/ssd/git/drm-core-next/drivers/gpu/drm/radeon/radeon_fence.c: In function ‘radeon_debugfs_fence_info’:
/ssd/git/drm-core-next/drivers/gpu/drm/radeon/radeon_fence.c:606:7: warning: format ‘%lx’ expects argument of type ‘long unsigned int’, but argument 3 has type ‘long long int’ [-Wformat]
Signed-off-by: NDave Airlie <airlied@redhat.com>

It never really belonged there in the first place.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Directly use the suballocator to get small chunks of memory.
It's equally fast and doesn't crash when we encounter a GPU reset.

v2: rebased on new SA interface.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Use one wait queue for all rings. When one ring progress, other
likely does to and we are not expecting to have a lot of waiter
anyway.

Also add a fence_wait_any that will wait until the first fence
in the fence array (one fence per ring) is signaled. This allow
to wait on all rings.

v2: some minor cleanups and improvements.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Some callers illegal called fence_wait_next/empty
while holding the ring emission mutex. So don't
relock the mutex in that cases, and move the actual
locking into the fence code.

v2: Don't try to unlock the mutex if it isn't locked.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Using 64bits fence sequence we can directly compare sequence
number to know if a fence is signaled or not. Thus the fence
list became useless, so does the fence lock that mainly
protected the fence list.

Things like ring.ready are no longer behind a lock, this should
be ok as ring.ready is initialized once and will only change
when facing lockup. Worst case is that we return an -EBUSY just
after a successfull GPU reset, or we go into wait state instead
of returning -EBUSY (thus delaying reporting -EBUSY to fence
wait caller).

v2: Remove left over comment, force using writeback on cayman and
    newer, thus not having to suffer from possibly scratch reg
    exhaustion
v3: Rebase on top of change to uint64 fence patch
v4: Change DCE5 test to force write back on cayman and newer but
    also any APU such as PALM or SUMO family
v5: Rebase on top of new uint64 fence patch
v6: Just break if seq doesn't change any more. Use radeon_fence
    prefix for all function names. Even if it's now highly optimized,
    try avoiding polling to often.
v7: We should never poll the last_seq from the hardware without
    waking the sleeping threads, otherwise we might lose events.
Signed-off-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NDave Airlie <airlied@redhat.com>

This convert fence to use uint64_t sequence number intention is
to use the fact that uin64_t is big enough that we don't need to
care about wrap around.

Tested with and without writeback using 0xFFFFF000 as initial
fence sequence and thus allowing to test the wrap around from
32bits to 64bits.

v2: Add comment about possible race btw CPU & GPU, add comment
    stressing that we need 2 dword aligned for R600_WB_EVENT_OFFSET
    Read fence sequenc in reverse order of GPU write them so we
    mitigate the race btw CPU and GPU.

v3: Drop the need for ring to emit the 64bits fence, and just have
    each ring emit the lower 32bits of the fence sequence. We
    handle the wrap over 32bits in fence_process.

v4: Just a small optimization: Don't reread the last_seq value
    if loop restarts, since we already know its value anyway.
    Also start at zero not one for seq value and use pre instead
    of post increment in emmit, otherwise wait_empty will deadlock.
Signed-off-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NDave 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: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

It's never used and so practically superfluous.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

As discussed with Michel that name better
describes the behavior of this function.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NMichel Dänzer <michel.daenzer@amd.com>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

We should signal the caller that we haven't waited at all.

v2: only change fence_wait_next not fence_wait_last.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NMichel Dänzer <michel.daenzer@amd.com>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Previusly multiple rings could trigger multiple GPU
resets at the same time.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

It makes no sense at all to have more than one flag.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Different rings have different criteria to test
if they are stuck.

v2: rebased on current drm-next
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Silly bad return path.

Reported-and-Tested-by: Mikko Vinni
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
CC: stable@vger.kernel.org
Signed-off-by: NDave 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: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Add a start fence driver helper function which will be call
once for each ring and will compute cpu/gpu addr for fence
depending on wether to use wb buffer or scratch reg.

This patch replace initialize fence driver separately which
was broken in regard of GPU lockup. The fence list for created,
emited, signaled must be initialize once and only from the
asic init callback not from the startup call back which is
call from the gpu reset.

v2: With this in place we no longer need to know the number of
    rings in fence_driver_init, also writing to the scratch reg
    before knowing its offset is a bad idea.

v3: rebase on top of change to previous patch in the serie
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Signed-off-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

That naming seems to make more sense, since we not
only want to run PM4 rings with it.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Split counting of emited fences out of power
management into a seperate function.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Emitting fences, semaphores and ib works differently
on different ring, so its is easier to maintain
separate functions for each ring.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Replace cp, cp1 and cp2 members with just an array
of radeon_cp structs.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave 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: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

For supporting multiple CP ring buffers, async DMA
engines and UVD.  We still need a way to synchronize
between engines.

v2 initialize unused fence driver ring to avoid issue in
   suspend/unload
Signed-off-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Better fix it before this obvious typo spreads even more.
Signed-off-by: NChristian König <deathsimple@vodafone.de>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Reviewed-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

abrt files a lot of bug reports when users get GPU lockups, but there's not really
enough context to do anything useful with them.  Given the lack of GPU context being
dumped, this patch removes the stack trace, so that abrt ignores the messages.
Signed-off-by: NDave Jones <davej@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

