I want to be able to use REF_CNT from other places in the kernel without
pushing a fence object onto the list of emitted fences.

The current code makes an assumption that every time the acked sequence is
bumped that there's at least one fence on the list that'll be signalled.

This will no longer be true in the near future.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

These are FIFO methods, it doesn't matter what subchannel is being used.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Fence lock needs to be initialized before any call to nouveau_channel_put
because it calls nouveau_channel_idle->nouveau_fence_update which uses
fence lock.

BUG: spinlock bad magic on CPU#0, test/24134
 lock: ffff88019f90dba8, .magic: 00000000, .owner: <none>/-1, .owner_cpu: 0
Pid: 24134, comm: test Not tainted 3.0.0-nv+ #800
Call Trace:
 spin_bug+0x9c/0xa3
 do_raw_spin_lock+0x29/0x13c
 _raw_spin_lock+0x1e/0x22
 nouveau_fence_update+0x2d/0xf1
 nouveau_channel_idle+0x22/0xa0
 nouveau_channel_put_unlocked+0x84/0x1bd
 nouveau_channel_put+0x20/0x24
 nouveau_channel_alloc+0x4ec/0x585
 nouveau_ioctl_fifo_alloc+0x50/0x130
 drm_ioctl+0x289/0x361
 do_vfs_ioctl+0x4dd/0x52c
 sys_ioctl+0x42/0x65
 system_call_fastpath+0x16/0x1b

It's easily triggerable from userspace.

Additionally remove double initialization of chan->fence.pending.
Signed-off-by: NMarcin Slusarz <marcin.slusarz@gmail.com>
Cc: stable@kernel.org
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

The HW will only accept the DMA_FROM_MEMORY class for DMA_SEMAPHORE without
asking the driver to intervene.

It appears that semaphores will work correctly even without DMA_IN_MEMORY,
so lets avoid the large amount of interrupts generated by x-chan sync.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

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

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

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

The DDX modifies DMA_SEMAPHORE on nv50 in order to implement sync-to-vblank,
things will go very wrong for cross-channel sync after this.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Commit 21e86c1c ("drm/nouveau: remove
cpu_writers lock") turned on lazy waits. Unfortunately
__nouveau_fence_wait was not optimized for this case and on HZ=100
kernel wasted up to 10 ms per call.

Depending on application, it led to 10-30% FPS regression.

Fix it.
Signed-off-by: NMarcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>

'mappable' isn't really used at all, nor is it necessary anymore as the
bo code is capable of moving buffers to mappable vram as required.

'no_vm' isn't necessary anymore either, any places that don't want to be
mapped into a GPU address space should allocate the VRAM directly instead.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

This gives a small, but noticeable performance gain at lower performance
levels, and unchanged at the higher ones.

With this commit, we're now using the same timeslice size as the NVIDIA
binary driver currently does, and dropping an unknown bit that NVIDIA
no longer appear to set.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

We may well be making more use of semaphores in the future, having the
entire VM available makes requiring DMA objects for each and every
semaphore block unnecessary.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Evil, evil chipset.  Worst of both worlds.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

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

These are the same semaphores nvc0 will use, and they potentially allow
us to do much cooler things than our current inter-channel sync impl.

Lets switch to them where possible now for some testing.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Current 3D driver expects this behaviour.  While this could be changed,
there's no compelling reason to reserve more than one subchannel for the
DRM.  If we ever need to use an object other then M2MF, we can just
re-bind subchannel 0 as required.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Just simple REF_CNT fencing for the moment.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

At some point in the future, this bo won't necessarily be backed by
a drm_mm_node, so use the start/size fields of the ttm_mem_reg instead.
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

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

Sleeping doesn't pay off for very short delays in comparison with the
minimum granularity of schedule_timeout().
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Reviewed-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

nouveau_fence_* functions are not type safe, which could lead to bugs.
Additionally every use of nouveau_fence_unref had to cast struct
nouveau_fence to void **.
Fix it by renaming old functions and creating static inline functions with
new prototypes. We still need old functions, because we pass function
pointers to ttm.
As we are wrapping functions, drop unused "void *arg" parameter where possible.
Signed-off-by: NMarcin Slusarz <marcin.slusarz@gmail.com>
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

The structs themselves, as well as the non-sw object creation function are
probably very misnamed now.  That's a problem for later :)
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Without it there's a potential race with nouveau_fence_update().
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

It needs a "strong" channel reference because it actually writes to
the channel pushbuf, otherwise the corresponding FIFO context could
get kicked off in the middle of nouveau_fence_sync().
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Fences didn't increment the channel reference count, and the fenced
channel could go away at any time. Fixes a potential race in
nouveau_fence_update().
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

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

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

Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Reported-by: NChristoph Bumiller <e0425955@student.tuwien.ac.at>
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Tested-by: NMaarten Maathuis <madman2003@gmail.com>
Tested-by: NXavier Chantry <chantry.xavier@gmail.com>
Tested-by: NBen Skeggs <bskeggs@redhat.com>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

Mainly to make room for inter-channel sync.
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

It fixes a race between the TTM delayed work queue and the GEM IOCTLs
(fdo bug 29583) uncovered by the BKL removal.
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

set_current_state() is called only once before the first iteration.
After return from schedule_timeout() current state is TASK_RUNNING. If
we are going to wait again, set_current_state() must be called.
Signed-off-by: NKulikov Vasiliy <segooon@gmail.com>
Signed-off-by: NFrancisco Jerez <currojerez@riseup.net>
Signed-off-by: NBen Skeggs <bskeggs@redhat.com>

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

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: NFrancisco Jerez <currojerez@riseup.net>

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: NFrancisco Jerez <currojerez@riseup.net>