not lock up.

The errors are still logged since they are useful for development.

This has been tested on NV49 and may not work on other cards.

To find out how things work on other cards, run the aforementioned
test using the blob with mmiotrace and grep for a read of the PGRAPH
source register.

Signed-off-by: Luca Barbieri <luca@luca-barbieri.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix thinko in nv04_instmem.c
Ben Skeggs [Fri, 15 Jan 2010 02:21:37 +0000 (12:21 +1000)]
drm/nouveau: fix thinko in nv04_instmem.c

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix a race condition in nouveau_dma_wait()
Ben Skeggs [Fri, 15 Jan 2010 02:08:57 +0000 (12:08 +1000)]
drm/nouveau: fix a race condition in nouveau_dma_wait()

Can be triggered easily on certain cards (NV46 and NV50 of mine) by
running "dmesg", the DRM's channel will lockup.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: check pushbuffer bounds in ioctl
Luca Barbieri [Sun, 10 Jan 2010 19:10:53 +0000 (20:10 +0100)]
drm/nouveau: check pushbuffer bounds in ioctl

Currently there is no check that the pushbuffer request bounds are inside
the TTM BO.

This allows to instruct the kernel to do relocations on user-selected
addresses, since the relocation bounds checking relies on the request
bounds.

This can oops the kernel accidentally and is easily exploitable.

This patch adds bound checking and alignment checking for ->offset and
->nr_dwords.

It also makes some variables unsigned, which should have no effect,
but prevents possible bounds checking problems.

Signed-off-by: Luca Barbieri <luca@luca-barbieri.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: reserve VGA area for the moment
Ben Skeggs [Thu, 14 Jan 2010 23:24:20 +0000 (09:24 +1000)]
drm/nouveau: reserve VGA area for the moment

This is to prevent things such as GART tables and other important GPU
structures being allocated there before we take over fbcon ourselves.

This is more of a workaround for the moment, a better solution will
require some more invasive changes, but it'll be done at some point.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Unset the EDID connector property when the EDID block goes away.
Francisco Jerez [Thu, 14 Jan 2010 14:54:23 +0000 (15:54 +0100)]
drm/nouveau: Unset the EDID connector property when the EDID block goes away.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Fallback to analog load detection when the EDID block is invalid.
Francisco Jerez [Thu, 14 Jan 2010 14:47:03 +0000 (15:47 +0100)]
drm/nouveau: Fallback to analog load detection when the EDID block is invalid.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix edid memleak in nouveau_connector
Xavier Chantry [Mon, 11 Jan 2010 21:42:21 +0000 (22:42 +0100)]
drm/nouveau: fix edid memleak in nouveau_connector

This was spotted by kmemleak.

Signed-off-by: Xavier Chantry <shiningxc@gmail.com>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Break some long lines.
Francisco Jerez [Thu, 14 Jan 2010 13:56:54 +0000 (14:56 +0100)]
drm/nouveau: Break some long lines.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: add NV18 device id to call_lvds_manufacturer_script
Andrea Tacconi [Thu, 14 Jan 2010 13:52:37 +0000 (14:52 +0100)]
drm/nouveau: add NV18 device id to call_lvds_manufacturer_script

This fixes imac black screen (NV18 card)

Signed-off-by: Andrea Tacconi <tacconet@libero.it>
Signed-off-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: Fix typo in PGRAPH initialisation.
Marcin Kościelnicki [Tue, 12 Jan 2010 18:21:56 +0000 (18:21 +0000)]
drm/nv50: Fix typo in PGRAPH initialisation.

This enables streamout functionality.

Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: less magic DCB 1.5 parsing
Ben Skeggs [Thu, 14 Jan 2010 07:53:04 +0000 (17:53 +1000)]
drm/nouveau: less magic DCB 1.5 parsing

This in the very least matches the parsing of all the previously known
entries, and hopefully (at least closer to) correct for any we haven't
seen yet.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: assume no nv04 board has a DCB table
Ben Skeggs [Thu, 14 Jan 2010 05:58:10 +0000 (15:58 +1000)]
drm/nouveau: assume no nv04 board has a DCB table

There's a report of a TNT2 where the DCB table pointer is *not* NULL
(it contains a part of a VBIOS data string), and we assume this means
a DCB table is present, causing all kinds of hilarity.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: remove PRIV0 check in nouveau_mem_close()
Ben Skeggs [Wed, 13 Jan 2010 01:22:33 +0000 (11:22 +1000)]
drm/nouveau: remove PRIV0 check in nouveau_mem_close()

We don't setup PRIV0 anymore, so this is unnecessary.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: wait on fence after bo move if validating for another channel
Ben Skeggs [Tue, 12 Jan 2010 05:28:19 +0000 (15:28 +1000)]
drm/nouveau: wait on fence after bo move if validating for another channel

Not an ideal solution, but it'll do for the moment for correctness.  We
need to come up with a nicer way to manage inter-channel sync, the hw
is unfortunately a little lacking in this area.