This is a resend from the original, changing the title from PATCH to
RFC(since this is a review for commit, and I should have put that the first go around).
and also removing some of the commit's with ia64 and bash since it is significant.
let me know if I might have missed anything etc..
Signed-off-by: NJustin P. Mattock <justinmattock@gmail.com>
Signed-off-by: NJiri Kosina <jkosina@suse.cz>

This allows us to move duplicated code in <asm/atomic.h>
(atomic_inc_not_zero() for now) to <linux/atomic.h>
Signed-off-by: NArun Sharma <asharma@fb.com>
Reviewed-by: NEric Dumazet <eric.dumazet@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: NMike Frysinger <vapier@gentoo.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The existing code assumed scratch registers in a number
of places while in most cases we are be using writeback
and events rather than scratch registers.
Signed-off-by: NAlex Deucher <alexdeucher@gmail.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

There's no need to pass kref_put() the address of a function (just the
function will do just fine) nor to cast its unused return to void.
Signed-off-by: NPaul Bolle <pebolle@tiscali.nl>
Signed-off-by: NJiri Kosina <jkosina@suse.cz>

This is necessary even with PCI(e) GART, and it makes writeback work even with
AGP on my PowerBook. Might still be unreliable with older revisions of UniNorth
and other AGP bridges though.
Signed-off-by: NMichel Dänzer <daenzer@vmware.com>
Reviewed-by: NAlex Deucher <alex.deucher@gmail.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Merge list_del() + list_add_tail() to list_move_tail().
Signed-off-by: NNicolas Kaiser <nikai@nikai.net>
Signed-off-by: NDave Airlie <airlied@redhat.com>

this adds a bo create, and fence seq tracking tracepoints.

This is just an initial set to play around with, we should investigate
what others we need would be useful.
Signed-off-by: NDave Airlie <airlied@redhat.com>

Coalesce long formats.
Align arguments.
Add missing newlines.
Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

On r6xx+ a newer fence mechanism was implemented to replace
the old wait_until plus scratch regs setup.  A single EOP event
will flush the destination caches, write a fence value, and generate
an interrupt.  This is the recommended fence mechanism on r6xx+ asics.

This requires my previous writeback patch.

v2: fix typo that enabled event fence checking on all asics
rather than just r6xx+.

v3: properly enable EOP interrupts
Should fix:
https://bugs.freedesktop.org/show_bug.cgi?id=29972Signed-off-by: NAlex Deucher <alexdeucher@gmail.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

When writeback is enabled, the GPU shadows writes to certain
registers into a buffer in memory.  The driver can then read
the values from the shadow rather than reading back from the
register across the bus.  Writeback can be disabled by setting
the no_wb module param to 1.

On r6xx/r7xx/evergreen, the following registers are shadowed:
- CP scratch registers
- CP read pointer
- IH write pointer
On r1xx-rr5xx, the following registers are shadowed:
- CP scratch registers
- CP read pointer

v2:
- Combine wb patches for r6xx-evergreen and r1xx-r5xx
- Writeback is disabled on AGP boards since it tends to be
unreliable on AGP using the gart.
- Check radeon_wb_init return values properly.
Signed-off-by: NAlex Deucher <alexdeucher@gmail.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Previous reset code leaded to computer hard lockup (need to unplug
the power too reboot the computer) on various configuration. This
patch change the reset code to avoid hard lockup. The GPU reset
is failing most of the time but at least user can log in remotely
or properly shutdown the computer.

Two issues were leading to hard lockup :
- Writting to the scratch register lead to hard lockup most likely
because the write back mecanism is in fuzy state after GPU lockup.
- Resetting the GPU memory controller and not reinitializing it
after leaded to hard lockup. We did only reinitialize in case of
successfull reset thus unsuccessfull reset quickly leaded to hard
lockup.
Signed-off-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave 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: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

Patch rename gpu_reset to asic_reset in prevision of having
gpu_reset doing more stuff than just basic asic reset.
Signed-off-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

This patch cleanup the fence code, it drops the timeout field of
fence as the time to complete each IB is unpredictable and shouldn't
be bound.

The fence cleanup lead to GPU lockup detection improvement, this
patch introduce a callback, allowing to do asic specific test for
lockup detection. In this patch the CP is use as a first indicator
of GPU lockup. If CP doesn't make progress during 1second we assume
we are facing a GPU lockup.

To avoid overhead of testing GPU lockup frequently due to fence
taking time to be signaled we query the lockup callback every
500msec. There is plenty code comment explaining the design & choise
inside the code.

This have been tested mostly on R3XX/R5XX hw, in normal running
destkop (compiz firefox, quake3 running) the lockup callback wasn't
call once (1 hour session). Also tested with forcing GPU lockup and
lockup was reported after the 1s CP activity timeout.

V2 switch to 500ms timeout so GPU lockup get call at least 2 times
   in less than 2sec.
V3 store last jiffies in fence struct so on ERESTART, EBUSY we keep
   track of how long we already wait for a given fence
V4 make sure we got up to date cp read pointer so we don't have
   false positive
Signed-off-by: NJerome Glisse <jglisse@redhat.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>