Should fix some resume corruption, as well as corruption that may be seen
while under memory pressure.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: trust init table registers are safe
Ben Skeggs [Tue, 12 Jan 2010 03:02:19 +0000 (13:02 +1000)]
drm/nouveau: trust init table registers are safe

Apparently the original reason for checking this was there were known
register accesses that caused hangs on some chipsets.  This was more
than likely because of incorrect parsing of previous opcodes, and I
hardly think aborting a script half way through is going to be any
better (in fact, we have had bug reports where this has been the cause
of s/r failures among other things).

This patch (which has been in Fedora 12 for a long time now) removes
all checking for known register ranges, and just leaves the check to
ensure the access is within the mapped aperture to avoid an oops.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: wait for pgraph to idle before unloading the context
Maarten Maathuis [Mon, 11 Jan 2010 20:18:53 +0000 (21:18 +0100)]
drm/nv50: wait for pgraph to idle before unloading the context

This should fix the problem with gpu hangs people have had when closing
channels.

Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv04: Fix set_operation software method.
Marcin Kościelnicki [Sun, 10 Jan 2010 17:09:14 +0000 (17:09 +0000)]
drm/nv04: Fix set_operation software method.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: initialise DMA tracking parameters earlier
Ben Skeggs [Fri, 8 Jan 2010 00:57:39 +0000 (10:57 +1000)]
drm/nouveau: initialise DMA tracking parameters earlier

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: use dma.max rather than pushbuf size for checking GET validity
Ben Skeggs [Fri, 8 Jan 2010 00:53:40 +0000 (10:53 +1000)]
drm/nouveau: use dma.max rather than pushbuf size for checking GET validity

Some upcoming G80 DMA changes will depend on this, but it's split out for
bisectibility just in case it causes some unexpected issues.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv04: differentiate between nv04/nv05
Ben Skeggs [Thu, 7 Jan 2010 03:47:57 +0000 (13:47 +1000)]
drm/nv04: differentiate between nv04/nv05

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: Fix null deref in nouveau_fence_emit due to deleted fence
Luca Barbieri [Wed, 6 Jan 2010 03:02:45 +0000 (04:02 +0100)]
drm/nouveau: Fix null deref in nouveau_fence_emit due to deleted fence

Currently Nouveau will unvalidate all buffers if it is forced to wait on
one, and then start revalidating from the beginning.  While doing so, it
destroys the operation fence, causing nouveau_fence_emit to crash.

This patch fixes this bug by taking the fence object out of validate_op
and creating it just before emit.  The fence pointer is initialized to 0
and unref'ed unconditionally.

In addition to fixing the bug, this prevents its reintroduction and
simplifies the code.

Signed-off-by: Luca Barbieri <luca@luca-barbieri.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: prevent a possible ctxprog hang
Ben Skeggs [Wed, 6 Jan 2010 02:00:02 +0000 (12:00 +1000)]
drm/nv50: prevent a possible ctxprog hang

The below is mainly an educated guess at what's going on, docs would
sure be handy...  NVIDIA? :P

It appears it's possible for a ctxprog to run even while a GPU exception
is pending.  The GF8 and up ctxprogs appear to have a small snippet of
code which detects this, and stalls the ctxprog until it's been handled,
which essentially looks like:

if (r2 & 0x00008000) {
r0 |= 0x80000000;
while (r0 & 0x80000000) {}
}

I don't know of any way that flag would get cleared unless the driver
intervenes (and indeed, in the cases I've seen the hang, nothing steps
in to automagically clear it for us).  This patch causes the driver to
clear the flag during the PGRAPH IRQ handler.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: have ttm's fault handler called directly
Ben Skeggs [Mon, 4 Jan 2010 05:52:20 +0000 (15:52 +1000)]
drm/nouveau: have ttm's fault handler called directly

There's no good reason for us to have our own anymore, this is left over
from an early port to these TTM interfaces.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: restore correct cache1 get/put address on fifoctx load
Ben Skeggs [Mon, 4 Jan 2010 23:41:05 +0000 (09:41 +1000)]
drm/nv50: restore correct cache1 get/put address on fifoctx load

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: create function for "dealing" with gpu lockup
Marcin Slusarz [Mon, 4 Jan 2010 18:25:09 +0000 (19:25 +0100)]
drm/nouveau: create function for "dealing" with gpu lockup

It's mostly a cleanup, but in nv50_fbcon_accel_init gpu lockup
message was printed, but HWACCEL_DISBALED flag was not set.

Signed-off-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: remove unused nouveau_channel_idle() function
Ben Skeggs [Mon, 4 Jan 2010 02:53:01 +0000 (12:53 +1000)]
drm/nouveau: remove unused nouveau_channel_idle() function

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nouveau: fix handling of fbcon colours in 8bpp
Ben Skeggs [Sun, 3 Jan 2010 23:10:55 +0000 (09:10 +1000)]
drm/nouveau: fix handling of fbcon colours in 8bpp

Depending on the visual, the colours handed to us in fillrect() can either be
an actual colour, or an index into the pseudo-palette.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv04: Context switching fixes.
Francisco Jerez [Sat, 26 Dec 2009 13:39:46 +0000 (14:39 +0100)]
drm/nv04: Context switching fixes.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
14 years agodrm/nouveau: Use the software object for fencing.
Francisco Jerez [Sat, 26 Dec 2009 01:09:36 +0000 (02:09 +0100)]
drm/nouveau: Use the software object for fencing.

This should avoid a race condition on nv0x, if we're doing it with
actual PGRAPH objects and a there's a fence within the FIFO DMA fetch
area when a context switch kicks in.

In that case we get an ILLEGAL_MTHD interrupt as expected, but the
values in PGRAPH_TRAPPED_ADDR aren't calculated correctly and they're
almost useless (e.g. you can see ILLEGAL_MTHDs for the now inactive
channel, with a wrong offset/data pair).

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
14 years agodrm/nouveau: Allocate a per-channel instance of NV_SW.
Francisco Jerez [Sat, 26 Dec 2009 01:42:45 +0000 (02:42 +0100)]
drm/nouveau: Allocate a per-channel instance of NV_SW.

It will be useful for various synchronization purposes, mostly stolen
from "[PATCH] drm/nv50: synchronize user channel after buffer object
move on kernel channel" by Maarten Maathuis.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
14 years agodrm/nv50: make the blocksize depend on vram size
Maarten Maathuis [Sat, 26 Dec 2009 20:46:36 +0000 (21:46 +0100)]
drm/nv50: make the blocksize depend on vram size

- This should be better than what we have now.
- I'm less sure about the non power of two path.

Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
14 years agodrm/nouveau: better alignment of bo sizes and use roundup instead of ALIGN
Maarten Maathuis [Fri, 25 Dec 2009 17:51:17 +0000 (18:51 +0100)]
drm/nouveau: better alignment of bo sizes and use roundup instead of ALIGN

- Aligning to block size should ensure that the extra size is enough.
- Using roundup, because not all sizes are powers of two.

Signed-off-by: Maarten Maathuis <madman2003@gmail.com>
14 years agodrm/nouveau: Don't skip card take down on nv0x.
Francisco Jerez [Tue, 22 Dec 2009 17:24:09 +0000 (18:24 +0100)]
drm/nouveau: Don't skip card take down on nv0x.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
14 years agodrm/nouveau: Implement nv42-nv43 TV load detection.
Francisco Jerez [Thu, 17 Dec 2009 17:57:44 +0000 (18:57 +0100)]
drm/nouveau: Implement nv42-nv43 TV load detection.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
14 years agodrm/nouveau: Clean up the nv17-nv4x load detection code a bit.
Francisco Jerez [Thu, 17 Dec 2009 17:52:44 +0000 (18:52 +0100)]
drm/nouveau: Clean up the nv17-nv4x load detection code a bit.

Signed-off-by: Francisco Jerez <currojerez@riseup.net>
14 years agodrm/nv50: fix fillrect color
Marcin Slusarz [Mon, 21 Dec 2009 22:00:41 +0000 (23:00 +0100)]
drm/nv50: fix fillrect color

struct fb_fillrect->color is not a color, but index into pseudo_palette
array

Signed-off-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
14 years agodrm/nv50: ignore vbios table's claim to the contrary if EDID says >8bpc
Ben Skeggs [Mon, 21 Dec 2009 02:16:52 +0000 (12:16 +1000)]
drm/nv50: ignore vbios table's claim to the contrary if EDID says >8bpc

Should fix dim panel issues reported on Dell M6400/M6500.

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>